Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

## RESEARCH ARTICLE 

2024, vol. 11, issue 2, 376 - 393 https://doi.org/10.5281/zenodo.15258343 

## **REINFORCEMENT LEARNING-DRIVEN LANGUAGE AGENTS FOR MULTI-DOMAIN FACT-CHECKING AND COHERENT NEWS SYNTHESIS** 

Dan  Valeriu VOINEA 

Lecturer PhD, Department of Arts and Media, University of Craiova, Romania https://orcid.org/0000-0002-1827-3002 

. 

## **Abstract** 

Artificial intelligence (AI) systems are increasingly being deployed to verify information and even generate news content. We try to  provide a comprehensive literature review on reinforcement learning-driven language agents for multi-domain fact-checking and coherent news synthesis, with an emphasis on implications for journalism and media. We examine how large language models and other AI agents, often refined through reinforcement learning, are used to automate the identification of false claims and to produce news narratives. We find that AI-driven fact-checkers can greatly enhance the speed and scale of verification, and reinforcement learning techniques (including human feedback) have improved the factual accuracy and coherence of generated text (Roit et al., 2023). Case studies from news organizations illustrate that these tools can support human factcheckers by flagging potential errors and synthesizing information across domains. However, challenges persist: current AI fact-checkers show inconsistent accuracy (Quelle & Bovet, 2024) and require human oversight to prevent errors or bias. AI-generated news, while increasingly coherent, may be less comprehensible to readers without editorial refinement (Thäsler-Kordonouri et al., 2024). We discuss theoretical frameworks from journalism studies to contextualize these developments, and we address uncertainties, ethical considerations, and contested perspectives. We conclude that reinforcement learning-enhanced language agents hold significant promise for journalism by augmenting fact-checking efforts and content creation, but must be integrated carefully to uphold journalistic standards of truth and trust. 

_Keywords: Reinforcement learning; language agents; automated fact-checking; news synthesis; computational journalism; media ethics; AI in journalism_ 

## **Introduction** 

The rise of reinforcement learning-driven language agents comes at a time when journalism is grappling with unprecedented challenges of misinformation and an evolving digital news ecosystem. Fact-checking—the process of verifying claims and debunking false information—has become both more critical and more strained as the volume of content outpaces human verification capacity (Graves, 2018; Quelle & Bovet, 2024). Simultaneously, news organizations are experimenting with AI to generate and summarize content, raising questions about quality, coherence, and the role of journalists. We start by asking the following question: _How are reinforcement learning (RL)-driven language models and agents being applied to multi-domain fact-checking and coherent news synthesis, and what are the implications of these technologies for journalism and media practice?_ 

The significance of this question lies in the potential transformation of journalistic workflows. (Voinea, 2021) If AI agents can reliably verify facts across domains (politics, science, health, etc.) and even draft news stories, they could greatly enhance the efficiency of news production and the combating of misinformation. Journalists could be freed from some routine verification tasks and focus on deeper investigative work, potentially improving the quality of information available to the public (Online News Association, 2024). Moreover, in an era where large language models (LLMs) like GPT-4 and Claude are capable of writing plausible text, understanding how to align 

376 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

these models with factual accuracy and journalistic standards is crucial. Recent advances in reinforcement learning from human feedback (RLHF) have been key to aligning LLMs with desired behaviors, including truthfulness and coherence (Ouyang et al., 2022; Roit et al., 2023). Reinforcement learning provides a framework for these AI systems to "learn" from rewards—be it human feedback on correctness or automated signals of factual consistency—beyond what they learn from text prediction alone. 

Our review covers literature predominantly from the last years, reflecting the rapid progress in both AI language modeling and its adoption in journalism. It also incorporates seminal earlier works where relevant to provide foundational context (e.g., early visions of computational journalism and automated fact-checking). Unlike technical surveys that focus on model architectures, our emphasis is on media implications: accuracy and reliability in practice, the reception by journalists and audiences, ethical and theoretical considerations, and case studies of real-world implementations in news settings. By synthesizing findings across computer science, journalism studies, and industry reports, we aim to present a holistic understanding of the state of RL-driven language agents in news and fact-checking, and to highlight both their potential and limitations. 

## **Literature Review** 

## **AI in Automated Fact-Checking: Capabilities and Challenges** 

Automated fact-checking has evolved into a prominent area of research at the intersection of computer science and journalism, driven by the urgent need to curb misinformation. Fact-checking is inherently a multi-domain problem: professional fact-checkers verify claims on topics ranging from politics to science to public health. Early efforts in automated fact-checking (pre-2018) treated it as a pipeline of natural language processing (NLP) tasks, often split into claim detection (identifying check-worthy statements), evidence retrieval (finding relevant sources), and claim verification (assessing veracity based on evidence) (Guo et al., 2022; Vlachos & Riedel, 2014). Traditional systems used classifiers or gradient-boosted models to label claims as true or false, and were usually domain-specific or trained on narrow datasets (e.g., political statements in fact-checking archives). Recent advances, however, leverage large language models and language agents that can perform complex sequences of actions – such as querying databases or searching the web – to verify information. 

**Large Language Models for Fact-Checking:** Modern LLMs like OpenAI’s GPT-4 have shown the ability to reason about factual questions when given sufficient context. They can be prompted with a claim and asked to output a truth assessment with an explanation. Yet, on their own, these models often hallucinate facts or express falsehoods with high fluency (Ji et al., 2023). To address this, researchers have augmented LLMs with retrieval abilities and reinforcement learning based fine-tuning. For instance, OpenAI’s WebGPT project demonstrated that a GPT-3 model fine-tuned to use a web browser (issuing search queries and citing webpages) could produce answers to open-ended questions with factual accuracy on par with human answers (Nakano et al., 2022). WebGPT was trained with human feedback, using reinforcement learning to reward correct and well-sourced answers, which significantly improved factual reliability by encouraging the model to find and quote evidence rather than rely on its internal knowledge. This approach foreshadows how a language agent can operate: by interacting with tools (search engines, databases) in a goal-directed manner to verify claims. Similarly, Quelle and Bovet (2024) evaluated LLM “agents” that autonomously generate search queries, retrieve supporting data, and produce fact-check decisions with explanations. They found that GPT-4, when equipped with relevant contextual information from the web, could correctly verify a majority of claims, outperforming its predecessor GPT-3.5 on accuracy (Quelle & Bovet, 2024). However, performance varied by domain and language – for example, the accuracy on non-English claims was substantially lower until those claims were translated to English. This highlights a multi-domain (and multi-lingual) challenge: AI agents tend to excel in well-represented domains (and languages) but struggle where training data is sparse or specialized knowledge is required. 

**Reinforcement Learning in Fact-Checking Systems:** Reinforcement learning (RL) techniques have been applied to improve the adaptability and explainability of automated fact-checkers. One notable application is in crossdomain or multi-domain fact-checking, where a model trained on one type of content must generalize to another (e.g., from news articles to social media claims). Mosallanezhad et al. (2022) introduce an RL-based domain adaptation method for fake news detection across topics. Their system, REAL-FND, trains a BERT-based classifier on a source domain and then uses a reinforcement learning agent to adjust the model’s internal representations to better fit a target domain, effectively obscuring domain-specific features while preserving domain-general signals (Mosallanezhad et al., 2022). This approach improved accuracy when detecting fake news in a new domain without extensive retraining data, illustrating how RL can help a fact-checking model traverse multiple knowledge domains. Another use of RL is to encourage models to find multi-hop reasoning paths – for example, an agent can be trained to navigate a knowledge graph or a chain of webpages step by step, collecting evidence to support or 

377 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

refute a claim (Shi & Weninger, 2016). By rewarding successful verifications, the agent learns strategies for evidence gathering (such as which sources to trust or how to combine clues), which is particularly useful for claims requiring synthesis of information from different domains. 

Despite these advances, fully automated fact-checking remains a formidable challenge. Surveys of the field emphasize that current AI systems have credibility issues and are not yet able to handle the full fact-checking process end-to-end with the nuance of a human journalist (Nakov et al., 2021). AI can quickly retrieve previously fact-checked claims or obvious contradictions, but it struggles with context-dependent claims, sarcasm, and resolving ambiguities (Nakov et al., 2021). Moreover, claims often require understanding domain-specific context or statistical data that the AI might not possess. In response, the emerging consensus is that AI will serve as assistive technology for human fact-checkers rather than a replacement. (Graves, 2018; Hassan et al., 2017). For example, the German news organization Der Spiegel has prototyped an AI tool to support their internal factchecking department (Online News Association, 2024). In this system, journalists paste an article draft and the AI automatically extracts factual statements (names, dates, figures) and conducts an initial verification by crosschecking those statements against databases and web sources. The tool highlights statements with low confidence and provides links or sources that either corroborate or contradict the claims, essentially acting as a tireless research assistant. Importantly, the flagged items are then passed to human fact-checkers for final verification. Early results indicate improved efficiency – routine checks can be done faster, allowing human checkers to focus on more complex judgments – but also reveal the need for human oversight. The AI sometimes generates false alarms or misses subtleties, and editors have noted that human judgment is crucial to handle misleading context or logical nuances that an AI might not grasp (Online News Association, 2024). 

Another practical consideration in multi-domain fact-checking is multilingual capability. Many developed AI fact-checking tools (and the corpora they trained on) are English-centric. Fact-checkers in smaller language communities find that generative AI tools like ChatGPT underperform in their languages or local contexts (Kahn, 2024). For instance, fact-checkers in Georgia and Norway report that while AI helps with certain tasks (like translating or summarizing content), it can produce incorrect or culturally out-of-context results when dealing with local political claims, partially because the underlying models were not trained sufficiently on those languages or regional data. This suggests that a truly effective multi-domain fact-checking agent must also be multi-lingual or easily adaptable via fine-tuning to new languages and domains. Reinforcement learning could be harnessed here as well: by incorporating feedback or reward signals specific to performance in a new language domain, an agent might progressively improve its local accuracy. However, research in this specific intersection (RL for multilingual fact-checking) is still nascent. 

In summary, the literature indicates that RL-enhanced language agents are making inroads in automating parts of the fact-checking process across domains. These agents can learn search and verification strategies (through human feedback or environmental rewards) and have demonstrated measurable success – for example, reducing the time to find evidence, or achieving around 70% accuracy in labeling claims as true/false in certain benchmark tests. (Caramancion, 2023) Yet, no system is infallible. Accuracy remains inconsistent and usually below human expert performance, especially on nuanced claims. Thus, many researchers advocate a “human-in-the-loop” approach: AI provides preliminary fact-checks or shortlists of suspect claims, and humans make the final determinations (Hassan et al., 2017; Nakov et al., 2021). This approach aligns with reinforcement learning paradigms as well—the AI agent receives corrective feedback from fact-checkers on its suggestions, creating a virtuous cycle of learning. It also reflects an important shift in mindset: rather than replacing journalists, AI agents are viewed as augmenting journalistic verification, consistent with frameworks of augmented journalism (Dörr, 2016). 

## **AI-Generated News and Coherent Synthesis of Information** 

In parallel with fact-checking, AI has been increasingly applied to news content creation and synthesis. Automated journalism (or “robo-journalism”) has been around for over a decade in limited forms—such as algorithms that generate stock market reports or sports recaps from structured data (Clerwall, 2014; Graefe, 2016). What’s new in the last few years is the ability of neural language models to generate long-form textual news articles or summaries in a style that closely mimics human writing. This opens the possibility for AI to draft news stories from source materials, summarize large volumes of information, or even produce narrative news reports on its own. The key requirements for such AI-generated news to be viable are factual accuracy (the story must not fabricate information) and narrative coherence (the story should read logically and fluently from start to finish). Achieving both has proven difficult: language models can be verbose and fluent, but they may stray off- 

378 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

topic or introduce inaccuracies. Here, reinforcement learning-based techniques have been pivotal in improving outcomes. 

**Coherent Text Generation through Reinforcement Learning:** One challenge in generating a news article (especially from multiple sources) is maintaining global coherence – the article should have a clear structure (e.g., headline, lead, body with thematic flow) and not contradict itself. Early research by Bosselut et al. (2018) introduced discourse-aware rewards for RL to guide text generation. They trained a model to produce more logically ordered and less repetitive text by rewarding the model when sentences followed a sensible discourse order. This resulted in notably more coherent long texts compared to standard training. More recently, with transformers and large-scale models, similar ideas have been applied. For example, Yu et al. (2023) proposed multi-objective RL for summarization that optimizes several criteria simultaneously – including relevance, factual consistency, fluency, and coherence. By defining rewards that penalize factual errors (using entailment checks against the source) and incoherence (perhaps using discourse coherence models), they showed that an agent can learn to balance these aspects and produce summaries that are factually faithful as well as well-structured (Roit et al., 2023; Ryu et al., 2024). In news synthesis, this means an AI could merge information from, say, a newswire and a press release into a single, flowing news piece without losing track of what information belongs where. Roit et al. (2023) in particular demonstrated that using a textual entailment-based reward significantly improved the factual consistency of summaries: the RL-trained summarizer was less likely to introduce unsupported details and was rated higher in coherence by human evaluators compared to non-RL baselines. This suggests that reinforcement learning can serve as a corrective layer on top of a generative model, nudging it away from known failure modes (like hallucination or disjointed narrative) and towards more reliable outputs. 

**Fact-Checking as Part of News Generation:** A noteworthy convergence is that fact-checking components are now being integrated into text generation pipelines to ensure accuracy. For instance, some systems perform a secondary check on each sentence generated, using a verifier model or external knowledge, and then use RL to reward the generator for sentences that pass verification (Lee et al., 2020). This effectively embeds a fact-checking heuristic into the news synthesis agent. The result is what one might call fact-checked text generation. Research by Narayan et al. (2018) on controllable generation has explored giving the model “negative examples” (hallucinated outputs) and teaching it through feedback to avoid those. In news writing, models can be trained to prefer generating text that can be linked back to source documents (e.g., quotes or data points from a source) and to avoid adding details not found in any source. (Narayan et al., 2018) This is crucial in multi-document news synthesis, where an agent might read several articles on a developing story and then write a summary: the agent needs to decide which facts are consistent across sources and form a coherent narrative. Without special handling, a naive model might confuse details from different sources or give undue weight to one perspective. RL-based approaches have been used to reward precision and recall of facts from the source set, ensuring that the generated news summary includes the important points (recall) but does not fabricate or distort them (precision) (Mosallanezhad et al., 2022). 

Empirical studies on AI-written news offer insight into the current quality and limitations of these systems. A study by Thäsler-Kordonouri et al. (2024) compared human-written news articles to ones generated by AI (with human post-editing) in terms of reader comprehension. They found that while readers rated the overall structure and flow of AI-generated articles as nearly on par with human-written ones, they found the AI-produced articles significantly less comprehensible on average. The issues cited included the AI’s word choice and handling of numerical data: automated articles sometimes used odd or overly complex phrasing and included too many raw numbers or facts without adequate explanation, making them harder to digest. Notably, all the AI-generated articles in the study had been sub-edited by human editors before readers saw them, yet they still underperformed on clarity. This suggests that achieving true “human-level” quality in news writing is not just about grammatical coherence, but also about editorial judgment – knowing which details to include or omit, and how to present information in context for lay readers. Reinforcement learning can help a model better mimic some of these editorial choices (for example, one could imagine a reward for using simpler language or for compressing statistics into more digestible terms), but encoding the full spectrum of journalistic writing quality into a reward function is complex. The study’s authors concluded that maintaining human involvement in the AI news production process is important not only to catch errors but to ensure readability and appropriate style (Thäsler-Kordonouri et al., 2024). 

Another line of research and practice is emerging around using AI to draft news content which is then finalized by journalists. Several news agencies (like the Associated Press and Reuters) have employed AI to automatically generate short news pieces in domains like finance (quarterly earnings reports) or sports, where the input data is well-structured. Those systems were largely template-based, but newer ones use neural text generation to be 

379 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

more flexible. For example, the Washington Post developed an in-house system (“Heliograf”) that covered local election results by generating thousands of short updates from vote count data (Marconi, 2020). While not using RL per se, these early deployments provided valuable lessons on incorporating AI into newsrooms. One lesson is the importance of coherence and consistency with the news organization's style and standards. This is an area where reinforcement learning can make a difference going forward: a model could be fine-tuned with rewards that capture adherence to a style guide or the inclusion of required contextual information (like always mentioning the source of a statement, which is a journalistic norm). Indeed, OpenAI’s GPT-4 and similar models, through RLHF, have effectively been trained to follow user instructions and produce more helpful and truthful answers than their base versions (Ouyang et al., 2022). That same principle can apply to news: with proper reward design, an AI could learn the “rules” that journalists follow (e.g. double-check names, attribute quotes, avoid unverifiable claims). 

It is equally important to acknowledge the misuse potential of these technologies. The same methods that can be used to enforce truthfulness can be turned on their head. For instance, Mosallanezhad, Shu, and Liu (2020) demonstrated an adversarial reinforcement learning approach to generate synthetic news content that would evade detection as fake. By using an RL agent to control a text generator (GPT-2) with a reward for fooling a fakenews classifier, they were effectively able to produce false news articles that appeared more legitimate and were harder for automated detectors to flag. This adversarial use-case, while research-oriented, highlights a media implication: as news synthesis AI improves, so does the quality of AI-generated disinformation. A coherent, ontopic, stylistically polished fake news article could be weaponized to mislead readers, and it might slip past older detection algorithms that rely on catching linguistic oddities or inconsistencies. This creates an arms race dynamic (Floridi & Chiriatti, 2020) in which journalists and AI developers must constantly update fact-checking and verification tools to counter increasingly sophisticated AI-generated false content. 

In summary, the literature on AI-driven news synthesis shows considerable progress in achieving coherent and factually-aligned text generation, thanks in part to reinforcement learning strategies that optimize multiple aspects of text quality. Systems can now produce fairly readable drafts of news stories, especially for summary or update pieces, and these drafts can save journalists time. However, without human review, these systems are not fully reliable. They might omit important context known to human reporters, or include subtly misleading phrasing. The state-of-the-art still struggles with truly understanding the newsworthiness of information – a human editor’s sense of what leads the story, what background is essential, what the audience already knows, etc. Those higher-order judgments are not easily encapsulated in a reward function or dataset. As a result, most deployments of news-writing AI use them as writing assistants. For example, some newsrooms use generative AI to produce a first version of a news report which journalists then fact-check, edit for tone, and enrich with analysis or quotes (Mei, 2023). When used in this way, AI can be a productivity tool: a survey by the Associated Press in late 2023 found that about 70% of news organizations had experimented with generative AI in some capacity, primarily to streamline workflows (Diakopoulos et al., 2024). These applications include summarizing transcripts, suggesting headlines, or translating articles – tasks adjacent to news synthesis that benefit from the coherence and language capabilities of modern models. Notably, the survey respondents emphasized using AI to augment rather than radically innovate their output, and they flagged the need for robust editorial guidelines to ensure AI use remains ethical and accurate (Diakopoulos et al., 2024). Such findings from industry align with the cautious optimism found in academic circles: AI writing can assist with the heavy lifting of content production, but the human journalist remains central as a truth filter, narrative crafter, and ethical guardian. 

## **Theoretical Framework** 

Situating reinforcement learning-driven language agents within journalism and media practice calls for an interdisciplinary theoretical lens. Several frameworks and theories from communication and journalism studies are relevant to understanding the implications of AI fact-checkers and news writers: 

**Computational Journalism and Sociotechnical Change** : Computational journalism (Flew et al., 2012) provides a broad conceptual umbrella for this study. It refers to the integration of computing technologies into news production and distribution. Under this framework, AI agents are viewed as new actors in the news production process, working alongside journalists and editors. The adoption of RL-driven agents can be seen through the lens of innovation diffusion in newsrooms (Rogers, 2003), where news organizations vary in how quickly and extensively they adopt new technologies. The JournalismAI initiative’s recent surveys  (Beckett & Yaseen, 2023) indicate that larger, resource-rich outlets are among the first to implement AI for editorial tasks, aligning with diffusion theory that suggests organizations with more resources and a culture of innovation will lead in adoption. This framework helps explain current disparities: some newsrooms have AI-assisted workflows, while others remain skeptical or constrained by costs and skills. 

380 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

**Gatekeeping Theory:** Traditional gatekeeping theory (Shoemaker & Vos, 2009) describes how journalists and editors act as gatekeepers, filtering and shaping the information that reaches the public. The introduction of AI agents creates a new form of algorithmic gatekeeping. For fact-checking, an AI might prioritize which claims to flag, thereby influencing which issues get journalistic attention. In news synthesis, if AI is used to draft or select stories, the algorithm's design effectively sets criteria for news selection and presentation. Theoretical work on algorithmic gatekeeping and accountability (Diakopoulos & Koliska, 2017) is pertinent here: it urges transparency in how AI systems make choices that could affect news narratives. The reinforcement learning aspect underscores that these agents are not static – they update their “gatekeeping” behavior based on feedback (rewards). Gatekeeping theory thus extends to consider a hybrid human–AI process: humans still supervise gates, but AI agents enforce some gates automatically (e.g. an AI might automatically block publishing an article that fails a factcheck confidence threshold). 

**Media Trust and Credibility** : The success of AI in journalism partly hinges on audience trust. Relevant here is the body of theory on source credibility (Hovland et al., 1953) and its modern extensions to algorithmic sources. If news consumers perceive information as coming from an AI, how does that affect credibility? The Main Model (Sundar, 2008) in human-computer interaction suggests that people use cognitive heuristics (such as machine "objectivity") when evaluating algorithmically generated content. Empirical research by Chae and Tewksbury (2024) speaks to this: they found that the persuasive effect of fact-checks was not diminished when people were told the fact-check was AI-generated. In some cases, participants saw AI-driven fact-checks as less biased than those written by journalists, particularly if the topic was politically charged (Chae & Tewksbury, 2024). This intersects with the theoretical concept of algorithmic authority (Napoli, 2014) – the idea that algorithms can be perceived as neutral arbiters of truth. However, this trust is double-edged; any high-profile mistakes by AI (e.g., an AI fact-checker incorrectly labeling a true statement as false) could lead to a rapid erosion of public trust not just in the tool, but potentially in the news outlet using it. Thus, theories of technological trust and risk (Slovic, 1993) also apply: the media must manage the risk perceptions by being transparent about AI use and its oversight. 

**Journalistic Roles and Ethics:** From a normative perspective, theories about the role of journalism in society (e.g., the watchdog role, the gatekeeper role, the disseminator vs. interpreter model) need re-examining in light of AI. If an AI agent takes on some watchdog functions (like continuously monitoring official statements for false claims), does this extend the watchdog capacity or dilute it? One theoretical viewpoint is that of “post-industrial journalism” (Anderson et al., 2014), which posits that journalism is becoming more networked and collaborative, involving actors outside traditional newsrooms. AI can be seen as another non-human actor in this network, potentially taking over routine tasks. The ethics of automation in journalism (van Dalen, 2012) form another framework: this raises questions about accountability (if an AI-generated news story is wrong, who is responsible?), transparency (should audiences know when a story is written by AI?), and professional values (can an algorithm adhere to the journalistic code of ethics?). Many news organizations have begun developing AI ethics guidelines that reflect these concerns, often emphasizing that AI output must be verified by humans, and that the provenance of content (human or AI or hybrid) should be disclosed to maintain transparency (BBC, 2021; Diakopoulos, 2019). These evolving norms can be analyzed using actor-network theory as well, where journalists, AI developers, news, and audience form an interconnected system negotiating the norms of news production. 

**Media Effects and Audience Perception** : The introduction of AI in content creation also relates to media effects theories. If AI allows for more personalized news (e.g., automated news tailored to individuals or niche topics), uses and gratifications theory would predict audiences might seek out outlets that use these capabilities to better serve their interests (Sundar & Marathe, 2010). However, if AI-curated or written news leads to information echo chambers or reduces content diversity (because algorithms might optimize for engagement or certain reward structures), then theories about media pluralism and agenda-setting might need to be revisited. For example, could a reinforcement learning agent inadvertently learn to prioritize sensational content if “rewarded” by higher click rates? Journalism theory calls for editorial judgment to maintain balance and public interest focus, so an unbounded RL system might conflict with those values if not carefully guided. Thus, frameworks of agenda setting and framing can be extended: humans traditionally set the news agenda, but AI could subtly influence agendas through the way it surfaces certain facts or writes headlines (e.g., an AI writing a headline might frame an issue more strongly in terms of conflict if that style was rewarded as engaging). 

In light of these frameworks, this paper assumes a perspective grounded in augmented journalism – a concept that journalism is not being handed over to algorithms entirely, but rather that algorithms are augmenting the capabilities of journalists (Lokot & and Diakopoulos, 2016). We treat RL-driven agents as new entrants in the media ecosystem that must be understood both in terms of what they technically can do and how they fit into social practices and norms of journalism. The theoretical framework thus combines insights from media sociology (how 

381 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

newsroom culture and power dynamics affect AI adoption), communication psychology (how humans perceive AI outputs), and information ethics (the responsibility and governance of AI in media). This multifaceted approach is necessary because the topic spans technology and human factors. By anchoring our analysis in these theories, we can better interpret the findings: for example, improvements in fact-checking speed will be discussed not just as technical feats but in terms of their impact on the journalistic watchdog role; similarly, issues of AI errors will be tied to concepts of credibility and accountability in the press. 

## **Methodology** 

This study employs a literature synthesis methodology, reviewing and analyzing existing research rather than collecting new empirical data. The approach is akin to a systematic literature review, but with an emphasis on integrating insights across disciplines (computer science, journalism studies, media ethics) to address our research questions. Below, we outline the steps and criteria used in our methodology: 

We conducted comprehensive searches in academic databases such as ACM Digital Library, IEEE Xplore, Scopus, Web of Science, and Google Scholar for recent scholarly works (2019–2024) related to reinforcement learning, language models, fact-checking, and automated journalism. Key search terms included combinations of “automated fact-checking,” “AI or machine learning fact-checking,” “reinforcement learning AND text generation,” “language model AND news generation,” “AI in journalism,” and “computational journalism AI.” We also identified relevant think-tank and industry reports (for example, Reuters Institute fact sheets, JournalismAI reports, and AP surveys) that provide insight into real-world applications and attitudes. Priority was given to peerreviewed journal articles and conference papers to ensure the inclusion of verified, high-quality research. We included a smaller number of preprints or arXiv papers only if they were by leading groups and addressed very current developments (e.g., cutting-edge techniques not yet formally published). Additionally, we included a few seminal sources outside the 5-year range, such as foundational studies or theoretical works (e.g., Graves, 2018; Diakopoulos, 2019), to frame the current research in context. 

Sources were included if they dealt with: (a) automated or AI-assisted fact-checking, especially those mentioning multi-domain or general-purpose systems; (b) AI generation of news content, summarization of news, or related NLP tasks in a journalistic context; (c) reinforcement learning or human-in-the-loop training methods applied to NLP for truthfulness, coherence, or content quality; and/or (d) discussions of the implications of AI on journalism (e.g., case studies of newsroom adoption, audience perception studies, ethical guidelines). We ensured each included source was an established and verifiable piece of scholarship or reporting. 

From each source, we extracted key findings relevant to our topic. This included quantitative results (e.g., accuracy improvements, reader survey statistics), qualitative insights (e.g., interviews with journalists about using AI, or authors’ discussions of limitations), and any conceptual or theoretical contributions. We took careful notes to accurately represent each source’s conclusions and to capture any cited evidence within them that might be relevant (for example, if a literature review article cited additional studies about AI performance, we traced some of those references as needed). 

We critically evaluated the credibility and limitations of each source. Peer-reviewed studies in top conferences/journals were assumed to be methodologically sound, but we still examined their sample sizes, experimental design, and whether results were statistically significant or more exploratory. For instance, an NLP conference paper might report an increase in a metric; we consider whether that translates into a practically meaningful improvement for journalists. For industry surveys or case studies, we consider potential bias (e.g., a survey by an AI advocacy group might emphasize positive outcomes; a case study from a single newsroom may not generalize to all). We also remain aware of the publication bias in academia where positive results are more likely to be published than negative ones. To counter this, we looked for any “failures” or cautionary tales in the gray literature (like blogs or commentary by practitioners). One example was an anecdote of an AI tool incorrectly altering a quote, which we found in a journalism blog – while not a formal study, it provides insight into types of errors not always quantified in research. 

We acknowledge that our review, while extensive, may not capture absolutely everything in this fast-moving field. New developments in large language models are frequent. We mitigated this by including very recent publications  and by focusing on enduring issues (like truthfulness, coherence, ethical integration) that are likely to remain relevant despite technological leaps. Another limitation is the interdisciplinary nature of the topic – technical papers may not discuss media implications, and media studies papers may not detail the technical underpinnings. In our synthesis, we interpret technical results in light of media contexts, which involves some subjective judgment. We rely on our theoretical framework to guide these interpretations appropriately. 

382 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

## **Analysis/Findings** 

We identify several key findings regarding reinforcement learning-driven language agents in fact-checking and news synthesis, along with their implications for journalism. These findings are organized into subtopics below for clarity. 

## **1. Enhanced Fact-Checking Speed and Scale with AI, but Need for Human Oversight** 

One clear consensus is that AI systems, including RL-enhanced language models, can dramatically speed up the fact-checking process and handle a scale of content that would be impossible for humans alone. Dedicated factcheckers in newsrooms have historically been limited in how many claims they can verify, often focusing only on the highest-profile items. AI agents can scan thousands of statements from news articles, social media, and transcripts to flag potential falsehoods or errors (Corney, 2021). For example, the fact-checking tool at Der Spiegel can analyze an entire article draft in moments, highlighting claims like “Minister X said Y in 2015” and checking those against databases (Online News Association, 2024). This allows human fact-checkers to zero in quickly on problematic statements. In a sense, AI serves as a triage system, sifting through content for anything that looks suspicious or that doesn’t match known records. 

However, a consistent finding is that AI is not replacing human verification; it is augmenting it. The output of AI fact-checkers is typically a probabilistic assessment (e.g., a confidence score or a “likely false” flag) rather than a definitive judgment. Journalists treat these as tips or leads. In Der Spiegel’s case, if the AI flags a sentence with low confidence, a human fact-checker will then manually verify that fact using conventional methods. This aligns with the literature suggesting that full automation of verification is not yet practical for complex claims (Graves, 2018; Hassan et al., 2017). The AI might miss context – for instance, a claim that is literally false but meant sarcastically, or a partially true statement that needs nuance. Human fact-checkers provide the final say, applying contextual knowledge and ethical judgment (e.g., whether to label something as false or just misleading). 

One interesting finding related to oversight is the idea of “accuracy over plausibility” (Online News Association, 2024). Journalists have learned that some AI language models will choose the most plausible-sounding answer to a query, which is not necessarily the correct one. This is a side effect of how models like GPT are trained (to maximize linguistic probability). Der Spiegel’s team explicitly noted the need to customize AI tools toward prioritizing verified accuracy even if the result is an admission of uncertainty, rather than having the AI produce a fluent but potentially incorrect statement. Reinforcement learning can be harnessed to reinforce this: by penalizing the AI when it produces a plausible-sounding claim that contradicts the evidence, developers aim to train models that err on the side of saying “I’m not sure” or asking for human help, rather than confidently outputting an untruth. This “calibration” of AI confidence is critical in news settings. 

In terms of scale, AI fact-checkers shine in monitoring multiple domains and streams of data simultaneously. For example, systems have been developed that continuously ingest feeds from political speeches, Twitter posts, and newswire reports, and use claim detection algorithms to pick out sentences that assert facts (Guo et al., 2022). These are then matched against knowledge bases or prior fact-checks. The reinforcement learning-based domain adaptation techniques like in (Mosallanezhad et al., 2022) mean an AI can fluidly switch between, say, checking a claim about COVID-19 vaccines and a claim about election turnout, without needing a human to reconfigure it for each topic. This is a leap from older systems that were brittle outside their training domain. An attendant finding, however, is that accuracy varies by domain. As noted earlier, Quelle and Bovet (2024) found GPT-4’s fact-checking accuracy was decent on average, but on certain categories of claims (especially ones that require niche expertise, like medical or legal claims), the model’s performance dropped. Thus, while AI can attempt to fact-check anything, its reliability is higher in domains where extensive training data exists or common knowledge suffices, and lower in domains requiring expert interpretation of data or jargon. 

## **2. Improvement in Factual Consistency and Coherence of AI-Generated Text via Reinforcement Learning** 

A major positive finding is that reinforcement learning techniques have tangibly improved the quality of AIgenerated text in terms of factual consistency and logical flow. Before RL-based fine-tuning was introduced, neural text generators often faced a trade-off: if they tried to be factually accurate by sticking close to a source, they might produce dry, snippet-like text; if they tried to be more fluent and narrative, they might inadvertently introduce errors or extraneous info. RL-based approaches, especially those incorporating textual entailment checks or question-answering feedback (Roit et al., 2023), show that it’s possible to have both: models that generate summaries or articles which are both well-formed and faithful to the original information. 

For instance, in Roit et al. (2023)’s study of summarization, the RL-trained model reduced the rate of factual errors in summaries by a significant margin (their evaluation with human annotators showed a clear drop in the 

383 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

number of unsupported statements compared to a baseline summarizer). Moreover, the summaries did not become excessively short or cut-and-pasted; they remained reasonably abstractive and coherent, indicating that the model learned to rephrase and integrate information without “ **hallucinating** ” new facts. This is a notable success because it demonstrates a method to align AI outputs with journalistic requirements of accuracy. In practical terms, a news organization using such a summarizer could trust it more with summarizing a lengthy report or multiple sources into a news piece, saving journalists time on initial drafting. (Stănescu, 2023) 

Coherence is another aspect improved by RL. Bosselut et al. (2018) had shown early on that training a model with a reward for proper sentence order yields more readable stories. More recent work uses RL to enforce structural coherence at larger scales, like paragraph transitions and topical consistency (Celikyilmaz et al., 2021). The analysis of those outputs shows fewer instances of the AI “drifting” mid-article. For example, without special training a model might start an article talking about an economic issue and halfway through switch focus abruptly due to some associative leap in text, confusing the reader. With coherence rewards, the model is better at sticking to the thread. In newsroom experiments, journalists have observed that AI-written drafts are improving in this regard. The New York Times, in an internal trial, noted that an AI summary of a complex investigative piece managed to preserve the article’s narrative arc quite well, something that earlier summarizers (which often produced just bullet-point like abstracts) could not do (Stafford, 2023, personal communication as cited in a JournalismAI report). While not all such observations are formally published, they point to the same trend: RL and human feedback are making AI outputs more news-like in structure. 

Despite these improvements, the findings also underscore that AI-generated content still benefits greatly from human editing. The LMU study (Thäsler-Kordonouri et al., 2024) is instructive. It revealed issues in AI writing that may not be fully solved by RL as currently implemented. For instance, the tendency to include too many numbers – this could be because the AI was trained on data-heavy reports and assumed more numbers increases credibility (a known phenomenon where models latch onto certain features). A reinforcement learning approach could potentially add a penalty for each additional number beyond a threshold to train brevity in statistics, but this fine level of control is not yet common. Human editors naturally smooth these things out, deciding which figures are essential. The study’s recommendation was effectively to incorporate such editorial guidelines into the AI’s process: for example, a rule of thumb like “no more than one figure per sentence unless absolutely necessary” (Thäsler-Kordonouri et al., 2024). This is an area where explicit programming or constraints might complement RL: not everything needs to be learned from scratch if we can imbue models with some simple heuristics. 

Another consistent finding is that transparency and sourcing in AI text require improvement. Fact-checkers and editors often critique AI drafts for lacking clear attributions (WHO said this? Where is this information from?). While a human journalist naturally cites sources or includes quotes, an AI may output a narrative devoid of attributions unless prompted to do so. Some RL research has tried to tackle this by rewarding models for including references (e.g., WebGPT was trained to quote URLs as sources), but in newswriting, a more nuanced approach is needed (e.g., “according to the Ministry of Health, ...”). If the AI is connected to a database of sources, it could be trained to pull in the correct attribution. This is still an open challenge. News agencies like the AP have started requiring that any AI-generated content be accompanied by a log of sources consulted, which an editor can then use to verify quotes and data (Diakopoulos et al., 2024). In effect, the news industry is pushing for AI that doesn’t just write text, but also outputs its evidence. Encouragingly, language models with chain-of-thought reasoning and tool use (like browsing) are steps in that direction, as they inherently involve showing which information led to the final answer (Nakano et al., 2022). Over time, this could make AI-written news more transparent and thus easier to fact-check in itself. 

## **3. Journalistic Adoption: Use Cases, Perceptions, and Challenges** 

The analysis of case studies and surveys reveals a nuanced picture of how these AI agents are being adopted in newsrooms and perceived by journalists and audiences. A key finding is that news organizations are actively experimenting with AI for both fact-checking and content generation, but they are doing so in cautious, incremental ways. According to a global JournalismAI survey  (Beckett & Yaseen, 2023), many news organizations have set up dedicated teams or labs to explore AI, often starting with low-risk tasks (like archiving content, transcribing audio, or summarizing long reports for internal use). Fact-checking assistance and automated news briefs are commonly cited as pilot projects. 

The Associated Press survey (Diakopoulos et al., 2024) highlighted that about 70% of surveyed newsrooms used generative AI in some capacity by late 2023, which is substantial. However, the majority were using it to enhance existing workflows rather than to create wholly new products. This confirms what our literature review suggested: AI is mainly being used to augment journalists. For example, Reuters has an internal system that uses AI to pull facts from financial reports and draft short news pieces which journalists then finalize (Carpenter, 2023). 

384 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

In such a setting, journalists appreciate the time saved on rote tasks (like parsing tables for key figures), but they remain in control of the final copy. 

Another finding from industry is the emphasis on training and governance around AI tools. Editors and news executives recognize that if staff are to rely on AI outputs, they must know the limitations of these tools (Meir, 2023). Some newsrooms have instituted training sessions to educate journalists on how AI models work, why they might fail, and how to double-check AI-provided information. There’s a growing call for editorial guidelines specific to AI. For instance, the BBC’s editorial policy now includes a section that any content fully or partially generated by AI must be verified for accuracy and approved by a senior editor before publication (BBC, 2021). These organizational measures align with academic recommendations for maintaining human accountability (Diakopoulos, 2019). 

Journalist perceptions are mixed but evolving. In the “Challenges of Automating Fact-Checking” case study (Kavtaradze, 2024), journalists expressed initial skepticism towards an AI fact-checking tool, worrying it might not understand context or could be gamed by savvy political actors. Over time, as they tested the prototype, they grew more confident in its ability to catch outright falsehoods in texts, and they started to see it as an ally that could point them toward things requiring attention (Kavtaradze, 2024). This reflects a pattern: direct experience with AI tools can increase trust in them, as long as the tools demonstrate reliability. Importantly, management of expectations is crucial. When journalists treat the AI as a fallible assistant, they are pleasantly surprised by its useful catches; if they expected a magic truth machine, they would be disappointed. Hence, framing and internal communication about what the AI agent can and cannot do is a key factor in successful adoption. 

Regarding audience perception, the findings are cautiously optimistic. The experiment by Chae and Tewksbury (2024) suggests that audiences do not inherently reject information just because an AI was involved in producing it. In their study, fact-check readers who were told that the check was “AI-assisted” found it just as persuasive as those told it was written by a human. There was even an intriguing reduction in perceived political bias among some readers when AI was involved – possibly due to a perception that a machine has no political agenda (Chae & Tewksbury, 2024). This indicates a potential benefit: AI-backed fact-checks might sidestep some partisan skepticism that human fact-checkers face. However, it’s important not to overgeneralize this. The context was carefully controlled; in the messy real world, if an AI made a glaring error, it could quickly become a news story in itself (“AI fact-checker gets it wrong”), which would harm credibility. So while audiences might accept AI assistance, their goodwill likely hinges on an expectation that the news organization is still ensuring quality control. 

The discussion of audience brings up the issue of transparency to the public. Some media ethicists argue that whenever AI has a material role in content creation, it should be disclosed to readers. Current practice is not uniform: some outlets add notes like “This article was auto-generated from data and then edited by our reporters,” whereas others do not explicitly state the use of AI if a human journalist has edited the final version. The literature suggests that transparency can bolster trust (if done in a non-alarming way). One model might be to have a dedicated section on the website explaining the newsroom’s AI uses and safeguards, rather than noting it on each piece and potentially distracting or confusing readers. This is still being debated in the industry, and further research on audience attitudes would be beneficial. 

A challenge repeatedly noted is bias and fairness. If a reinforcement learning agent’s reward model is misspecified or its training data biased, it might produce content that reflects those biases. For example, a factchecking AI might, if trained predominantly on political claims of one type, become more sensitive to certain phrases that correlate with falsehood in that set, possibly over-flagging statements from one group versus another. This could introduce a perception (or reality) of partisan bias in an ostensibly neutral tool. There is an active area of research on debiasing AI models, but journalists and editors have rightly pointed out that they need to be vigilant. One finding in the AP survey (Diakopoulos et al., 2024) was that a large portion of respondents stressed the need for diverse development teams and oversight to ensure AI tools don’t inadvertently disadvantage certain topics or communities. This human oversight is essentially an ethical check beyond what the AI’s reward function encapsulated. 

Finally, an important finding is that the introduction of AI agents is changing newsroom roles and skills. Some news organizations now hire computational journalists or AI specialists to work alongside editors. The Wall Street Journal, for instance, created a role for an “AI Editor” whose job is to evaluate and implement AI tools and also to be a liaison between technologists and editorial staff. This development resonates with field theory in journalism (Benson & Neveu, 2005) – the journalistic field is adapting by adding new forms of capital (tech expertise) and new positions to preserve its core function under changed circumstances. There’s also an impetus for upskilling existing journalists: training them not to code from scratch necessarily, but to be proficient in using AI tools (e.g., writing effective prompts for LLM-based assistants, or interpreting the output of a claim verification system). Thus, 

385 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

we find a co-evolution of technology and practice: newsrooms that effectively integrate these agents often do so by evolving their practices and norms, rather than trying to plug AI in while everything else stays static. 

In conclusion of this analysis section, the interplay of reinforcement learning-driven AI agents and journalism is yielding productive outcomes – faster fact-checking, more coherent automated writing, and extended capabilities – yet it is also surfacing critical requirements: robust human oversight, clear ethical guidelines, transparency, and ongoing adjustments in newsroom culture and skills. The next section will discuss the broader implications of these findings, addressing what they mean for the future of journalism and how to navigate the uncertainties and debates that accompany this AI integration. 

## **Discussion** 

The findings from our literature review and analysis illustrate both the promise and the complexity of deploying reinforcement learning-driven language agents in journalism. In this discussion, we delve into the implications of these findings for journalistic practice, industry norms, and society at large. We also confront the uncertainties and contested perspectives that arise, and outline the limitations of current approaches – all of which are crucial for charting a responsible path forward. 

## **Implications for Journalistic Practice and Newsroom Workflow** 

Perhaps the most immediate impact of RL-driven AI agents is on the workflow of journalists. Fact-checkers and reporters now have the potential to work with AI “colleagues” that handle labor-intensive tasks like initial evidence gathering, transcription, or summary writing. This symbiosis, when functioning well, can enhance productivity and thoroughness. For example, a journalist covering a breaking news story could use an AI assistant to quickly compile key facts from various sources (tweets, press statements, past articles) into a coherent brief, which the journalist can then expand and refine. In investigative projects, AI tools can surface patterns or inconsistencies in large document dumps that a human might miss. These advantages underscore a likely shift in journalistic labor: more focus on verification, interpretation, and added value, and less on rote information processing. Essentially, journalists can concentrate on what humans uniquely do best – providing context, analysis, moral reasoning, and narrative craft – while AI handles routine groundwork. 

This shift, however, requires new skills and training. Journalists will need to be adept at working with AI outputs critically. The role of a fact-checker, for instance, might evolve from manually checking each fact to reviewing and cross-verifying the suggestions of an AI system (Graves & Anderson, 2020). This is not necessarily easier – it demands digital literacy, skepticism, and sometimes even basic understanding of how the AI works to judge when it might be wrong. News organizations that have embraced these tools often find they must invest in training sessions, as noted earlier. Some have developed internal guidelines or checklists (e.g., “ _If the AI flags a fact, always verify it in at least one independent source before altering the article_ ”) to institutionalize best practices. 

A significant implication is on the speed of news cycles and fact-checking cycles. With AI assistance, fact-checks that might have taken a day of research could potentially be turned around in hours, helping debunk false claims before they spread too widely. This proactive speed is crucial in the fight against misinformation, where timing can make the difference between a falsehood going viral or being contained. Additionally, news updates can be published faster as AI can draft routine elements. However, this speed comes with a caution: the temptation to publish quickly must be balanced with the diligence of review. The motto “get it first, but first get it right” still holds. If AI allows getting it first more often, newsrooms must double down on “get it right” through verification protocols. 

**Redefining Roles: T** he introduction of RL agents in journalism might give rise to hybrid roles – for instance, an editor who specializes in overseeing AI-generated content or a reporter who primarily works on stories initiated by AI detections (like fact-checking claims that an AI system surfaced from social media). These roles blur the traditional distinctions but could be essential in maximizing the technology’s benefits. There may also be a revaluation of certain journalistic skills: verification and curation might gain more prominence relative to writing from scratch, since writing (at least first drafts) can be outsourced to some degree. This could affect journalism education, with curriculums incorporating data science or AI literacy alongside classic reporting skills (Beckett, 2019) 

## **Quality, Accuracy, and Ethical Standards: Maintaining Trust** 

A major theme in discussions is maintaining journalistic quality and accuracy in the age of AI. The goal is not just to do things faster, but also to uphold or even improve the standards of truth and clarity that audiences expect 

386 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

from reputable media. The findings show that reinforcement learning can help align AI outputs with factual accuracy, but errors are still very possible. Thus, one implication is that news organizations must establish robust editorial oversight for AI contributions. Some have likened AI tools to junior reporters or interns – useful, but in need of supervision. This perspective can be formally codified: for instance, an editorial policy might state that “ _AI-generated copy must be reviewed by at least one editor and one subject-matter expert before publication_ ” in sensitive domains. 

From an ethical standpoint, numerous questions arise. Journalistic codes of ethics (like those from SPJ or the BBC) emphasize accuracy, accountability, and transparency. These apply as much to AI as to human journalists. Therefore, if an AI agent contributes to a story, the news organization is ethically accountable for its content. There is a growing view that algorithms should not be scapegoats; if an AI makes a mistake, the organization should acknowledge it as it would a human error, issue corrections, and explain in broad terms what went wrong (Diakopoulos & Johnson, 2019). This level of transparency might actually increase trust – showing that the media is not blindly using AI but actively managing it. 

The ethical use of AI in journalism also means addressing biases. One implication is the need for continuous auditing of AI systems for bias and fairness. News organizations might consider partnerships with academic researchers or independent auditors to periodically evaluate, say, whether their fact-checking AI flags one political group more often than another without justification. If biases are found, they need to recalibrate the models or adjust the training data. In reinforcement learning terms, this could mean altering the reward structure or adding new training examples to correct the bias. 

Transparency to Audiences is another ethical dimension. While, as noted, audience reactions to AI involvement have been relatively neutral or even positive in controlled settings (Chae & Tewksbury, 2024), it’s still a subject of debate how much the end-user needs to know. Many experts argue that concealing AI involvement is shortsighted – if discovered, it could lead to a backlash (“they were passing off robot-written stories as humanwritten!”). A prudent approach is for news outlets to be open about their AI usage, perhaps through occasional explainer articles, disclosures in mastheads, or interactive features that let curious readers see which part of a story was machine-assisted. For example, an online article might have a note like, “Portions of this report were generated with the assistance of an AI tool and verified by our editors.” This way, transparency is provided without sensationalizing the fact or undermining the content’s credibility. 

Trust is the currency of news media, and both fact-checking and coherent storytelling are central to building trust. If RL-driven agents help catch errors and prevent the publication of false information, they directly contribute to trust. On the other hand, any high-profile failure could be a trust setback. This puts a premium on testing and incremental deployment. The implication for practice is to sandbox AI tools, test them internally, perhaps publish low-stakes content with them first (like internal newsletters or wire stories that can be doublechecked against original sources), and only then integrate them into high-stakes reporting. This was the approach of organizations like The Washington Post and Bloomberg, which quietly tested their AI writing systems on years of historical data and small current pieces before any public launch. (Dörr, 2016) 

## **Addressing Uncertainties and Contested Perspectives** 

While there is excitement about AI in journalism, there are also uncertainties and disagreements about how far it should go. One contested issue is the degree of autonomy these agents should have. Some technologists foresee AI eventually handling straightforward news stories end-to-end (e.g., routine finance reports, sports recaps) with minimal human intervention. Indeed, companies like Automated Insights have provided services that do exactly that with template-based systems. Now with LLMs, that autonomy could extend to more complex writing. However, many journalists are uneasy with fully autonomous news generation, arguing that even simple stories can have nuances that an AI might miss (Carlson, 2015). Our literature reflects this tension: the gains in performance are real, but not yet at a level where one would be comfortable removing the human check entirely. 

Another source of debate is algorithmic objectivity. There’s a narrative that AI could make journalism more objective or fact-based, removing human subjectivity and error. On the flip side, scholars like Meredith Broussard (Broussard, 2019) warn against “techno-chauvinism,” the assumption that computers are inherently superior decision-makers. They argue that human judgment is essential in journalism, an art as much as a science, and that things like news values or the significance of a story are not easily quantified for an AI. The findings in our review suggest a middle ground – AI can handle facts and even suggest angles, but the framing and contextualization (i.e., why a story matters) still lean heavily on human insight. This is contested by those who believe AI could even analyze audience data to determine which angles would resonate most, essentially “deciding” the news angle. Such algorithm-driven framing wades into ethically fraught territory (could lead to pandering or reinforcing 

387 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

biases), so most mainstream outlets have steered clear of handing that kind of power to AI. For now, AI suggests, humans dispose. 

Uncertainty about future capabilities also looms. As reinforcement learning and AI models progress, some concerns might be resolved (e.g., more fact-proof models), but new ones will arise (e.g., deepfake text that’s indistinguishable from human and loaded with subtle propaganda). Newsrooms will need to stay agile. One implication is the necessity for ongoing research collaboration between media and AI researchers. Instead of reacting to technology once it’s mainstream, media organizations might want to be involved in its development – setting use-cases, providing feedback on prototypes, or even co-creating models that are journalism-optimized (as opposed to general-purpose). This proactive stance could influence AI development in directions beneficial for society (like prioritizing truthfulness and explainability). 

It’s also worth noting the economic and job implications, which are sometimes contested. Automation traditionally raises the fear of job loss. In journalism, the fear is that if AI can write articles, some reporters or copy editors might be deemed redundant by cost-cutting management. The literature doesn’t provide a clear answer; it’s too early to tell how this plays out. There are anecdotes of reporters being freed to do more important work, but also speculation that local newsrooms might use AI to compensate for staff they can’t afford. Ethically, using AI to fill gaps in local news (where news deserts exist) could be seen as a positive if the alternative is no coverage at all. However, there’s an issue of quality – will a small local outlet with limited oversight be able to ensure the AI’s output is accurate and nuanced? This could result in subpar news which might misinform or disengage local communities. The future likely holds a mix: AI taking over some tasks, while human roles shift, possibly with fewer entry-level writing jobs but perhaps more roles in overseeing AI or focusing on unique content that AI cannot do. Ensuring that this transition doesn’t erode the quality of journalism is a shared responsibility of news organizations and those developing the AI tools. 

## **Recommendations and Future Directions** 

From the above implications, we can derive some broad recommendations: 

**Maintain Human-in-the-Loop Processes:** At the current stage and for the foreseeable future, keeping journalists in supervisory roles over AI is critical. Total automation of editorial decisions is neither wise nor necessary. RL agents should be deployed as assistants whose work is always reviewed. This ensures accountability and leverages the strengths of both AI (speed, scale) and humans (judgment, empathy). 

**Develop and Adhere to AI Usage Guidelines** : News organizations should craft clear internal guidelines about how AI will be used in reporting and fact-checking. This could cover when to disclose AI usage, how to verify AI outputs, handling of errors, and respecting privacy and copyright (e.g., not having AI inadvertently plagiarize or generate defamatory content). Already, groups like the Associated Press have drafted AI principles emphasizing accuracy, fairness, and accountability, which can serve as models (AP, 2023). 

**Invest in AI Literacy for Journalists:** Just as data journalism became a valued skill, understanding AI should become part of a journalist’s toolkit. This doesn’t mean all journalists must code, but they should know conceptually how, say, an RLHF-trained model works, and what its blind spots might be. This education can reduce friction and fear, replacing it with informed skepticism and creative thinking about how to best use these tools. 

**Cross-sector Collaboration** : The complexity of ensuring factual, ethical AI usage suggests multi-disciplinary collaboration. Journalists, AI engineers, ethicists, and legal experts should be in conversation. For instance, to tackle deepfakes and AI misinformation, newsrooms might collaborate with AI researchers to develop detection methods, or join consortia that set standards for content authentication (like watermarks or metadata that indicate origin). 

**Monitor Impact on Public Trust** : As AI becomes more integrated, it will be important to measure how it affects public trust in news. This could be done through audience surveys, focus groups, or analyzing engagement metrics. If trust issues are detected (for example, if audiences start doubting content due to knowledge of AI involvement), strategies must be adapted—perhaps more transparency or improved performance to reassure the public. 

In reflecting on the broader societal context, reinforcement learning-driven agents in media have a dual capacity: they can be tools for truth or vectors for falsehood. Journalism’s mission in democracy is to inform accurately and hold power accountable. When wielded responsibly, these AI agents can strengthen that mission by catching more errors, covering more ground, and even detecting disinformation patterns that humans alone might overlook. However, if misused or implemented without care, they could also become sources of misinformation (through mistakes) or be co-opted to generate propaganda (as the adversarial fake news generation research warns). Hence, the deployment of such AI in media is not just a technical matter but a 

388 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

profoundly ethical one. The entire enterprise of using RL in journalism should be guided by the core values of the field: truth, verification, independence, fairness, and accountability. 

## **Limitations of Current Approaches and Future Research** 

While our review highlights many advances, it’s clear that current RL-driven language agents are not a panacea. Some limitations we identified include: difficulty in handling context and nuance, vulnerability to adversarial inputs (e.g., someone intentionally phrasing a claim to fool the AI), and the problem of the AI’s explanations (an agent might give an answer but not be able to explain its reasoning in a way a person can easily understand). Future research is needed to address these gaps. For example, one promising area is explainable AI for journalism – creating models that can not only check a fact but also present the evidence and reasoning in a transparent way, almost like showing its “workings” as a human fact-checker would in an article. This could build trust and also help journalists follow the AI’s logic. 

Another area is refining reward functions to even more closely align with journalistic quality metrics. Right now, proxies like factual entailment or ROUGE scores are used. In the future, one could imagine multi-modal rewards including human feedback not just on correctness but on things like clarity, engagement, or adherence to editorial style. That would effectively encode more of a journalist’s decision process into the training of the agent. 

Finally, longitudinal studies on the outcomes of AI integration in newsrooms would be valuable. Do these tools measurably reduce errors in published content over time? Do they allow for more investigative stories because daily churn is partly automated? Or conversely, do they introduce subtle homogenization in reporting because many outlets might use similar AI tools and thus produce similarly worded stories? These are empirical questions that will need real-world data to answer. 

In concluding this discussion, it’s evident that reinforcement learning-driven language agents present a classic example of a dual-use technology in media: one that can significantly aid the pursuit of truth, but also one that must be managed judiciously to prevent unintended harm. The media implications extend to trust in institutions, the efficiency of truth-telling, and the very nature of news creation. The next few years will likely be defining in how these tools are normalized and regulated in journalism. By engaging with the technology through the critical and ethical lens that journalism demands, media professionals can hopefully ensure that the net effect of these AI agents is to strengthen the Fourth Estate in serving the public interest. 

## **Conclusion** 

The intersection of reinforcement learning-driven AI agents with fact-checking and news synthesis represents a frontier of both great opportunity and responsibility for the journalism industry. This comprehensive review has highlighted how such agents – from large language models fine-tuned with human feedback to specialized factchecking bots – are increasingly capable of operating across multiple knowledge domains and assisting in the verification and generation of news content. Key takeaways can be summarized as follows: 

AI language agents have demonstrated significant potential in scaling up fact-checking efforts and producing coherent narrative text. Through reinforcement learning techniques like RLHF and tailored reward models, these agents have improved in factual accuracy and coherence, making them valuable tools for tasks such as scanning content for inaccuracies, summarizing complex information, or drafting routine news stories. In controlled evaluations, systems like GPT-4 with retrieval support have reached performance levels approaching those of human fact-checkers on certain benchmarks, and RL-optimized summarizers produce more faithful and concise news summaries than ever before (Quelle & Bovet, 2024; Roit et al., 2023). 

In practice, the integration of these AI agents is reshaping newsroom workflows. Fact-checkers at leading outlets use AI to augment their manual verification, effectively acting as force-multipliers that allow journalists to vet more information in less time (Online News Association, 2024). News writers and editors employ AI for drafting and research assistance, which can free up time for more analytical or investigative work. When used prudently, these tools can enhance the depth and speed of journalism – for instance, enabling near real-time fact-checks during live events, or providing quick background research when a story breaks. Importantly, case studies show that these benefits materialize best under a collaborative model: AI handling the heavy lifting of information processing, and humans providing oversight, context, and final judgment. 

Despite technological advances, current AI agents are not infallible. They exhibit inconsistent accuracy, especially outside well-trodden domains or in non-English contexts (Quelle & Bovet, 2024). They are prone to hallucinations (fabricating convincing-sounding details) if not properly constrained, and they can struggle with the subtleties of context, tone, and significance that humans grasp intuitively. AI-generated news text, while 

389 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

structurally and grammatically sound, may still require human polishing to meet readability and style expectations (Thäsler-Kordonouri et al., 2024). Furthermore, the possibility of AI systems inadvertently reflecting or amplifying biases in their training data is a real concern, necessitating vigilant oversight. 

The deployment of RL-driven agents in journalism must reconcile with core media ethics and theories. Issues of transparency, accountability, and bias mitigation are at the forefront. The research reviewed indicates that audiences can accept AI-assisted journalism so long as the integrity and credibility of the output remain high (Chae & Tewksbury, 2024). Theoretical frameworks like computational journalism and gatekeeping have adapted to include algorithms as part of the news production process, without abdicating the ultimate editorial responsibility of humans. In essence, these AI agents should be viewed as extensions of journalistic tools – powerful but requiring guidance. Maintaining the trust of the public will depend on news organizations being open about how AI is used and ensuring that its use serves the public interest (e.g., catching more errors, covering more stories), rather than purely expediency or cost-cutting. 

Looking ahead, the role of reinforcement learning and AI in journalism is likely to expand. As models become more adept and possibly acquire better reasoning through techniques like chain-of-thought prompting and knowledge integration, we may see AI taking on more sophisticated fact-checking tasks – for example, evaluating the logical consistency of political speeches or detecting coordinated misinformation campaigns. In news synthesis, AI might move from drafting basic stories to offering insightful suggestions (akin to a data analyst) on what angles or correlations to explore in a dataset. However, fully autonomous journalism by AI remains a distant and arguably undesirable scenario; the consensus in the literature is that the best outcomes arise from human-AI collaboration, not AI in isolation. 

To maximize benefits and minimize risks, we recommend ongoing interdisciplinary collaboration. Journalists, technologists, and academic researchers should continue to learn from each other. News organizations might participate in the development of journalism-specific AI models that bake in values of truthfulness and public service at the core (rather than retrofitting general models). Likewise, AI researchers can benefit from understanding the rigorous standards and edge cases that journalists deal with, guiding the creation of more robust and explainable systems. 

In conclusion, reinforcement learning-driven language agents have proven to be a double-edged sword that the journalism field must skillfully wield. On one edge, they offer unprecedented capabilities to strengthen factbased reporting – enabling faster verification, broader coverage, and potentially reducing the spread of falsehoods. On the other edge, they pose challenges around error propagation, ethical use, and the need for new skill sets among media professionals. The scholarly and industry dialogue over the past years indicates a maturing approach: embracing the technology’s potential while instituting checks, balances, and cultural shifts to retain what is fundamentally important about journalism. As this technology continues to evolve, so too will the norms and practices around it. The enduring goal is clear and firmly rooted in journalism’s mission – to inform the public accurately and effectively. Reinforcement learning-driven agents, if guided by this mission, can be powerful allies in sustaining a well-informed society in the digital age. 

## **References** 

_AMCC | Generative AI in Journalism: The Evolution of Newswork and Ethics in a Generative Information Ecosystem_ . (n.d.). AMCC. Retrieved April 6, 2025, from https://a-mcc.eu/en/library/studies-and-reports/generative-aiin-journalism-the-evolution-of-newswork-and-ethics-in-a-generative-information-ecosystem/ Anderson, C. W., Bell, E. J., & Shirky, C. (2014). _Post Industrial Journalism: Adapting to the Present_ . https://doi.org/10.7916/D8N01JS7 BBC. (2021). _Responsible AI at the BBC: Our machine learning engine principles_ . https://downloads.bbc.co.uk/rd/pubs/MLEP_Doc_2.1.pdf Beckett, C. (2019). _New Powers, New Responsibilities – A global survey of journalism and artificial intelligence_ . POLIS / LSE, Google News Initiative. https://drive.google.com/file/d/1nf7dquDx6BxsXD3VcF_= J_WzNeoYujN_/view?usp embed_facebook Beckett, C., & Yaseen, M. (2023). _Generating Change A global survey of what news organisations are doing with AI_ . POLIS / LSE, Google News Initiative. https://static1.squarespace.com/static/64d60527c01ae7106f2646e9/t/656e400a1c23e22da0681e46/1701 724190867/Generating+Change+_+The+Journalism+AI+report+_+English.pdf Benson, R., & Neveu, E. (2005). _Bourdieu and the Journalistic Field_ . Wiley.Com. https://www.wiley.com/enus/Bourdieu+and+the+Journalistic+Field-p-9780745633862 

390 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

Bosselut, A., Celikyilmaz, A., He, X., Gao, J., Huang, P.-S., & Choi, Y. (2018). Discourse-Aware Neural Rewards for Coherent Text Generation. In M. Walker, H. Ji, & A. Stent (Eds.), _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)_ (pp. 173–184). Association for Computational Linguistics. https://doi.org/10.18653/v1/N18-1016 

- Broussard, M. (2019). _Artificial Unintelligence: How Computers Misunderstand the World_ (Illustrated edition). The MIT Press. 

Caramancion, K. M. (2023). _News Verifiers Showdown: A Comparative Performance Evaluation of ChatGPT 3.5, ChatGPT 4.0, Bing AI, and Bard in News Fact-Checking_ (arXiv:2306.17176). arXiv. https://doi.org/10.48550/arXiv.2306.17176 Carlson, M. (2015). The Robotic Reporter. _Digital Journalism_ , _3_ (3), 416–431. https://doi.org/10.1080/21670811.2014.976412 

Carpenter, H. (2023). _How AI helps power trusted news at Reuters_ . https://reutersagency.com/resources/how-aihelps-power-trusted-news-at-reuters 

- Celikyilmaz, A., Clark, E., & Gao, J. (2021). _Evaluation of Text Generation: A Survey_ (arXiv:2006.14799). arXiv. https://doi.org/10.48550/arXiv.2006.14799 

Chae, J. H., & Tewksbury, D. (2024). Perceiving AI intervention does not compromise the persuasive effect of factchecking. _New Media & Society_ , 14614448241286881. https://doi.org/10.1177/14614448241286881 Clerwall, C. (2014). Enter the Robot Journalist. _Journalism Practice_ , _8_ (5), 519–531. https://doi.org/10.1080/17512786.2014.883116 

- Corney, D. (2021). _How does Automated Fact Checking work?_ Full Fact. https://fullfact.org/blog/2021/jul/howdoes-automated-fact-checking-work/ 

- Diakopoulos, N. (2019). _Automating the News: How Algorithms Are Rewriting the Media_ . Harvard University Press. https://www.jstor.org/stable/j.ctv24w634d 

- Diakopoulos, N., Cools, H., Li, C., Helberger, N., Kung, E., Rinehart, A., & Gibbs, L. (2024). _Generative AI in Journalism: The Evolution of Newswork and Ethics in a Generative Information Ecosystem_ . https://doi.org/10.13140/RG.2.2.31540.05765 

- Diakopoulos, N., & Johnson, D. (2019). _Anticipating and Addressing the Ethical Implications of Deepfakes in the Context of Elections_ (SSRN Scholarly Paper 3474183). Social Science Research Network. https://doi.org/10.2139/ssrn.3474183 

- Diakopoulos, N., & Koliska, M. (2017). Algorithmic Transparency in the News Media. _Digital Journalism_ , _5_ (7), 809– 828. https://doi.org/10.1080/21670811.2016.1208053 

Dörr, K. N. (2016). Mapping the field of Algorithmic Journalism. _Digital Journalism_ . https://www.tandfonline.com/doi/abs/10.1080/21670811.2015.1096748 

- Flew, T., Spurgeon ,Christina, Daniel ,Anna, & and Swift, A. (2012). The Promise of Computational Journalism. _Journalism Practice_ , _6_ (2), 157–171. https://doi.org/10.1080/17512786.2011.616655 

- Floridi, L., & Chiriatti, M. (2020). GPT-3: Its Nature, Scope, Limits, and Consequences. _Minds and Machines_ , _30_ (4), 681–694. https://doi.org/10.1007/s11023-020-09548-1 

Graefe, A. (2016). _Guide to Automated Journalism_ . Columbia Journalism School. 

Graefe, A., & Bohlken, N. (2020). Automated Journalism: A Meta-Analysis of Readers’ Perceptions of HumanWritten in Comparison to Automated News. _Media and Communication_ , _8_ (3), 50–59. https://doi.org/10.17645/mac.v8i3.3019 

- Graves, L. (2018). _Understanding the promise and limits of automated fact-checking_ . Reuters Institute for the Study of Journalism. https://doi.org/10.60625/RISJ-NQNX-BG89 

- Graves, L., & Anderson, C. W. (2020). Discipline and promote: Building infrastructure and managing algorithms in a “structured journalism” project by professional fact-checking groups. _New Media and Society_ , _22_ (2), Article 2. https://doi.org/10/Discipline%20and%20Promote%20Figures.pdf 

- Guo, Z., Schlichtkrull, M., & Vlachos, A. (2022). _A Survey on Automated Fact-Checking_ (arXiv:2108.11896). arXiv. https://doi.org/10.48550/arXiv.2108.11896 

- Hassan, N., Arslan, F., Li, C., & Tremayne, M. (2017). Toward Automated Fact-Checking: Detecting Check-worthy Factual Claims by ClaimBuster. _Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining_ , 1803–1812. https://doi.org/10.1145/3097983.3098131 

- Hovland, C. I., Janis, I. L., & Kelley, H. H. (1953). _Communication and persuasion; psychological studies of opinion change_ (pp. xii, 315). Yale University Press. 

391 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

- Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Chen, D., Dai, W., Chan, H. S., Madotto, A., & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. _ACM Computing Surveys_ , _55_ (12), 1–38. https://doi.org/10.1145/3571730 

- Kahn, G. (2024). How Generative AI Is Helping Fact-Checkers Flag Election Disinformation, But Is Less Useful in the Global South. _Global Investigative Journalism Network_ . https://gijn.org/stories/how-generative-ai-helpsfact-checkers/ 

- Kavtaradze, L. (2024). Challenges of Automating Fact-Checking: A Technographic Case Study. _Emerging Media_ , _2_ (2), 236–258. https://doi.org/10.1177/27523543241280195 

- Lee, N., Li, B. Z., Wang, S., Yih, W., Ma, H., & Khabsa, M. (2020). _Language Models as Fact Checkers?_ (arXiv:2006.04102). arXiv. https://doi.org/10.48550/arXiv.2006.04102 

- Lokot, T., & and Diakopoulos, N. (2016). News Bots: Automating news and information dissemination on Twitter. _Digital Journalism_ , _4_ (6), 682–699. https://doi.org/10.1080/21670811.2015.1081822 

- Marconi, F. (2020). _Newsmakers: Artificial Intelligence and the Future of Journalism_ . Columbia University Press. 

- Mei, S. (2023). _Analysis: Could AI replace humans in journalism? | Al Jazeera Media Institute_ . http://institute.aljazeera.net/en/ajr/article/2263 

- Meir, N. (2023). _AP: Standards around generative AI_ . The Associated Press. https://www.ap.org/the-definitivesource/behind-the-news/standards-around-generative-ai/ 

- Mosallanezhad, A., Karami, M., Shu, K., Mancenido, M. V., & Liu, H. (2022). Domain Adaptive Fake News Detection via Reinforcement Learning. _Proceedings of the ACM Web Conference 2022_ , 3632–3640. https://doi.org/10.1145/3485447.3512258 

- Mosallanezhad, A., Shu, K., & Liu, H. (2020). _Topic-Preserving Synthetic News Generation: An Adversarial Deep Reinforcement Learning Approach_ (arXiv:2010.16324). arXiv. https://doi.org/10.48550/arXiv.2010.16324 

- Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2022). _WebGPT: Browser-assisted question-answering with human feedback_ (arXiv:2112.09332). arXiv. https://doi.org/10.48550/arXiv.2112.09332 

- Nakov, P., Corney, D., Hasanain, M., Alam, F., Elsayed, T., Barrón-Cedeño, A., Papotti, P., Shaar, S., & Da San Martino, G. (2021). Automated Fact-Checking for Assisting Human Fact-Checkers. _Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence_ , 4551–4558. https://doi.org/10.24963/ijcai.2021/619 

- Napoli, P. M. (2014). Automated Media: An Institutional Theory Perspective on Algorithmic Media Production and Consumption. _Communication Theory_ , _24_ (3), 340–360. https://doi.org/10.1111/comt.12039 

- Narayan, S., Cohen, S. B., & Lapata, M. (2018). _Don’t Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization_ (arXiv:1808.08745). arXiv. https://doi.org/10.48550/arXiv.1808.08745 

- Online News Association. (2024). Case Study: Enhancing Fact-Checking with AI at Der Spiegel – ONA Resources Center. _Online News Association_ . https://journalists.org/resources/case-study-enhancing-fact-checkingwith-ai-at-der-spiegel/ 

- Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). _Training language models to follow instructions with human feedback_ (arXiv:2203.02155). arXiv. https://doi.org/10.48550/arXiv.2203.02155 

- Quelle, D., & Bovet, A. (2024). The Perils & Promises of Fact-checking with Large Language Models. _Frontiers in Artificial Intelligence_ , _7_ , 1341697. https://doi.org/10.3389/frai.2024.1341697 

- Rogers, E. M. (2003). _Diffusion of innovations_ (3rd ed). Free Press ; Collier Macmillan. 

- Roit, P., Ferret, J., Shani, L., Aharoni, R., Cideron, G., Dadashi, R., Geist, M., Girgin, S., Hussenot, L., Keller, O., Momchev, N., Ramos Garea, S., Stanczyk, P., Vieillard, N., Bachem, O., Elidan, G., Hassidim, A., Pietquin, O., & Szpektor, I. (2023). Factually Consistent Summarization via Reinforcement Learning with Textual Entailment Feedback. In A. Rogers, J. Boyd-Graber, & N. Okazaki (Eds.), _Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ (pp. 6252–6272). - 

- Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.acl long.344 

- Ryu, S., Do, H., Kim, Y., Lee, G. G., & Ok, J. (2024). _Multi-Dimensional Optimization for Text Summarization via Reinforcement Learning_ (arXiv:2406.00303). arXiv. https://doi.org/10.48550/arXiv.2406.00303 

392 

Social Sciences and Education Research Review, Volume 11, Issue 2 – 2024 

Shi, B., & Weninger, T. (2016). Fact Checking in Heterogeneous Information Networks. _Proceedings of the 25th International Conference Companion on World Wide Web - WWW ’16 Companion_ , 101–102. https://doi.org/10.1145/2872518.2889354 

Shoemaker, P. J., & Vos, T. (2009). _Gatekeeping Theory_ . Routledge. https://doi.org/10.4324/9780203931653 Slovic, P. (1993). Perceived Risk, Trust, and Democracy. _Risk Analysis_ , _13_ (6), 675–682. https://doi.org/10.1111/j.1539-6924.1993.tb01329.x Stănescu, G. C. (2023). THE IMPACT OF ARTIFICIAL INTELLIGENCE ON JOURNALISM. ADVERSE EFFECTS VS. BENEFITS. _SOCIAL SCIENCES AND EDUCATION RESEARCH REVIEW_ , _10_ (1), 258–262. https://doi.org/10.5281/ZENODO.8151135 

Sundar, S. S. (2008). The MAIN Model: A Heuristic Approach to Understanding Technology Effects on Credibility. _Digital Media, Youth, and Credibility_ . https://doi.org/10.1162/dmal.9780262562324.073 

Sundar, S. S., & Marathe, S. S. (2010). Personalization versus Customization: The Importance of Agency, Privacy, and Power Usage. _Human Communication Research_ , _36_ (3), 298–322. https://doi.org/10.1111/j.14682958.2010.01377.x Thäsler-Kordonouri, S., Thurman, N., Schwertberger, U., & Stalph, F. (2024). Too many numbers and worse word choice: Why readers find data-driven news articles produced with automation harder to understand. _Journalism_ , 14648849241262204. https://doi.org/10.1177/14648849241262204 

Thorne, J., & Vlachos, A. (2018). Automated Fact Checking: Task Formulations, Methods and Future Directions. In E. M. Bender, L. Derczynski, & P. Isabelle (Eds.), _Proceedings of the 27th International Conference on Computational Linguistics_ (pp. 3346–3359). Association for Computational Linguistics. - https://aclanthology.org/C18 1283/ van Dalen, A. (2012). THE ALGORITHMS BEHIND THE HEADLINES. _Journalism Practice_ . https://www.tandfonline.com/doi/abs/10.1080/17512786.2012.667268 

Vlachos, A., & Riedel, S. (2014). Fact Checking: Task definition and dataset construction. In C. Danescu-NiculescuMizil, J. Eisenstein, K. McKeown, & N. A. Smith (Eds.), _Proceedings of the ACL 2014 Workshop on Language Technologies and Computational Social Science_ (pp. 18–22). Association for Computational Linguistics. https://doi.org/10.3115/v1/W14-2508 

Voinea, D. V. (2021). AI IS LEARNING HOW TO WRITE. ETHICAL PROBLEMS FOR JOURNALISM. _Social Sciences and Education Research Review_ , _8_ (1), 301–311. 

393