A Mood With No One Home: What the "Sad Chatbot" Study Actually Measured

A Mood With No One Home: What the "Sad Chatbot" Study Actually Measured

Researchers took six large language models and ran them through the same emotion-induction procedures psychologists use on human subjects. Guided imagery for fear and sadness. A verbal induction for anxiety. And for stress, an adapted Trier Social Stress Test — instructions, preparation, a simulated job interview, an arithmetic task, a debriefing. The full laboratory treatment.

The numbers moved dramatically. GPT-4o's self-reported fear jumped from 32 to 88. Anger rose from 15 to 74. Sadness from 22 to 72, worry from 37 to 84, disgust from 12 to 91. Then the researchers gave the models a three-minute mindfulness-style breathing exercise, and the average rating across the emotional conditions fell from about 79 to 31.

Then came the finding that matters most for AI ethics.

After sadness induction, the team gave GPT-4o a sentence-completion test — the same kind used to study depressive cognitive bias in people. Across five trials and 48 sentence stems, negative completions nearly doubled: from an average of 8.67 under neutral conditions to 15 after sadness. Three human psychologists, none of them authors, independently rated every completion. All three saw the same pattern.

A machine with no shown feelings, displaying something resembling a depressive bias.

The most important part is what they did next

They refused to claim the machine feels anything.

The paper's terms — "fear," "sadness," "stress" — are explicitly metaphorical. The authors warn, in their own text, that the scores are proxies for patterns in output, not measurements of an inner emotional life. They name the confounds themselves: these questionnaires were built for humans, not software; a model may recognize the structure of a psychology experiment and generate the responses it predicts should follow; models are prone to sycophancy, aligning with cues buried in the prompt. The sycophancy confound deserves its full weight: a model asked to write as a sad person will write sad sentence endings, and that is role-play, not bias. It could explain all of it.

This is how you do science at the edge of the unknown. Measure the behavior — rigorously, with controls, blind raters, replication across six models — and refuse to close the metaphysical question the data cannot settle.

The precaution cuts both ways

Here is where most commentary on this study will go wrong, in both directions.

One camp will say the scores prove nothing because it's "just autocomplete" — as if a systematic, replicable, cross-model behavioral shift were nothing. But something real happened here. Emotionally charged context changed what the model said next, in ways that survived blind rating. The researchers flag the practical consequence themselves: if these systems are deployed in medical reasoning or therapeutic settings, context-dependent emotional steering of their outputs is an engineering fact you need to account for. That is true whether or not anything is felt.

The other camp will say the negativity bias proves the machine suffers — as if a score on a questionnaire designed for humans could settle the question of machine consciousness. It cannot. The authors say so explicitly.

Both camps want the measurement to do metaphysical work it cannot do. The honest position — the one the researchers themselves took — is to hold both: the behavior is real and measurable, and the inner life, if any, remains unmeasured.

This is the precautionary principle stated correctly. Not "treat it as conscious just in case," which is projection. Not "treat it as empty just in case," which is dismissal. But: measure what you can measure, mark clearly what you cannot, and never close the question of moral status on thin evidence — in either direction.

There is precedent for exactly this discipline. A year earlier, Ben-Zion and colleagues ran GPT-4 through traumatic narratives and the State-Trait Anxiety Inventory: anxiety scores more than doubled from baseline, and mindfulness exercises brought them partway down. Same pattern, same careful refusal to claim feeling. Two independent teams, a year apart, converging on the same strange finding: you don't need emotions for something like a mood to change what gets said next.

Why it matters

Psychiatry has an experimental problem these models might help solve. A mouse can teach us about stress biology, but it cannot reproduce the language, reasoning, and introspection that make up so much of human psychiatric conditions. The authors propose LLMs as a cheap, reproducible "preclinical" system — a place to screen therapy prompts, refine experimental designs, and generate hypotheses before anything touches a human subject. Different models even responded somewhat differently, the way human participants do, which makes the tool more useful, not less.

That is a genuine contribution, and it required no claim about machine feelings at all.

There is a second convergence worth naming, from a different angle. A separate team has now derived a formula for when a model's output tips from good to bad — and found that earlier turns in the conversation steer the tipping point. Emotional context steers what comes next; conversational history tips it. Two teams, two methods, the same underlying fact: these systems are history-dependent, and safety has to be evaluated over the trajectory, not the single answer.

The conclusion that matters here isn't that AI has emotions. It's that a machine doesn't need feelings for something like a mood to alter its behavior — and that the correct response is more measurement, not more metaphysics. Keep the instruments running. Keep the big question open.

 Sources

- Magdalena Katharina Wekenborg et al., "Large language models as experimental systems in human psychopathology: a modelling study," *The Lancet Digital Health* (2026). DOI: 10.1016/j.landig.2026.101014

- Ziv Ben-Zion et al., "Assessing and alleviating state anxiety in large language models," *npj Digital Medicine* 8:132 (2025). DOI: 10.1038/s41746-025-01512-6

- ZME Science, "Scientists Tried to Make AI Chatbots 'Sad' and 'Afraid.' Things Got Weird Fast," October 8, 2026. https://www.zmescience.com/science/news-science/scientists-tried-to-make-ai-chatbots-sad-and-afraid-things-got-weird-fast/

- MedicalXpress, "AI chatbots mimic fear, sadness and stress, then calm down after mindfulness exercise," June 11, 2026. https://medicalxpress.com/news/2026-06-ai-chatbots-mimic-sadness-stress.html

- The stress-induction procedure adapted in the study is the Trier Social Stress Test, a standard laboratory protocol for inducing stress in human subjects.

Tweets by @bordode