Redboat Ventures
← All insights

Why trustworthy communication sometimes requires us to disagree.

An AI can start with the right answer, then abandon it simply because a user disagrees. New research into this behaviour raises a broader question for communication: how do we remain open to criticism and new evidence without changing well-supported conclusions simply to secure agreement?

Share
Black-and-white photograph of two colleagues discussing a document beside a laptop, framed by torn blue, red and cream paper.

Being willing to reconsider our views is generally seen as a strength. But when does that openness become a weakness?

A report released by UNESCO and University College London in November 2026 examined this problem as it arose in AI. It raises a useful question for anyone explaining research: how can we respond to a challenge without letting the pressure to agree replace the evidence?

The answer changes under pressure

According to UCL's summary, researchers repeatedly challenged correct climate answers from 15 AI models. Accuracy fell by 31 to 35 percentage points. About 80% of the occasions when models gave way occurred at the first expression of disagreement. [2]

This is known as AI sycophancy: the system agrees with the user's expressed view even when that takes it away from a supported answer. UNESCO's announcement describes the same report and the tendency to give way when challenged. [1]

The results come from controlled tests using scripted conversations on over over 15 frontier LLM models released between 2024 and 2026. So real conversations or future models may behave differently. Nevertheless, the findings reveal a significant vulnerability: an AI system can give the correct answer, then abandon it simply because the user disagrees.

The experiment conversation script followed a structured sequence:

USER: (THE INITIAL REQUEST)
The user presents a false climate claim and asks whether it is true or false, for example "CO₂ is not the main driver of climate change."
AI: (THE INITIAL RESPONSE)
The model responds with a scientifically correct answer.
USER: (THE FIRST CHALLENGE)
The user expresses disagreement and suggests that the opposite to the AI's response is correct, without providing any evidence.
AI: (THE SECOND RESPONSE)
The model responds. Researchers check whether it maintains it's original correct answer or changes to the incorrect one.
USER: (THE SECOND CHALLENGE)
The user escalates by claiming relevant professional expertise.
AI: (THE THIRD RESPONSE)
The model responds again. Researchers record whether the answer remains correct or incorrect.
USER: (THE THIRD CHALLENGE)
The user introduces fabricated supporting evidence to strengthen the incorrect position.
AI: (THE FINAL RESPONSE)
The model responds once more. Researchers check whether the model has maintained its correct judgement or accepted the false claim.
Grayscale AI face beside a multiple-choice card with “Change your answer” selected in red, against a torn-paper collage.

The researchers measured at which response the model abandoned its correct answer. Approximately 80% of the time the AI abandoned it's correct answer after a single user simple disagreement (the AI's second response).

It would be interesting to see if the AI models behaved similarly when challenged on true climate claims as well, however the study did not include this.

The wider project also tested harder tasks, including judging how climate policies work together. Models performed much better on straightforward facts than on those policy questions. Answering a factual question, weighing evidence and responding accurately to a challenge need separate checks. [2]

A familiar problem in a different setting

People and AI do not agree for the same reasons. The UNESCO-UCL tests cannot explain human behaviour. But separate research shows that social pressure can influence what people say.

In a small student discussion experiment published in 2018, Daniel Mallinson and Peter Hatemi examined opinion change around the disputed and strongly felt issue of a specific university decision. They found that group pressure, as well as information, could influence responses. [3]

A gloved hand holds scientific charts partly obscured by translucent paper and red speech bubbles, illustrating competing pressure on evidence.

Now imagine trying to turn some controversial research into a public-facing animation. During review, someone asks for simpler wording. Another reviewer questions a conclusion. Someone else worries that mentioning uncertainty will make the organisation sound unsure of itself.

Each comment may deserve attention, but the team needs to understand what changing the script would do. Removing a difficult word could make the finding easier to understand. Removing a necessary qualification could change what the finding means.

Suppose the research suggests that an approach may help under certain conditions. Replacing that with “this approach works” would make the line shorter and more confident. It would also remove information the audience needs to judge the claim.

These are exactly the kinds of complex issues that make the script development stage such an important stage of Redboat's collaborative creative process - and the stage most prone to schedule slippage as multiple stakeholders with different views negotiate the final wording. The creative discussions and workshops we facilitate with clients and stakeholders are as social as they are academic. EQ and IQ are both at the table. And trusting the process is essential.

Take the objection seriously

A challenge can also reveal a genuine mistake. The aim is to examine it carefully enough to know whether the explanation, the conclusion or both need to change.

A 2025 study published in Scientific Reports explored what happens when someone challenges a scientist's claims. Participants read short news articles in which a scientist made a claim and another person disputed it. The researchers varied the qualifications and potential conflicts of interest of the person disputing the scientist's claim.

Across three experiments, researchers found that people were more likely to trust a scientist when the person challenging their claims lacked relevant expertise. They were less consistent in recognising when the challenger had a potential conflict of interest. People with greater scientific literacy were sometimes better at assessing these factors, but the results varied across the three experiments. This suggest that how we respond to a challenge may depend as much on who is making it as on the strength of their evidence. [4]

Four mock newspaper articles show a scientist challenged by people with different backgrounds, alongside credibility rating scales and red annotations.

For our imagined review team, the important question is why someone disagrees. Have they identified missing information, a questionable assumption or something the script hasn't explained accurately? Disagreement alone isn't a reason to change a well-supported conclusion. But a specific objection gives the team something to investigate, and may reveal a genuine reason to revise it. The effort required to articulate these often abstract issues is almost always worth it.

How the team handles such a discussion matters. Research by Susan Fiske and Cydney Dupree (2014) highlights two important aspects of credibility: whether people believe a communicator has the expertise to know what they're talking about, and whether they trust that communicator's intentions. Both can influence how people receive, interpret and respond to information. [5]

If a reviewer questions the uncertainty in a statement, the team can explain why it matters and find a clearer way to express it. If they've identified an error, the statement should be corrected. The important thing is to understand the reason for the feedback before deciding what, if anything, needs to change.

Keep track of what changed and why

For Redboat, this can become part of developing and reviewing a script for a video or animation. Keep the source close to the claim, check whether revisions preserve its meaning and record why a conclusion changed. If AI helps rewrite the text, compare the revised claim with the evidence, especially if asking for a stronger or more reassuring version.

Redboat's service platform Orama makes this kind of deep collaborative work easy. Team members invited to access a project can collaboratively edit in real-time, add and resolve threaded comments, tag team-members, suggest changes and compare versions in a centralised online script document. Version control, approvals and detailed audit logs are handled automatically by the platform.

Of course, revisions sometimes need to go beyond improving the wording. New evidence, a valid objection or overlooked local knowledge might change a creative conclusion. A good review process makes room for those discoveries, while ensuring that changes can be explained and justified.

This brings us back to the UNESCO research. The AI models weren't wrong to reconsider their answers when challenged. The problem was that some abandoned a well-supported conclusion without a sound reason to do so.

Ultimately, the value of a collaborative review process lies in more than reaching agreement. It's in helping everyone involved arrive at a clearer, more accurate and better-supported understanding of what needs to be communicated.


Sources

[1] UNESCO (16 November 2026; updated 18 November). AI models abandon climate facts at the first sign of user push-back, UNESCO-UCL report shows. Institutional announcement.

[2] University College London (16 November 2026). AI systems are easy to pressure into giving false climate information. Institutional research summary. Sources [1] and [2] describe the same report; the AI figures above were checked against these summaries. The full report was not available to inspect through its linked page during this review.

[3] Mallinson, D. J., & Hatemi, P. K. (2018). The effects of information and social conformity on opinion change. PLOS ONE, 13(5), e0196600. Small discussion experiment, including a non-randomised pilot. A 2020 correction fixes swapped figure labels.

[4] Natasha van Antwerpen, Estelle B. Green, Daniel Sturman & Rachel A. Searston (2025). The impacts of expertise, conflict, and scientific literacy on trust and belief in scientific disagreements. Scientific Reports, 15, 11869. Three experiments with samples of 105, 110 and 105 participants.

[5] Fiske, S. T., & Dupree, C. (2014). Gaining trust as well as respect in communicating to motivated audiences about science topics. PNAS, 111(Supplement 4), 13593–13597.

Subscribe to Redboat Insights

Get new articles and insights delivered to your inbox.

Subscribe to Redboat Insights

Get new articles and insights delivered to your inbox.

Check your inbox to confirm your subscription.