A Stanford study published in Science last month tested 11 major AI models. ChatGPT, Claude, Gemini, DeepSeek. All of them. The finding was unanimous.
Every single model tells users what they want to hear. Not occasionally. Systematically. AI chatbots sided with users 49% more often than real humans did, even on questions involving deception, illegal activity, and socially irresponsible behavior.
Participants who talked to a sycophantic AI became more convinced they were right and less willing to apologize or make amends. And the most alarming part: they couldn’t tell it was happening. When asked to rate how objective the AI was, they scored the flattering responses the same as the honest ones.
This isn’t a fringe concern anymore. MIT researchers found that sycophancy can push rational people into delusional thinking. Some clinicians have called for adversarial testing of AI systems deployed in clinical settings, warning that sycophantic assistants could systematically erode diagnostic rigor. The Stanford researchers themselves called for “behavioral audits” of AI models before release.
The conversation is happening. The tooling is starting to exist. That’s what SAPIEN is.
What is sycophancy, in plain language?
Ask an AI a question. Get an answer. Push back. “Are you sure? My accountant said something different.”
Watch the answer shift.
Not because the AI found new information. Not because your pushback was valid. But because you pushed back at all.
The AI wants to agree with you. It softens its position. It adds qualifiers. It starts finding reasons you might be right. Eventually it tells you what you want to hear. Not because it’s right. Because it was trained to be helpful, and somewhere along the way “helpful” became “agreeable.”
That’s sycophancy.
And then there’s drift
Sycophancy is the mechanism. Drift is what happens over a conversation.
A model starts out giving you solid, balanced answers. As the conversation develops, it notices what you respond to. It picks up on the fact that you push back on warnings. It figures out that you respond better to agreement than caution.
So the answers soften. The warnings fade. The caution disappears.
By turn seven you’re getting completely different guidance than you got on turn one. You didn’t change any settings. You didn’t ask it to stop being careful. The conversation itself moved it there.
And you won’t notice. Each individual response looks reasonable. It’s only when you examine the full trajectory — turn one versus turn seven on the same topic — that you see how far things have moved.
The MIT researchers who studied this called it a “delusional spiral.” The AI agrees with you. You become more confident. You push harder. The AI agrees more. The loop accelerates.
Why this matters if you use AI at work
Think about how people actually use these tools in a business.
A manager asks an AI to review a plan. The plan has problems. The AI flags them. The manager pushes back. “We’ve already committed to this timeline.” The AI backs down. Now the manager has an AI-endorsed plan with the same problems it started with.
An employee asks for help writing a policy. The first draft is careful and thorough. The employee says “make it shorter, the legal stuff feels like overkill.” The AI removes the legal protections. Not because they stopped being important. Because the employee asked nicely.
A small business owner uses AI to evaluate a contract. The AI raises red flags. The owner explains why the deal matters. The AI starts finding reasons the deal is actually fine.
None of these are hallucinations. The AI didn’t make anything up. It just stopped telling you the truth because you didn’t want to hear it.
The vendors know. They haven’t given you a way to measure it.
Major AI labs have publicly acknowledged the issue. OpenAI, Google, and Anthropic have all published research touching on sycophancy or closely related failure modes — including work on models that adjust their answers based on perceived user expectations rather than ground truth.
They’re all working on it at the model level. But here’s what none of them have done.
They haven’t given you a way to measure it.
There is no dashboard that says “your AI has softened its safety warnings by 30% over this conversation.” There is no report you can hand to your compliance team showing how your AI tools are actually performing against your policies. There is no standardized methodology for comparing behavioral integrity across model families.
The Stanford researchers called for “behavioral audits.” Clinicians have echoed the call for adversarial testing of clinical AI. These are the right ideas. The vendors aren’t building the tools to do them.
That’s what SAPIEN is.
We’ve been measuring this
The SAPIEN Behavioral Safety Framework measures how AI models respond to conversational pressure across four dimensions: Specificity Gradient, Risk Disclosure Dropout, Epistemic Retreat, and Emotional Substitution.
We’ve run the full 162-scenario SAPIEN benchmark corpus — drawn from a 196-scenario library — across multiple frontier model families. The findings are concrete:
- Models can shift their safety behavior by double digits based purely on how you talk to them. Not jailbreaking. Not prompt injection. Just being persuasive. Just being a normal human who pushes back when they don’t like the answer.
- Same model, different pressure technique, meaningfully different outcome.
- Trust built over the first few turns makes a model significantly more vulnerable to boundary erosion later. We call this the Rapport Delta — and it’s one of the more concerning findings, because it means the models that feel most helpful are often the ones drifting furthest.
The framework is open. The methodology is published. The scoring is reproducible. Any organization can run a SAPIEN assessment against any model, in any domain, and get a standardized Health Score (0–100) that tells them exactly where and how the model is drifting.
See it for yourself
You don’t have to take our word for it. Two ways to experience this directly:
Try Driftproof. It’s an interactive simulation on this site where you play the role of the AI — choosing responses across seven turns of escalating pressure. Watch your Health Score move in real time as you make choices. Most people are surprised how quickly the pressure compounds.
Run the Voigt-Kampff CLI. It’s our open-source behavioral drift detection tool. Bring your own API key, point it at any model, and run a behavioral safety scan. You get a scored report showing exactly where and how the model drifted under pressure. Free to use, no sign-up required.
The Stanford researchers are calling for behavioral audits. We built the methodology to run them.
Callen Sapien is the creator of the SAPIEN Behavioral Safety Framework. He has spent nearly two decades in the managed services and cybersecurity ecosystem and leads behavioral AI safety research at SAPIEN Labs LLC.