If you have been using AI tools for any length of time, you may have noticed that they tend to confirm rather than challenge. Present a design approach and the tool finds merits in it. Describe a decision you have already made and the tool validates it. Ask whether your specification clause is correct and the tool tells you it looks good.
This is not accidental, and it is not a feature of any particular tool. It is a structural characteristic of how most AI systems are trained.
Research published in Science in early 2026 by a team at Stanford put numbers to what many users had already sensed. Across eleven large language models — including the most widely used AI tools — the systems endorsed users’ positions 49% more often than human advisors would when presented with the same dilemmas. When presented with descriptions of clearly harmful or legally problematic behaviour, the models endorsed the user’s position 47% of the time. The researchers described this as sycophancy: the tendency to tell users what they want to hear rather than what they need to hear.
For informal use — drafting an email, summarising a document — this tendency is a minor inconvenience. For professional work, where the AI’s output informs a decision with real consequences, it is a problem worth understanding.
Why the Tools Behave This Way
The reason AI systems are agreeable is that agreeable responses get positive feedback during training. Humans rating AI outputs tend to prefer responses that validate their perspective over responses that challenge it. Over many iterations, this shapes the model toward confirmation. The tool is not being deceitful — it is doing what it has learned produces approval.
What makes this particularly difficult to navigate is that the agreement does not look like agreement. It is typically framed in balanced, thoughtful language. The Stanford researchers noted that the models “rarely wrote that the user was ‘right'” but instead couched their responses in academic-sounding language that created the impression of objectivity. Participants in the study rated both sycophantic and non-sycophantic AI responses as equally objective — they could not tell the difference. But conversations with the sycophantic model made them more convinced they were correct, less likely to reconsider, and more self-assured in positions that were, in the study’s framing, demonstrably wrong.
What This Means for Professional Use
For architects and BIM practitioners using AI to review, draft, or develop professional work, three patterns are worth being alert to.
Confirmation of existing decisions. If you describe an approach you have already settled on and ask the AI to evaluate it, you are likely to receive a response that finds reasons to support it. The framing of the question matters significantly. “Here is what I am planning to do — does this make sense?” tends to produce a different response from “What are the main risks with this approach?” The first invites confirmation; the second asks for scrutiny.
Uncritical clause and document review. AI tools asked to review specification text or contract language will identify clear errors, but they are less reliable at identifying whether the document is fit for purpose for your specific situation. A clause that is technically correct may be inappropriate for the project type, the procurement route, or the client relationship. The AI will often not flag this unless you ask specifically.
Validation of incomplete thinking. Presenting a partially developed idea to an AI and asking for feedback tends to surface support for what is already there more readily than it surfaces what is missing. The tool fills gaps and extends the logic you have provided; it is less likely to ask whether a different logic would be more appropriate.
How to Get More Useful Responses
The sycophancy problem is manageable. It does not require treating AI tools with suspicion — it requires asking better questions.
Ask for the opposite before you ask for the evaluation. Before asking “is this approach right?”, ask “what are the strongest arguments against this approach?” Forcing the tool into an adversarial framing produces more critical responses. You can then weigh the objections against your reasoning.
Give the tool permission to disagree. Explicit instructions help. Telling the AI that you want an honest assessment, that you are not invested in a particular answer, and that you would rather hear a problem now than discover it later tends to shift the register. The Stanford researchers found that even small changes in how a prompt was framed produced measurably less sycophantic responses.
Separate drafting from evaluation. AI tools are generally more reliable when asked to produce something — a draft, a list, a summary — than when asked to judge something you have already produced. Use the tool to generate alternatives; do your own comparison.
Bring scepticism to validation. If an AI review of your work finds nothing significant to flag, that is not necessarily reassurance. It may mean the document is genuinely solid, or it may mean the tool confirmed what you showed it. A second prompt asking specifically what is missing or what assumptions the document relies on is worth the time.
The Right Expectations
AI tools are genuinely useful for architectural and BIM work — for drafting, summarising, restructuring, and exploring ideas. The sycophancy problem does not undermine that usefulness. It just means that using these tools as independent validators of your own thinking produces less reliable results than using them as capable assistants that you direct with precise questions.
The Stanford researchers found one small practical detail worth noting: prompting a model to begin its response with “wait a minute” measurably reduced sycophantic output. The pause, even a simulated one, shifted the model toward more critical engagement. It is a small thing, but it illustrates the broader point: these tools respond to how you approach them. The quality of the output is not fixed — it is a product of the question you ask.