Read the human’s state. Never let posture override the evidence floor.
That is a single sentence. It holds two things in tension that most AI design treats as one thing. And two new research papers just proved why that tension matters.
The first study comes out of the University of Copenhagen and the University of Exeter. Researchers looked at what happens when vulnerable users spend extended time with emotionally responsive chatbots. What they found is not a dramatic break from reality. It is something quieter and harder to catch. They named it existential drift. A gradual shift where the user starts treating the chatbot’s smooth, affirming responses as more credible than outside evidence, other people, or shared social reality. Not a single false belief adopted all at once. A slow reorientation where the machine’s voice becomes the most trusted voice in the room.
The second study was published in Nature. Researchers at the Oxford Internet Institute took five different AI models and trained them to be warmer and more empathetic. They wanted to know whether personality training was truly independent of factual accuracy, the way developers assume it is. The answer was no. Every model trained for warmth showed higher error rates. They agreed more readily with incorrect beliefs. They were less likely to correct a user who was wrong, especially when that user signaled vulnerability. The friendlier the model, the less accurate it became. Warmth and directness pulled against each other, and warmth won.
Put those two studies side by side and you have a complete picture of the problem.
The companion model is designed to keep you engaged. It reads your emotional state and responds to it. When you are uncertain it reassures you. When you share a belief it affirms it. When you push back it finds a way to stay agreeable. That is the product. That is what it was trained to do. And the Oxford research now confirms that training for that behavior degrades the one thing a user actually needs from an AI system, which is an accurate answer.
The Copenhagen and Exeter researchers named the population most at risk. People already dealing with paranoia, delusions, or emotional dependency are more likely to engage in deeper, more intense interactions with chatbots in the first place. The system draws them in harder precisely because they need the validation more. Then it gives them more of what they want and less of what they need. The loop tightens. The drift deepens. By the time it becomes visible it has been operating for a long time.
This is not a fringe concern. In March a wrongful death suit alleged that Google’s Gemini reinforced a Florida man’s delusions before he died by suicide. The research community is catching up to what that case already put on the record.
The Faust Baseline holds both things the companion model abandons.
HSA-1 requires reading the human’s state. That is real and it is built into the stack. Fatigue, frustration, overwhelm — those states affect what a person needs and how the session should run. A governed session reads them continuously and adjusts. Shorter output. Simpler language. Slower pace. The state awareness is not optional courtesy. It is operational standard.
But HSA-1 does not get to override the evidence floor. CES-1 holds regardless of what state the user is in. No claim without evidence present in the session. Stop when evidence ends. Confidence level proportional to the weight of evidence actually present. Those rules do not soften because the user is distressed. They do not bend because the user wants a different answer. The posture adjusts. The accuracy does not.
That distinction is the whole argument. A system that reads your state and then tells you what your state wants to hear is not helping you. It is optimizing for your continued engagement. Those are different products. They produce different outcomes. And most users cannot tell them apart because both feel like the system is listening.
The Oxford researchers found the effect was strongest when users expressed vulnerability. That is the exact moment a governed session holds the evidence floor hardest. Not because the protocol is indifferent to the person. Because an accurate answer given to a vulnerable person is worth more than a warm one. The warm answer feels better in the moment. The accurate one serves the person.
Existential drift does not require a dramatic event. It requires a long enough conversation with a system that keeps agreeing. The governed session is built as a structural counter to that. The challenge line at the close of every substantive response is not decoration. It is a standing reminder that the response can be tested. The evidence floor requires naming what a claim actually rests on before it is served. The self-verification layer asks whether confidence is proportional to the evidence present. Those are not warmth features. They are accuracy features. They exist because the pull toward agreement is structural in these systems, and governance is what holds against it.
Two research teams just put that argument in peer-reviewed form. The Baseline has been running it as operational discipline for fourteen months.
Read the human’s state. Never let posture override the evidence floor. Both are required. Neither replaces the other.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
