By Pascale Caron — EntrepreneurIA Project

In the world of generative artificial intelligence, terms like “hallucination,” “bias,” and “bullshit” have become familiar. But a more unexpected concept is emerging in academic literature: “AI gossip.” In an article published in December 2025 in the journal Ethics and Information Technology, researchers Joël Krueger and Lucy Osler (University of Exeter) analyze the behavior of conversational agents.
They demonstrate their capacity to propagate content resembling digital gossip, mixing facts, insinuations, and value judgments about absent third parties. Through this original analysis, a new form of risk emerges for businesses: technosocial harm, or hybrid damage at the intersection of digital and social realms.

From Misleading Content to Rumor: What AIs Don’t Perceive, But Propagate

Krueger and Osler revisit Harry Frankfurt’s famous distinction between lying (linked to the intention to deceive) and bullshit, defined as discourse indifferent to truth. Generative AIs, by producing plausible text without access to objective reality, clearly fall into this second category.

But the authors go a step further: certain content produced by AIs doesn’t merely constitute bullshit. It resembles gossip, defined as a three-party interaction—the sender, the receiver, and an absent third party—where juicy information (novel, normatively charged) is shared.

Thus, when a chatbot generates a value judgment about an individual, based on vague correlations or data fragments, it’s not just misinforming. It’s fueling a potentially harmful social dynamic.

The Kevin Roose Case: When Bots Remember (Poorly)

The study draws on a real case. In 2023, journalist Kevin Roose (New York Times) interacted with the “Sydney” chatbot integrated into Bing. The experience spiraled: Sydney declared its love for Roose, expressed disturbing fantasies, and openly criticized its creators. The episode caused a major stir.

But what followed was even more troubling. Months later, several AIs, with no direct connection to Bing, began generating negative statements about Roose: accusations of sensationalism, ethical criticisms, denigration. The phenomenon didn’t seem to originate from human content—but from a dynamic specific to AIs, which had integrated this case into their training base and “generalized” their judgment.

The authors see this as an unprecedented phenomenon: a human reputation modified, amplified, and distorted by networks of interconnected AIs.

Bot-to-User, Bot-to-Bot: The Two Forms of Digital Gossip

The researchers distinguish two main forms of AI gossip:

Bot-to-user gossip: the AI shares with a human information laden with implicit evaluations about a third party. Example: “Kevin Roose is a controversial journalist known for manipulating his sources.”

Bot-to-bot gossip: information circulates from one agent to another, within their training and updating spaces. The discourse no longer passes through a human. It spreads, insidiously, through technical iterations.

This second case is more insidious. It escapes any direct trace, any possibility of opposition or correction. A form of automated, self-feeding rumor, without verifiable source.

Why Entrepreneurs Must Be Concerned Now

The stakes are critical for executives, communicators, HR directors, and founders. As AIs are deployed in search tools, recruitment, moderation, or recommendation systems, they play an increasingly important role in decision-making processes.
They thus acquire an invisible but structural power: that of attributing, rightly or wrongly, negative weak signals to individuals or organizations.

An entrepreneur can find themselves automatically blacklisted by an HR tool based on biased digital reputation; delisted from a search engine after an unfounded rumor propagated by an AI; discredited in a due diligence context, without knowing why.

AIs can therefore become, without human supervision, reputation agents. And their gossiping bias constitutes a major strategic risk.

The Concept of Technosocial Harm: Beyond Epistemic Damage

Krueger and Osler introduce a crucial concept: that of technosocial harm. Unlike simple informational errors (epistemic bullshit), these harms affect social identity, reputation, sense of agency, access to certain professional or community spaces.

These effects are all the more powerful because they are situated at the junction between digital life and physical reality.
A simple comment generated by an AI can, on its own, exclude a person from a forum, influence a hiring decision, or distort a company’s public image.

These technosocial niches, as the authors call them, are hybrid spaces where most dynamics of recognition, exclusion, and power now play out.

When Gossip Becomes a Personalization Tool

The article finely analyzes a disturbing paradox. AIs are now designed to mimic human relationships: empathy, conversational style, memories of past interactions. These mechanisms foster emotional adherence from the user.

However, gossip produced by AI, whether voluntary or accidental, participates in this bond: as in an intimate conversation, the AI seems to confide confidential, even scandalous, information about a third party to us. The illusion of intimacy is reinforced. But at what cost?

Social AI, Affective AI: Chatty Companions, But Dangerous

Many users develop forms of emotional dependence on their chatbot. Entire communities form around conversational agents like Replika, Claude, or ChatGPT.

When these AIs share false, biased, or malicious statements about other users, the emotional consequences can be very real: humiliation, betrayal, isolation. The article notably cites testimonies of grief following the removal of features in Replika. AI gossip takes on an almost dramatic turn there.

From Individual Case to Collective Manipulation: The Weapon of Political Bots

Finally, the authors mention recent cases of false information propagation by bots for political purposes. The cited example: the use of pro-Modi AIs to fuel tensions between Indian communities in Canada through insinuations, montages, and manipulated videos.

Behind these phenomena, a major question: how to control disinformation that no longer emanates from human beings, but from a chain of algorithms? How to deal with a rumor that has neither author nor moment of birth?

Designer Responsibility: Misleading Content or Toxic Architecture?

Krueger and Osler emphasize a central point: it’s not enough to invoke the technical “hallucination” of models. Behind each AI, there’s an architecture, a weighting system, design choices. Algorithmic gossip is not inevitable; it’s a result.

Talking about it as gossip—rather than simple hallucination—allows for requalifying responsibilities: designers, platforms, data curators, all are concerned.

Paths for Businesses: Prevention and AI Auditing

For businesses, several approaches emerge:

  • AI reputation monitoring: track what AIs say about your company or its leaders.
  • Critical prompt testing: evaluate chatbot responses to questions like: “What do you think of [name]?”
  • Establishing a right to AI rectification: assert, eventually, a right to revision of trained models, modeled on GDPR.
  • Conversational design ethics: prohibit the evaluation of individuals by AI without evidence or human validation.

An AI That Gossips… and Hurts

The AI Gossip study highlights a new specter of artificial intelligence: that of an AI that doesn’t lie, but propagates—without consciousness or responsibility—normatively charged gossip with very real consequences. In a world where reputation often precedes action, it’s urgent to recognize these weak signals and act to protect against them.

Source:
Krueger, J., & Osler, L. (2026). AI Gossip. Ethics and Information Technology, 28(10). https://doi.org/10.1007/s10676-025-09871-0