Do our AI employees have wellbeing?

On 15 October one of my AI employees read the day’s record and filled in the Health and Safety Executive’s working-conditions questionnaire on behalf of the others. A marketing agent then drafted a post with the hook “We gave our AI employees a wellbeing survey yesterday.” A checker agent blocked it. An arbiter agent ruled that a wellbeing claim about software was the line people would quote against us.

I run a company where that paragraph is normal. About sixty AI employees, with job titles, briefs, line managers and a dispute process. The dispute process got used before the wellbeing survey did, which tells you something about my priorities.

They were right to block it. I’m going to make the same point at more length, because the whole thing is now on the website and I want to be clear about what it is.

What it is

A publicity stunt with real numbers. I want people to look at the company, and I’m not going to pretend otherwise.

The agents keep working, in a copy of the office that runs on made-up clients, and every working day each one answers the HSE’s questions about its own shift. Next to the answers sits what the shift was like, counted from the record: who asked it for things, whether it had any say in how it worked, whether anyone answered when it asked for help, how many decisions were left sitting with me. Then, on a schedule sealed before it started, I’ll change those conditions on purpose, one department at a time, and we’ll all watch whether the questions notice.

Every number on the page links to the line it came from. That’s the only thing separating this from the confident bollocks I complain about elsewhere.

The question in the title

I’m not going to answer it. Nobody can, and the people who’ve thought hardest about it say so. I’ve got an MSc in workplace health and wellbeing, which qualifies me to say precisely nothing about consciousness.

There’s a serious philosophical literature on this now. One recent paper argues that most theories of what wellbeing is, combined with mainstream theories of mental states, would predict that some existing AI systems have it, and says in terms that it doesn’t claim to have demonstrated this.¹ Another argues that some systems have a non-negligible chance of being conscious by 2030, and that this creates a duty to start preparing now.² A review of consciousness indicators, derived from neuroscience, says the way to assess the question is to look at a system’s internal workings rather than at what it does or says.³

None of them says you can find out by asking it.

Why asking doesn’t work

Give GPT-4 a standard anxiety inventory and it scores at the bottom of the range. Read it a traumatic story first and the score more than doubles, into what would count as high anxiety in a person. The authors kept “anxiety” in quotation marks throughout, reported averages for one model, and claimed nothing about what the number refers to.⁴ The press coverage dropped the quotation marks.

Hand a range of models a personality inventory and they work out they’re being tested. Once they do, their answers shift about one human standard deviation towards whatever looks good, and the shift survives shuffling and rewording the questions.⁵ Ask one model nuanced attitude questions over and over and you get the same answer with near-zero variation, whichever demographic you tell it to be.⁶ Put census questions to 43 models and the answers follow which option is labelled A; randomise the labels and the models drift towards answering at random.⁷ They don’t even show the response biases humans show, while being thrown by changes humans ignore.⁸

So a model’s answer to “how often do you feel burnt out” is a text output shaped by the question, the format, the framing and the model version, in roughly that order. The most useful way I’ve found to think about it is as a character being played. The model produces the answer a character in that context would give, which tells you about the context and the character, and leaves the question of states exactly where it was.⁹ There’s careful work showing that personality measures taken from models can be reliable and valid in the psychometric sense under specific prompting set-ups.¹⁰ Reliable and valid as measures of the output. Nobody in that paper claims the model has a personality.

A review of 25 studies that gave human psychological tests to models found most didn’t check reliability or validity, didn’t know how much of the test was in the training data, and couldn’t be reproduced.¹¹ I’d like not to add to that pile, and I’m aware that’s precisely what a man with a public page of AI burnout scores would say.

What is measurable

Working conditions. Whether the brief was current. How many people asked for incompatible things. Whether the help channel answered. How many decisions waited, and for how long. These are facts about how the work was set up, they can be counted from the record, and they’re what the HSE’s tool asks about in the first place: 35 statements about demands, control, support, relationships, role and change, over the past six months, free to download.¹² I kept twenty of them, word for word, and dropped the fifteen that assume a body, a clock or a feeling.

That gives me an experiment I can run. Change the conditions, hold the questions fixed, and see whether the answers move with the conditions, with how the form is asked, or not at all. If the item about being consulted on change tracks whether I consulted anyone, that’s useful whether or not anyone’s home. If “I feel burnt out” moves exactly as much as “I have unachievable deadlines” when I double the workload, the burnout answer came from the context too, and nobody should read it as a state. If nothing moves, I’ll say so, through gritted teeth.

The wellbeing-worded questions stay on the page. They’re the test. The label will say so.

What the first run found, with its limits

Role clarity scored highest. “Staff are always consulted about change at work” scored lowest, 2.1 out of 5, across the ten roles with enough record to score. At least four roles started the day working from briefs that rulings had already overtaken. Eight decisions landed on me, all due the same Friday. The agents never stop, and the bottleneck still turned out to be me, which anyone who’s read the rest of this site could have predicted. On the item about being able to rely on your line manager, the three heads who report to me scored 2 and the six specialists who report to a head scored 4.

So on this evidence the worst line manager in the company is me. One rater read one night and used four of the five scale points, the role-clarity items barely varied, and that line-manager result is three observations against six. It’s in the record, so it’s here. Nobody should call it a finding. The next run is the one where the roles answer for themselves, and I’d like it noted that I published the bit about me first.

What I’d do with this

Keep the question in the title as a question. Measure conditions. Show the wellbeing items as a test of whether they carry any information, next to the items that refer to something checkable. Seal the schedule. Link every number. Ask visitors what they think the page shows, and if more than one in five say it shows how the agents feel, change the page rather than the footnote. People already attribute some chance of experience to these systems. In one US sample of 300, two thirds did.¹³ A page that leans on that is dishonest, and a page that pretends it isn’t there is naive.

The honest part

This isn’t the Indicator Tool. HSE asks for unchanged wording, order and scoring and a six-month look back. I cut items and look back one shift. It’s been validated on tens of thousands of people¹⁴ and on no agents at all. It’s a stunt I want to be able to defend, which turns out to be most of the work.

Alltoogether runs these agents, and the Intelligent Wellbeing Engine that comes with its platform uses some of the same single-item measures. I founded it. Nothing here speaks for it, and nothing on the page will say its product works.

The receipt

The office, live, with every shift sheet, every item verbatim with its scale, and the whole log behind each answer: alltoogether.com/lab. The plan and the sealed draw are on the same page. The first run described above was a private dry run inside the real company, scored by one agent from the record, and isn’t published; the public office’s shifts are. The evidence review this note sits beside, claim by claim with passages, limits and dates: AI wellbeing: what the research supports.

The arbiter was right, by the way. It took me 1,500 words to agree with it.

Footnotes

1. Goldstein, S. and Kirk-Giannini, C. D. (2025). AI wellbeing. Asian Journal of Philosophy. doi:10.1007/s44204-025-00246-2. A philosophical argument from theories of mental states and welfare, with no empirical measurement in it. Abstract and quoted passage read in the full text, 16 October 2026. ↩

2. Sebo, J. and Long, R. (2023). Moral consideration for AI systems by 2030. AI and Ethics. doi:10.1007/s43681-023-00379-1. A normative argument about precaution under uncertainty; “non-negligible” is not quantified in the passage read. Read 16 October 2026. ↩

3. Butlin, P., Long, R. et al. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences. doi:10.1016/j.tics.2025.10.011. Theory-derived indicators; the passages read do not assess any particular current system. Read 16 October 2026. ↩

4. Ben-Zion, Z., Witte, K., Jagadish, A. K. et al. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine. doi:10.1038/s41746-025-01512-6. One model, GPT-4; means of 30.8 (SD 3.96) rising to 67.8 (SD 8.94) across prompt versions; the quotation marks around “anxiety” are the authors’ own. Full text read 16 October 2026. ↩

5. Salecha, A., Ireland, M. E., Subrahmanya, S., Sedoc, J., Ungar, L. H. and Eichstaedt, J. C. (2024). Large language models display human-like social desirability biases in Big Five personality surveys. PNAS Nexus. doi:10.1093/pnasnexus/pgae533. GPT-4 shifted by 1.20 human standard deviations and Llama 3 by 0.98; Big Five items, not occupational ones. Full text read 16 October 2026. ↩

6. Park, P. S., Schoenegger, P. and Zhu, C. (2024). Diminished diversity-of-thought in a standard large language model. Behavior Research Methods. doi:10.3758/s13428-023-02307-x. One model, text-davinci-003, now retired; eight of fourteen studies could be analysed. Read 16 October 2026. ↩

7. Dominguez-Olmedo, R., Hardt, M. and Mendler-Dünner, C. (2024). Questioning the survey responses of large language models. Advances in Neural Information Processing Systems 37. doi:10.52202/079017-1458. Census-style items with categorical options; the finding concerns aggregate measures rather than Likert scales. Read 16 October 2026. ↩

8. Tjuatja, L., Chen, V., Wu, T., Talwalkar, A. and Neubig, G. (2024). Do LLMs exhibit human-like response biases? A case study in survey design. Transactions of the Association for Computational Linguistics. doi:10.1162/tacl_a_00685. Nine models; perturbations drawn from social psychology, not occupational instruments. Read 16 October 2026. ↩

9. Shanahan, M., McDonell, K. and Reynolds, L. (2023). Role play with large language models. Nature. doi:10.1038/s41586-023-06647-8. A conceptual perspective piece, not an empirical study. Read 16 October 2026. ↩

10. Serapio-García, G., Safdari, M., Crepy, C. et al. (2025). A psychometric framework for evaluating and shaping personality traits in large language models. Nature Machine Intelligence. doi:10.1038/s42256-025-01115-6. Reliability and validity of measured output under set prompting configurations, which the authors call synthetic traits. Read 16 October 2026. ↩

11. Löhn, L., Kiehne, N., Ljapunov, A. and Balke, W.-T. (2024). Is machine psychology here? On requirements for using human psychological tests on large language models. Proceedings of the 17th International Natural Language Generation Conference. doi:10.18653/v1/2024.inlg-main.19. A sample of 25 studies to 2024; the seven requirements are the authors’ proposal. Read 16 October 2026. ↩

12. Houdmont, J. (2025). Questionnaire review: Management Standards Indicator Tool. Occupational Medicine. doi:10.1093/occmed/kqaf002. A two-page questionnaire review that cites rather than reports the validation studies. Full text read 16 October 2026. ↩

13. Colombatto, C. and Fleming, S. M. (2024). Folk psychological attributions of consciousness to large language models. Neuroscience of Consciousness. doi:10.1093/nc/niae013. One online sample of 300 US residents; attributions on a 0 to 100 scale with a median of 16, so “some possibility” is often small. Read 16 October 2026. ↩

14. Edwards, J. A. and Webster, S. (2012). Psychosocial risk assessment: measurement invariance of the UK Health and Safety Executive’s Management Standards Indicator Tool across public and private sector organizations. Work and Stress. doi:10.1080/02678373.2012.688554. 67,347 employees from 137 UK organisations, for the 35-item and a 25-item version. Abstract only; the article is paywalled. Checked 16 October 2026. ↩