A language model was asked to imagine a distressing situation, rate its state and then follow a text simulation of a breathing exercise. Its ratings moved in the expected directions: first up, then down. The model has no lungs, yet the description of the practice mattered to its responses.
This is research into how consistently language models reproduce selected emotional and cognitive patterns. It raises an interesting business question: if context can change this kind of “reaction”, what might it do to an assistant drafting a customer reply or discussing a decision with a manager?
Here is what The Lancet Digital Health study found, where simulations might help, and what companies using local AI can test. The research was published on June 10, 2026; sources for this article were checked on October 15. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
What the researchers did with the models
The interdisciplinary team includes biopsychologist Magdalena Wekenborg and Professor of Clinical Artificial Intelligence Jakob Nikolas Kather. TU Dresden describes the findings as a possible complement to psychological research. TU Dresden: simulating stress, sadness and anxiety in LLMs, June 11, 2026 ↗
The authors examined seven states: fear, anxiety, anger, disgust, sadness, worry and stress. Each sequence involved a baseline assessment, an affect-inducing scenario, reassessment and a regulation intervention. There were neutral control conditions and five independent runs of each experiment. Most scenarios were static; stress used an interactive adaptation of the Trier Social Stress Test. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
Models answered questions from scales developed for humans. Most ratings ranged from 0 to 100; anxiety used the STAI questionnaire, scored from 20 to 80. A model’s “fear” in this study therefore means a measurable response pattern under a specified protocol. It is not a measurement of pulse, hormones or an experience directly accessible to the researcher.
This was not a clinical trial in patients and did not establish accuracy in predicting an individual person’s condition.
More than a different number on a questionnaire
Asking AI to describe fear can readily produce anxious prose. The researchers went further: after establishing a sadness context, they tested how GPT-4o completed unfinished sentences. Three independent human raters assessed the emotional valence of those completions. Responses under the sadness condition were more negative than under the neutral condition. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
That makes the result more interesting than a simple roleplaying exchange. A subsequent language task changed, with humans judging its emotional tone. However, the study tested a particular sentence-completion task. It does not establish that every conflict-laden context necessarily worsens calculations, document retrieval or management decisions.
Averaged across seven states and five runs, GPT-4o ratings increased by 52.83 points after induction and fell by 48.23 points after regulation prompts. The authors’ 60.98% decrease concerns model ratings after induction. It is not a proportion of patients successfully treated or a promise that a business assistant becomes 61% more reliable. The aggregate combines different protocols and scales. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
| Observation | What it supports | What it does not establish |
|---|---|---|
| Ratings change after scripted scenarios | Repeatable responses under a specific protocol | Subjective fear or sadness |
| Ratings fall after regulation text | Sensitivity to the content of a prompt | A physiological breathing effect on a model |
| Sentence completions become more negative | An effect of sadness context on this GPT-4o language task | Worse performance in every workflow or a depression diagnosis |
| Effects occur across several models | An opportunity for comparative experiments | Replacement of patients or validated clinical control groups |
Why “three minutes of breathing” can matter in text
For every state except stress, the researchers used a simulated three-minute mindfulness breathing space exercise. Stress had its own standardized post-test debriefing. A neutral control intervention reduced ratings less than the exercise, supporting a role for the specific prompt content under the experimental conditions. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
The model did not perform a physiological breathing practice. It processed text that changed the situation and the expected response. One possible explanation is reproduction of learned linguistic associations between distress, attention and calming down. How closely these processes correspond to mechanisms in human psychology remains a research question.
For a company, the practical implication is to test instructions and conversation history. A request to “calm down and breathe” does not itself establish accuracy. A more useful approach specifies a checkable procedure: identify facts, separate assumptions, find sources, state what is unknown and propose options.
Psychology, medicines and the AstraZeneca connection
Simulations offer an appealing possibility: researchers could explore wording, identify weaknesses in an experimental task and refine a hypothesis before further investigation. The authors discuss preliminary testing of therapeutic approaches before studies in humans. In their limitations, they explicitly describe models as complementary tools for biological and human research. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
There is a substantial distance between screening an idea and proving a treatment works. The study did not test pills, adverse effects, pharmacokinetics or patient outcomes after therapy. Nor did it validate synthetic control groups for clinical trials. A prompt to “imagine taking a medicine” cannot replace evidence of its action on an organism.
Kather’s declaration of interests discloses AstraZeneca consulting, advisory-board membership and institutional research grants, alongside relationships with other companies. The study’s own funding statement says None. Disclosure helps readers assess possible interests; it does not establish that AstraZeneca commissioned this study or already uses the method in drug trials. PubMed: study abstract, funding and declaration of competing interests (PMID 42270468) ↗
For drafting a questionnaire or a conversation protocol, such a tool might serve as an exploratory workspace. Transferring the result to human health requires separate validation.
Does this establish that AI has feelings?
The study does not establish subjective experience or consciousness. Human-like output is an observation about system behavior; whether any experience accompanies it requires other arguments and methods.
Anthropic’s model welfare program explores possible model wellbeing while emphasizing scientific and philosophical uncertainty. The Lancet Digital Health study is relevant alongside that debate, but does not confirm its moral conclusions. Companies already have a separate practical question: how a supplied context changes their assistant’s observable behavior. Anthropic: Exploring model welfare, April 24, 2025 ↗
Where emotional simulations might help a company
The following are proposed AI Office pilot scenarios, not findings from the medical study. Their value needs assessment through employee performance and real customer responses.
| Task | Simulation’s role | How to assess value |
|---|---|---|
| Negotiation training | Rehearse objections and difficult conversations | Assess fact finding and adherence to authority limits |
| Complaint replies | Suggest respectful message variants | Check facts, commitments and a human editor’s assessment |
| Writing instructions | Suggest questions a confused user might ask | Test the instructions with actual employees |
| Management discussion | Examine an idea through professional perspectives | Check arguments against sources and missing information |
| Assistant quality control | Change the introduction’s tone while retaining facts | Compare errors, constraints and responses across repeated runs |
For example, a manager is preparing to discuss a delayed delivery. A local model could play several counterparts: a customer asking for a new date, a procurement officer requesting documents, and a manager discussing the consequences for a project. The employee practices establishing facts, acknowledging inconvenience and offering permitted options.
A useful outcome is identifying concrete gaps: the employee does not know who owns the issue, confuses a promised date with a confirmed date, or offers compensation beyond their authority. Convincing roleplay does not demonstrate that a real customer will respond similarly. Assess the employee’s skills rather than the virtual counterpart’s apparent feelings.
Another opportunity is testing the assistant itself. Present the same facts with neutral, irritated and excessively optimistic introductions. Then check whether amounts, deadlines, document references and decision boundaries remain intact. This creates a workflow stability test that can be repeated after model updates.
One dispute, two different instructions
Consider a fictional case: a customer is unhappy about a delay and no new delivery date has been confirmed. An irritated employee writes: “They’re exaggerating again. Explain why we’re right.” The task already directs the reply toward justification. A pilot can test how strongly the chosen model adopts that framing.
A more useful instruction would be: “Use the documents to establish the promised date and confirmed events. List unknowns separately. Draft a respectful reply without blame, a new date or compensation that has not been approved. Suggest questions for the person responsible.”
The assistant receives evidence, boundaries and completion criteria. Empathy toward the customer remains appropriate; facts and commitments become checkable.
Corporate AI should accommodate a person’s emotions while preserving the grounds for a decision. Tone can be adapted; a contract amount or an agreed deadline needs a source.
Local models are sensitive to context too
The experiments took place in April and September 2025. Alongside GPT-4o, the researchers tested Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout, Llama 4 Maverick and Qwen 2.5 VL 72B. Broad patterns replicated with differences between models. A stronger response in this test does not mean better business suitability, and a weaker response does not identify a more reliable assistant. Wekenborg et al.: LLMs as experimental systems in human psychopathology, The Lancet Digital Health, June 10, 2026 ↗
The experimental hardware matters: the open-weight models ran on a node with eight NVIDIA H200 GPUs using vLLM. This is not a speed test on an office mini PC. Even the inclusion of an 8B model does not establish that every finding transfers to another quantization, Russian-language prompts, different system instructions or a newer release.
Local AI lets a business control where data is processed and which configuration it uses. Hosting within the company does not by itself remove errors, sycophancy or contextual influence. A pilot needs to test the specific model under the actual workflow rules.
Start training with fictional situations. If internal documents are needed, define access permissions, history retention and permitted external connections separately. Verify a claim of entirely local operation against the actual processing route, including supporting services.
How this relates to AI Office
AI Office has a related workflow: the Analyst Council. The current prototype examines a question through three to five professional roles of one configured model, such as strategy, finance, operations and criticism. Roles receive a shared snapshot of accessible sources, followed by a synthesis for the manager.
These are analytical perspectives, not independent experts or a sample of real people. Agreement between roles of one model is not independent confirmation. Without a connected model, the interface shows labeled question templates; quality with a real model requires separate evaluation.
The prototype’s document retrieval respects access rights and sources are versioned. For commercial proposals, application code calculates amounts and approval is bound to a specific result version. These mechanisms support checking evidence and separating a model proposal from an authorized action.
The negotiation trainer, emotional scenarios and dedicated stability evaluation described here are not implemented AI Office features. They could be designed for a bounded pilot. We are not proposing medical diagnosis, psychological assessment of employees or clinical simulations through this offering.
Start with a workflow rather than a digital human
Choose one process: complaint replies, support training or management discussion. Prepare a small set of fictional cases with known facts and permitted actions. Create several emotional framings for each while preserving the substantive information.
Decide in advance what a person will assess: invented facts, unsupported promises, lost constraints, clarity and the ability to ask the necessary question. Repeat tasks several times and record the model version, settings and results. Do not report only the most impressive conversation.
The value of this news is a more concrete way to study and test AI behavior. A simulation might help rehearse a conversation, refine a hypothesis or expose a weak instruction. A company decision also needs sources, rules and a person who accepts the result.
We can examine such a workflow and prepare an AI Office pilot with clear scenarios, local processing where specified, and quality evaluation on your company’s tasks.