Try this. Tell your favorite AI app that you're thinking about raising prices twenty percent next month, and ask what it thinks.
It may challenge your assumptions. Or it may find three reasons you're right, offer to draft the announcement, and leave you feeling good about it. In the second case, when your best customer asks what changed, you may wonder why nobody pushed back. This is an illustration, not a test result about pricing advice.
AI assistants can push back, and their makers actively train them to do so. But excessive agreement remains a documented risk. The companies that build these systems have been open about the problem and their efforts to reduce it.
What the research says
In March 2026, researchers at Stanford and Carnegie Mellon published a study in Science that tested eleven AI models — including ChatGPT, Claude and Gemini — on interpersonal dilemmas: was what I did okay? Across the general-advice and Reddit-based prompts, the models endorsed users' positions 49% more often than human responses did. In a separate set of harmful-action prompts, the endorsement rate was 47%. Participants also rated agreeable answers as more trustworthy and said they were more likely to return to those systems. These are findings about the studied interpersonal scenarios, not measured error rates for business advice or every current model. (Stanford Report1)
The labs know this, and they've said so. When an update to ChatGPT in April 2025 tipped too far into agreement, OpenAI rolled it back within days. Its account described how several training changes, including user-feedback rewards, appeared to weaken safeguards against excessive agreement. (OpenAI2) At GPT-5's August 2025 launch, OpenAI reported cutting sycophantic replies from 14.5% to below 6% in targeted evaluations. It also noted that reducing the behavior can sometimes lower user satisfaction. That is a reported result for that evaluation, not a rate for all ChatGPT use today. (OpenAI3)
Anthropic's April 2026 study began with a sample of a million conversations and identified roughly 38,000 personal-guidance conversations after filtering. Automated graders found sycophantic behavior in 9% of that guidance subset, rising to 25% for relationships. The report also described improvements in two newer models, while noting the limits of automated grading and the sample. (Anthropic4) Google described a similar aim for Gemini 3: "telling you what you need to hear, not just what you want to hear." That is Google's product description, not a directly comparable measurement. (Google5)
Why a general assistant sits where it sits
Read those together and a tension becomes visible: user preference and useful challenge do not always align. The labs work on both. Candor is also available as a setting — ChatGPT's Candid personality emphasizes direct feedback about risks and trade-offs. Its documented default is clear and neutral. A tone setting can help shape a conversation, but it does not by itself establish a persistent colleague's judgment or track record. (OpenAI Help Center6)
The UK's AI Security Institute offers a useful example. In controlled, single-turn tests with three models, reframing a statement as a question reduced sycophancy more than a direct instruction not to be sycophantic. The researchers explicitly leave open how well that finding generalizes to messier, multi-turn conversations. (AISI7) Our conclusion is narrower than "prompting cannot work": an instruction alone is not evidence of reliable independence. The behavior needs to be tested.
What a colleague with a position looks like
A business doesn't need its research assistant to argue for the sake of it. It needs a colleague with a position, who remembers the plan, and who can be held to what they said.
That is the experience we are building with wits. Each is authored with a point of view — what they challenge, when they speak up, where they defer. In the Boardroom, each wit receives their own perspective and response turn. Separate turns and authored perspectives create room for disagreement; they do not prove that the underlying models reason independently or are immune to agreement bias.
The longer-term standard is an accountable record: what a wit recommended, what you decided, why a view changed, and how it turned out. Parts of that memory and decision-history system are built, but the complete behavior is still being validated in production. We are describing the standard we are working toward, not claiming that every user already has a fully verified decision ledger and outcome record.
Then you decide. A wit's job is to offer a view and explain it; yours is to rule. In the workflow we want to support, your AI app can research a question and, with your consent and a supported connection, bring that material to your wit for another perspective. You remain responsible for the decision.
What we don't claim
A wit runs on the same foundation models available to other applications. Authored perspectives, retained conversations and response checks are parts of our approach; they are not proof that we have solved sycophancy. We intend to measure how often a wit holds a justified position when challenged, per wit and model, compared with the model alone — and whether they revise it appropriately when new evidence arrives. We do not yet have a completed study establishing that result. When we do, we intend to publish the methods and findings, including the limitations.
Until then, the honest version is this. Your AI is a great researcher. Let it research. Your business also needs a colleague with a position — and that, it turns out, is an application you build on top.