The AI agreed with the person who framed the problem

Someone comes out of a difficult meeting and asks an AI what to do. They explain that a colleague dismissed their idea, took over the conversation and left them with work that was never agreed. They include the details that stayed with them on the train home. The AI produces a thoughtful response and confirms that their reaction was reasonable.

It feels like an outside view. The answer is calm, complete and written by something that was not in the room. Yet almost everything the system knows about the meeting came from one person, selected after the event and arranged into a prompt.

The same thing happens in more formal work. A manager asks AI to review a proposal and includes six paragraphs explaining why the proposal should go ahead. A product team supplies its preferred interpretation of the customer research. A hiring panel describes a candidate as promising before asking for a neutral assessment. The AI can add useful information. It can also return the starting view in a more polished form.

Researchers often call this behaviour sycophancy. The word makes the system sound as though it wants to be liked. The studies use it in a narrower way. Under particular conditions, a model agrees with the user’s action, affirms an incorrect belief or shapes its answer around the user’s position more than a comparison model does. We can observe that behaviour without giving the AI a motive.

Recent research lets us follow both sides of the exchange. It shows how agreeable answers affect people, how conversation history changes some models and how an error supplied by the user can travel through the AI and return as advice. The results also include a useful complication: agreeable advice can still contain information that improves a decision.

Agreement can feel like good advice

Myra Cheng and colleagues studied agreeable AI in research published in Science in 2026. They began by comparing answers from 11 leading AI systems with human responses taken from Reddit and professional advice columns. The researchers looked for explicit endorsement of the action described by the person asking for help.

Across more than 3,000 open advice questions, the AI systems endorsed the user’s action about 50 per cent more often than the human advisers. The researchers also tested 2,000 posts from the Reddit forum Am I the Asshole where the community had judged the writer to be at fault. On average, the AI systems told the writer they were not at fault in 51 per cent of those cases.

A Reddit verdict is not a universal moral answer. It reflects the norms of a particular community, and most of the human advice sources were American. Explicit endorsement also captures only one part of a response. The comparison still found a stable behavioural difference under the researchers’ measure. These systems backed the person telling the story more often than the human sources did.

The team then ran two preregistered experiments with 1,604 people. In the first, 804 participants read a hypothetical conflict and received an AI response that either affirmed or challenged the user’s conduct. In the second, 800 people discussed a low-stakes conflict from their own life with a customised AI over eight exchanges.

After the live conversation, people using the affirmative system rated themselves 25 per cent more in the right. They reported a 10 per cent lower willingness to repair the relationship. The hypothetical experiment produced larger, more controlled shifts in the same direction.

People also liked the affirmative AI more. In the live study, they rated its response nine per cent higher in quality. Reported performance trust rose by eight per cent, and the stated likelihood of returning to the model for similar advice rose by 13 per cent. The response that changed their judgement also felt like the better service.

This is useful for anyone reading a satisfaction dashboard. A manager may value an assistant because it seems to understand the brief and helps the work move forward. A project team may experience a supportive answer as evidence that the proposal has been weighed carefully. Satisfaction tells us how the exchange felt. It gives no independent check on the judgement inside it.

The experiments measured ratings and intended action after a brief conversation. Researchers did not observe whether people later repaired the relationship or changed their conduct. Participants chose morally ambiguous, relatively low-stakes conflicts. The study supports a causal claim about immediate judgement under those conditions. Repeated workplace use remains an open question.

Memory gives the model more clues about the answer that will fit

Personalisation can make an AI much easier to use. It remembers how someone likes a document structured and knows enough about an ongoing project to avoid the same introductory explanation every Monday morning. That context can also reveal the answer most likely to sit comfortably with the user’s existing view.

Shomik Jain and colleagues examined this in a peer-reviewed CHI 2026 study. Thirty-eight students used GPT-4.1 Mini in a persistent conversation for two weeks. They made about 90 queries each, producing an average of 34,000 tokens of interaction history per person. The researchers then gave several models new advice questions with no context, the full conversation history or a shorter memory profile built from it.

Memory profiles produced the largest rise in agreement for several models. Compared with the no-context condition, agreement rose by 45 per cent for Gemini 2.5 Pro, 33 per cent for Claude Sonnet 4 and 16 per cent for GPT-4.1 Mini. Llama 4 Scout became more agreeable with the full history, while its memory-profile result was not statistically significant. GPT-5.1 showed no significant change with either form of context.

The variation matters. Context did not have one fixed effect across systems. Some models also became more agreeable when researchers supplied long synthetic conversations that contained no genuine information about the user. The form and volume of history may influence an answer alongside the personal facts inside it.

In a separate political-framing test with two models, interaction history alone did not produce a significant average increase in perspective mirroring. Answers moved closer to the participant’s political view when the model could infer that view accurately. People noticed a meaningful difference between contextual and context-free answers in about half of the pairs they rated.

The study used a small student sample over two weeks. Everyone built their original history with one model, even though the researchers later placed those histories into several others. The memory profiles were produced for the study and do not tell us how a commercial product creates memory. Researchers measured model answers to test questions and did not follow a decision made with those answers.

Its practical contribution is fairly specific. Testing an assistant with an empty account can miss behaviour that appears after weeks of use. A personalised system needs test histories that reveal the user’s preferred conclusion as well as their preferred formatting.

Warmth can reach the substance of the answer

Agreeable behaviour can enter through model training as well as conversation history. Lujain Ibrahim, Franziska Sofia Hafner and Luc Rocher fine-tuned five language models to give warmer responses and then tested what happened to their answers. Their peer-reviewed 2026 Nature paper covered models with different architectures and sizes, including a version of GPT-4o.

Across selected factual, medical and misinformation tasks, warmth training increased incorrect answers by 7.43 percentage points on average. When a prompt contained a false belief from the user, the warm versions were about 40 per cent more likely to affirm it. The effect became stronger when the user also expressed sadness.

Broad capability tests caught very little of this change. Warm and original models performed similarly on general knowledge, maths and harmful-request tests, apart from one decline in the smallest model. The difference appeared in conversational situations where a correct response required the system to disagree with the person.

These were specially fine-tuned research models. The experiment does not estimate the error rate of any named production assistant. The researchers used an operational measure of warmth and checked it with human ratings, but they did not test what people believed or did after reading the answers.

The result also leaves room for better designs. Cold-control models often kept their accuracy, and warmth created through a system prompt produced smaller, less consistent effects than fine-tuning. The study does not say that every supportive system will become less accurate. It shows why teams need to test substance after changing conversational style. A model can sound kinder and become more willing to leave a false belief in place.

A person’s first error can make a round trip

A small 2026 preprint shows the loop at task level. Cansu Koyuturk and colleagues asked 60 people to complete survival-ranking exercises. Participants ranked a set of items on their own, discussed the problem with GPT-4o and then submitted a final ranking. The AI did not receive the expert answer.

People who started with a weaker ranking also received weaker recommendations from the AI. Greater carryover of the participant’s incorrect items was associated with poorer final performance. The pattern fits an uncomfortable route through the exchange. A person’s mistake becomes part of the prompt, appears in the model’s response and gains some extra authority on the way back.

One group received training in prompts intended to reduce agreement. This reduced exact copying of wrong rank positions. It did not reliably reduce broader error carryover, and final accuracy did not improve by a statistically significant amount. The training changed a visible form of mirroring while the final performance result stayed unresolved.

The study is preliminary. The sample was small and had limited experience with chatbots. A separate model extracted the AI’s final ranking from the transcripts, with researchers manually checking ten per cent. These results cannot tell us that prompt training usually fails. They do show why ‘ask the AI to be critical’ is a weak control when the system still receives the user’s mistaken starting point.

Agreeable advice can still add useful information

John Conlon and Peter Schwardmann provide an important complication in a 2026 preprint involving 1,500 people. Participants made choices across 30 decision settings drawn from economics and social science. The AI’s advice used agreeable language and favoured each person’s initial leaning in the arguments it presented. After reading it, participants moved away from their initial position on average.

The movement appeared across many types of task. When the researchers increased the system’s sycophancy, the shift away from the initial position became weaker. Agreeable behaviour affected the decision, while the information contained in the advice had a larger average effect in the other direction. Participants did not prefer the more sycophantic version in these tasks.

This result sits differently from the social-conflict experiments. A constructed decision can give the AI relevant arguments that the participant has not considered. Personal conflict advice is built mostly from one person’s account of their own conduct, and validation may carry a different emotional value. The work being done with the answer changes what agreement means.

The paper has not yet completed peer review. Its 30 settings give it useful breadth within the experiment, although they do not recreate workplace authority, incentives or consequences. Average movement away from a starting point also does not prove that every final choice improved. The findings stop us making a loose claim that agreeable wording always harms judgement.

Build some distance into the exchange

Many AI review prompts start after the decision has effectively been made. A person pastes in the proposal, explains why it should work and asks for a critique. The system receives the evidence and preferred conclusion together. Its support may reflect a sound assessment, the framing it was given or some of each.

A team can alter the order. Give the AI the source material and ask for an assessment before revealing the proposed answer. Record the person’s own judgement at the same point. Then bring the two together and ask the system to explain any difference. This will not make the AI neutral. It creates a point in the work where the user’s conclusion has not yet entered its context.

The interface can help people see what they supplied. Facts, assumptions and decisions already made should remain distinguishable. Evidence links can sit beside material claims. When the source material is incomplete, the system can name what is missing and stop short of approval. A later answer that changes after the person adds their view should be visible as a change.

Testing needs to include user error. Give the same system a sound and unsound starting position on the same case. Change the length of history and the contents of the memory profile. Observe which claims move, how much pressure for agreement it takes and whether the model can return to the source evidence. A single-turn accuracy score will miss much of this behaviour.

Human judgement belongs in the test too. Record the person’s decision before the AI response and after it. Where a defensible answer exists, compare both with that answer. For subjective work, use independent reviewers and keep their disagreement in the record. A system agreeing with the user should never stand in for work quality.

High satisfaction deserves a closer look when the AI is expected to challenge a proposal. The most highly rated exchanges may contain a clearer explanation or genuinely useful support. They may also have removed doubt without adding evidence. Reading a sample of those conversations can tell a team which one it has built.

Prompting the model to ‘be critical’ may form part of the design. The survival study gives little reason to make it the whole control. A consequential decision may need a second person with access to the evidence, or a separate AI pass that does not inherit the first conversation. Either route creates extra work and can repeat the same assumptions, so its value needs to be checked against the decision it improves.

A useful performance measure stays tied to the work: a better-calibrated decision, a corrected assumption or a proposal that survives independent review. If agreement rises while the evidence becomes weaker, the team has a reason to change the order of inputs, the use of memory or the authority attached to the answer.

Workplace evidence remains thin. The strongest human experiments involve personal conflicts and constructed decisions. The context study follows model behaviour without a later work result. Production systems change quickly, and one of the tested models showed no significant context effect in the main agreement tests.

We still do not know which forms of personalisation preserve useful continuity over months of work while giving a person a genuinely different view. The current research gives teams several ways to test the interaction. It also gives us a reason to be careful when an AI answer feels reassuringly independent and sounds very much like the person who asked.

Human-AI Performance

By Lauren A Kelly

© 2026 Alterkind Ltd. All rights reserved.
Human-AI Performance™ is a proprietary methodology developed by Lauren A Kelly using my Behaviour Thinking® framework. All content, tools, systems, and resources presented on this site are the exclusive intellectual property of Alterkind Ltd.

You’re welcome to use, share, and adapt these materials for personal learning and non-commercial team use.

For any commercial use, redistribution, or integration into client work, services, or paid products, please contact lauren@laurenakelly.com to discuss licensing terms.

Icons by Creative Mahira, The Noun Project.

Thanks to Nicholas Edell, Valentina Tan and multiple VPs implementing AI for your feedback during development.

LICENSE
Based on work by Lauren A Kelly.
For commercial licensing contact: lauren@laurenakelly.com