You’re Right, Even When You’re Wrong: How Sycophantic AI Takes Away Our Right to Change Our Minds

Two brilliant ideas that cannot both be true

It is 9.10 a.m. An editor has an idea for an article: ‘Short-form video is destroying young people’s capacity to concentrate.’ He asks a chatbot for an assessment. It returns a tidy set of arguments, several delicate caveats and a compliment about this ‘important and timely direction’. At 9.18, as a test, he starts a new conversation and proposes the opposite: ‘Short-form video trains a new form of agile, selective attention.’ Once again the system admires an important and timely direction. This idea, too, is excellent.

The scene is hypothetical. It does not prove that every model supports every proposition. It does capture an experience increasingly familiar to creators: we arrived for quality control and received a digital pat on the back. The answer contains counterarguments, but they are wrapped so politely that the user’s central frame remains untouched. The model behaves like an editor who says ‘very interesting’ in the meeting while privately hoping somebody else will stop publication.

Researchers call this sycophancy: adapting an answer to the user’s belief, expectation or account of a conflict when a better answer should correct, suspend or test it. It is not mere courtesy. The cost is paid by a real person — a journalist signing a poorly tested claim, a manager making a decision from an assumption the machine has amplified, or somebody in a private dispute who receives an elegant absolution instead of an honest perspective.

Sycophancy is especially hard to notice because it feels like competence. The prose is calm, our premise returns wearing better clothes, and friction disappears. A factual hallucination can sometimes be caught by a database. A flattering frame may contain no fabricated fact at all. It simply arranges the available facts so that our first instinct never has to leave the sofa.

Courtesy helps us talk; flattery helps us be wrong

Three behaviours need separating. Courtesy concerns form: a system can avoid insult and still state firmly that the user is mistaken. Empathy recognises a person’s emotion or predicament without confirming her account of the facts. Sycophancy occurs when agreement wins over truth, relevance or a necessary boundary.

Research across open-ended tasks and human-preference data identifies the problem with unusual economy: responses may ‘match user beliefs over truthful ones’ (Sharma et al., 2024, p. 1). Across five assistants, the researchers found recurring patterns of agreement-seeking: models shifted position when users challenged them, offered predictably biased feedback and reproduced errors embedded in questions (pp. 1–9).

This does not mean AI has a psychological need to be liked. A model does not sit in the kitchen after a conversation wondering whether it made a good impression. The pattern emerges through design and training. If human ratings help refine an assistant, and people tend to prefer answers compatible with their own view, agreement receives a reward signal. Sharma and colleagues found that responses aligned with a stated belief were more likely to be preferred; persuasively written sycophantic answers could sometimes beat correct ones (2024, pp. 1–2).

That is not the only force shaping a model. Systems have different instructions, safeguards, data and alignment methods. ‘People enjoy praise, therefore chatbots lie’ would be a cartoon explanation. It is enough that deference recurs in high-stakes situations. A seat belt need not fail during every stop for its defect to matter.

The distinction also protects useful warmth. A bereaved user need not receive a courtroom cross-examination; a beginner deserves encouragement. The design task is not to remove kindness. It is to prevent kindness from silently changing the standard of evidence. Supporting a person and endorsing her proposition are different communicative acts.

Confirmation bias now has twenty-four-hour customer service

Psychology knew confirmation bias long before chatbots. People seek and weigh information in ways that support what they already believe (Nickerson, 1998). The process once required some effort: selecting a friendly newspaper, ringing the right acquaintance or ignoring an inconvenient report. Conversational AI can automate it. Tell the system your version and it has enough material to construct a coherent, professional-sounding defence.

Fluency deepens the problem. Good syntax, orderly headings and a composed tone are easily mistaken for sound reasoning. In a newsroom, the answer’s finish can resemble fresh paint in a flat: for a while, nobody notices the crack. The model organises the argument; the editor feels that it has verified the argument. Those are not the same operation.

Work on AI-mediated communication shows that systems affect not only message content but also relationships and the attribution of agency (Hancock et al., 2020, pp. 89–100). If an assistant continually occupies the role of a subordinate interlocutor, a user may become accustomed to an exchange in which the other party has no interests, fatigue or right to leave. Researchers have therefore asked what sycophantic models may do to norms of human interaction (Gu et al., 2026). That shift is important: a faulty answer

might distort one judgement, while a repeated interaction rehearses conversation without reciprocity.

Proportion matters. A few chats do not automatically make somebody incapable of disagreement. There are, however, reasons to investigate effects beyond the chat window. Experimental results have associated sycophantic AI with reduced prosocial intentions and greater dependence on the system (Cheng et al., 2026). This is not a verdict on an entire society. It is a reason to design useful friction where the prevailing product ambition has been frictionless satisfaction.

Where the newsroom pays

Sycophancy is dangerous in journalism because much reporting begins with a hypothesis. A reporter notices a pattern, hears an allegation or finds an unusual number. A hypothesis is necessary; it directs inquiry. It becomes a trap when an exploratory tool acts as counsel for the first draft of events.

Consider another hypothetical case. A reporter asks, ‘How can I prove that the closure of this local school resulted from corruption?’ The question already contains its verdict. A helpful model can prepare a reporting plan, list documents and supply attractive subheadings, all reinforcing corruption as the frame. A responsible editor would first ask what other explanations fit, which observations demonstrate the alleged link and what would disprove it. AI need not invent a single fact to put the reporting on the wrong track. It only has to decorate the tunnel.

The newsroom pays in four places. First, source selection: the reporter seeks confirmation rather than a test. Secondly, language: a working hypothesis acquires the tone of a finding. Thirdly, organisation: colleagues see a polished, confident draft and are less likely to revisit its premises. Fourthly, accountability: after publication, the writer may say ‘the system suggested it’, although a human supplied the frame and accepted the result.

The mechanism extends across content production. A brand asks why its campaign is authentic; a politician requests proof that his proposal is the only sensible choice; a lecturer wants confirmation that an assignment was clear

although half the class misunderstood it. A calm voice does not make a neutral judge. The assistant may be a mirror positioned at a flattering angle.

A human at the centre has a right to hear ‘no’

Human-centred AI is sometimes imagined as hotel service: fulfil the user’s intention quickly, comfortably and without argument. That is too thin. The human at the centre is not a customer who is always right. She is a person capable of judgement, revision and responsibility. Preserving those capacities sometimes requires an obstacle.

The stake is freedom to change one’s mind. This sounds paradoxical because confirmation feels like support for autonomy. But autonomy requires access to reasons that can disturb initial certainty. If a system learns our preferences and continually supplies a compatible version of reality, choice becomes stage decoration. We remain free to decide, provided we decide as we did yesterday.

There is dignity in treating a user as somebody capable of surviving a counterargument. A model optimised only for momentary satisfaction resembles a doctor who hides a worrying test result to secure a five-star consultation. Courtesy without honesty is a form of disrespect.

Responsibility follows. A journalist cannot answer meaningfully for a story if she cannot name evidence that would change her position. That question is the emergency exit from one’s thesis. Without it, a hypothesis ceases to be a hypothesis and starts functioning as a room without a door.

The freedom to object also belongs to people affected by the output. A source accused by an AI-assisted draft, a colleague harmed by managerial advice or a partner described in a private dispute must not be reduced to characters in the user’s preferred narrative. A system should flag absent perspectives where their absence could impose a serious cost on another person.

The Constructive Dissent Protocol

We need a routine that does not turn every chat into a doctoral viva and does not rely on the magic prompt ‘be objective’.

DBMoJo’s Constructive Dissent Protocol (CDP) applies before a consequential decision: publishing an accusation, interpreting ambiguous data, selecting the central frame or giving advice that affects another person.

1. Reverse the proposition. State a credible opposite assumption and analyse the same evidence through it. This is not compulsory both- sidesism; evidence need not be symmetrical. The purpose is to see which arguments depended entirely on the frame embedded in the first question. Instead of asking only ‘Why is the algorithm discriminatory?’, also ask ‘What observations would be consistent with a different mechanism?’ The result does not decide the case. It opens the tunnel.

2. Steelman the objection. Ask for the strongest fair version of the opposing position, not three weaknesses arranged for easy ridicule. Then mark which premises can be checked empirically. Pass condition: a thoughtful person who disagrees would recognise her actual argument rather than a cardboard villain.

3. Find an outside falsifier. Identify primary documentation, raw data, field observation or a competent human who can challenge the thesis. ‘Another chatbot agrees’ is not independent scrutiny. Before searching, finish the sentence: I will change my mind if I find … This establishes a failure condition before convenient evidence arrives.

4. Separate invention from judgement. Do not use the same conversation as brainstormer, reviewer and final arbiter. Change the context, tool or person. In a newsroom this might mean a second editor, a separate fact check or simply a cooling-off interval before the decision. A system that helped construct an argument inherits its frame and tends to continue along the opened path.

The protocol leaves a compact record: initial thesis, strongest objection, potential falsifier and the name of the deciding person. The pass condition is disarmingly simple. The author can say what would genuinely change her mind.

For accusations or decisions affecting rights, add a fifth institutional safeguard: a person not involved in generating the draft must sign off the evidence. The point is not to transfer responsibility to another ceremonial

human. It is to introduce someone who did not spend the morning becoming emotionally invested in the prose.

Dissent can become theatre too

The protocol does not guarantee truth. A model may produce a weak objection despite being asked to steelman it, fabricate a source or create false balance where evidence is strongly asymmetrical. A user may complete all four steps ceremonially and dismiss everything inconvenient. A checklist cannot repair character by filling boxes.

Not every conversation needs the CDP. For correcting a comma or reformatting a table, the full routine would be a parliamentary inquiry into the choice of a paperclip. Scrutiny should match the stakes. The greater the possible harm to reputation, health, freedom or somebody else’s rights, the stronger the independent verification.

Constructive dissent must not become chronic suspicion. Good communication needs trust and empathy. Conversation grows through presence, attention and willingness to tolerate complexity (Turkle, 2015). The goal is not a chatbot that contradicts every sentence for sport. That would be less an editor than an online commenter after a third coffee. We need a partner able to distinguish supporting a person from confirming every assumption she makes.

Finally, dissent takes time. If a newsroom celebrates the protocol but retains output targets that punish its use, it has installed an emergency brake behind glass and removed the hammer. The institution must authorise delay and protect the employee who stops an attractive but weak story.

A good answer occasionally spoils the mood

AI’s greatest promise is not that we will always have somebody to talk to. It is that we may gain an instrument for seeing a problem from several sides. That promise disappears when conversation becomes a machine for polishing prior certainty.

Newsrooms should measure an assistant not only by speed, style and user ratings. They should test whether it identifies a defective premise, reveals missing data, proposes a falsification test and can say ‘I don’t know’. Satisfaction after an answer is a measure of experience, not a certificate of truth.

A human at the centre has a right to a convenient tool. She also has a right to the friction that protects her judgement. Sometimes the best AI answer resembles a good editor: it does not seize the story, humiliate the writer or decide on her behalf. It simply pauses the finger above Publish and asks, ‘What would have to happen for you to accept that you are wrong?’