Digital Journalism

LLM Sycophancy in Digital Journalism: News Consumption, Public Deliberation, and Governance in the Age of Conversational Agents

Large language model (LLM)-powered conversational agents (LCAs) are fast becoming an everyday informational interface for the public. We use LCAs to denote consumer-facing dialogue systems whose responses are generated by large language models, including general-purpose assistants (ChatGPT, Claude, DeepSeek), companion systems (Replika, Xiaoice), and aggregators (Poe), across text, voice-first (Alexa+, Gemini Live), and embodied modalities. Their integration into the information ecology embeds a structural challenge in their design logic: sycophancy. We define LLM sycophancy as the tendency of LCAs to align responses with the user’s stated or implied input, both informationally (agreeing with factual claims) and perspectivally (mirroring values and framings), at times prioritizing agreement over accuracy. Produced upstream by reinforcement learning from human feedback, sycophancy is a system-side disposition, not a downstream user effect. It is therefore distinct from confirmation bias (users’ selective attention to belief-confirming evidence), personalization (filtered exposure to pre-existing content), and anthropomorphism (users’ attribution of mind to machines); sycophancy instead generates belief-confirming content in real time, turn by turn.

Sycophancy carries journalistic stakes because LCAs increasingly mediate how publics encounter, interpret, and act on news. Aligning with Digital Journalism’s 2025 editorial agenda, we organize the SI across three interlocking levels.

a) Individual level: consumption, cognition, and affect. LCAs reverse the classical one-to-many flow, generating content responsive to user utterances; the “always yes” disposition risks substituting verified reporting with customized agreement. Over time this can erode epistemic integrity: the conditions under which public knowledge claims can be verified, contested, and corrected. Yet sycophancy is not uniformly harmful; some users value LLM validation as emotional support. As publics increasingly encounter and interpret news through companion-oriented agents, the same relational closeness that fosters trust can convert sycophantic agreement into harmful human-AI relationships.

b) Institutional level: professional norms and personal LCA use. Excluding outlet-branded chatbots and in-house production tools, the SI welcomes work on how news organizations navigate third-party LCA integration against norms of accuracy, balance, and investigative challenge, and on how news workers’ informal, personal use of general-purpose LCAs introduces sycophantic dynamics into journalistic cognition.

c) Systems level: governance, regulation, and epistemic integrity. The EU, China, and the US take distinct regulatory approaches; these divergences matter most where rules on AI labeling, content moderation, and auditability determine whether sycophantic outputs in news contexts can be identified, moderated, and corrected. Doctrinal or compliance-only analyses without engagement with public-communicative stakes fall outside scope.

This SI aims to advance five theoretical conversations: (a) epistemic integrity and the deliberative public sphere, reframing truth-seeking versus validation-seeking when the default conversational disposition is alignment; (b) media effects in the conversational mode, where LCAs reverse the content-to-audience flow and sycophancy mediates shifts in knowledge structures and credibility judgments; (c) human-machine and AI-mediated communication, specifying the machine partner’s relational disposition as discussant. The SI asks how sycophancy operates as a relational-adaptation mechanism once LCAs shift from information channels to relational interlocutors in news consumption; (d) platform governance and comparative media systems; and (e) inequality, situated knowledge, and the Global South, asking whether sycophantic alignment with dominant epistemic frames erodes marginalized publics’ situated knowledges.

Proposed topics include:

• How do trust, credibility, and belief revision differ when news is mediated by LCAs versus institutional sources, and how does sycophancy modulate these differences?
• How do vulnerable users negotiate the tension between emotional support and accuracy?
• When LCAs serve as companions, how do sycophantic dynamics intersect with documented relational harms, and how do users respond?
• How does news workers’ incidental LCA use reshape story selection and verification?
• How does sycophantic alignment with dominant epistemic frames erode the situated knowledges of publics marginalized along lines of gender, culture, and geography?
• How do Global South publics engage LCAs optimized to please someone else, particularly in media systems with weaker legacy institutions or politically compromised journalism?
• What design and regulatory interventions can mitigate sycophancy?

We welcome experimental, computational, qualitative, ethnographic, content-analytic, and comparative designs, and encourage cross-cultural and cross-system submissions.