Introducing AI without overlooking the strain: the risk assessment that goes with it

This questionnaire is a BETA version: soundly built on the professional literature, but not yet conclusively validated in occupational psychology terms. It can be used, but not yet as the sole basis for an assessment under Section 5 of the German Occupational Safety and Health Act (ArbSchG).

Introducing AI changes the work — and with it the psychological strain. That is precisely where a legal duty applies: Section 5 ArbSchG requires hazards to be assessed whenever working conditions change substantially. An AI rollout is such a change.

What is known about this internationally The findings are remarkably consistent, and they do not come from the opinion pages. In 2026, the [International Labour Organization](https://www.ilo.org/resource/news/ai-driven-intrusive-surveillance-and-loss-autonomy-work-linked-psychosocial) warned of psychosocial risks arising from AI-driven work management: work intensification, surveillance, loss of autonomy. The [European Agency for Safety and Health at Work](https://osha.europa.eu/en/oshnews/implementing-safer-ai-worker-management-through-policy-and-prevention) (EU-OSHA) states explicitly that employers must assess the psychosocial risks of AI-driven management and take measures — this is existing occupational safety law, not a new special regime. And occupational health research now supplies effect sizes: in a [study on algorithmic management](https://www.sjweh.fi/article/4270), each additional level of management intensity raised psychosocial risk by roughly one fifth.

The efficiency arrives — the relief does not The most common sentence from workforces working with AI runs, in essence: it goes faster, but there is no less of it. That impression is well documented. [Analyses of time saved through AI](https://c3.unu.edu/blog/the-ai-productivity-paradox-why-your-ai-powered-workday-isnt-making-you-richer) show that by far the largest share of the time saved flows straight into further tasks; of the productivity gains, only a small single-digit percentage reaches wages. In work psychology terms this is an effort-reward imbalance: high effort, missing return — one of the best-documented patterns behind stress-related illness. The efficiency gain is real. The question for the risk assessment is where it ends up.

Two paths by which AI acts on the psyche [More recent work](https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2026.1798423/full) shows that the link between AI use and psychological strain runs almost entirely through two intermediate steps. The first is worry about one's own job — described in the research as technology-related job insecurity and measurable with dedicated instruments since [Brougham and Haar](https://doi.org/10.1017/jmo.2016.55). The second is loneliness at work: someone who now puts questions to the chatbot that they used to ask a colleague, while sitting alone in a home office, loses the informal contacts that otherwise cushion strain. Both paths can be shaped — but only if they are measured.

Training your own replacement A particular strain arises where employees train AI systems with their own experience while suspecting that this very experience is thereby becoming replaceable. That is not a diffuse unease but a role conflict: doing good work and weakening your own position coincide here. The questionnaire asks this question directly, because it is almost never voiced unprompted in discussion rounds.

Why this belongs in the risk assessment and not in a mood survey A mood survey collects opinions. A risk assessment records stressors and the strain that follows from them — and it obliges you to act where a need for action shows up. This module therefore follows the same twin axis as the standard survey: every statement is answered for how strongly it applies and how much it burdens. Not every change is a hazard. What matters is what it does to people.

What the twelve questions cover Six areas with two questions each: work intensification and pace. Relief and reward — does the time gained reach the employees. Control, transparency and participation — do I decide on the use, do I understand the results, was I asked. Job security and professional future. Social integration and isolation. Competence, trust and responsibility — including the question of who is liable when an AI result is wrong. Plus an opening question on how strongly AI shapes the working day at all: when analysing, it allows a comparison between intensive and occasional users without recording any personal characteristic.

Why there are deliberately only twelve questions The module is meant as a supplement, not a replacement. Anyone already carrying out the psychological risk assessment attaches it to the standard survey or to the COPSOQ — and there every extra minute counts against the participation rate. A workforce will answer four to five minutes; twenty minutes will be answered by part of it.

Anonymous, and firmly so The survey cannot be traced back to individuals; the analysis appears only as a group picture. On a topic that is about the fear of losing your own job, that is the precondition for honest answers. For the same reason the AI interview mode is locked for this module: an assessment under Section 5 ArbSchG is collected in the form — and having a survey about AI conducted, of all things, by an AI would also be the wrong gesture in substance.

How to get started concretely Create the survey, adjust the questions for this run if needed, invite anonymously, read the results as a picture by area. Conspicuous areas are the starting point for measures — and here those often have little to do with technology: binding rules on where saved time goes; participation before the next rollout; fixed on-site times against isolation; clarified responsibility for faulty results. Repeat the survey after the next major AI step: only the comparison shows whether the introduction went better than the last one.

Sources [ILO — AI-driven surveillance and loss of autonomy at work (2026)](https://www.ilo.org/resource/news/ai-driven-intrusive-surveillance-and-loss-autonomy-work-linked-psychosocial) · [EU-OSHA — Safer AI-driven worker management through prevention](https://osha.europa.eu/en/oshnews/implementing-safer-ai-worker-management-through-policy-and-prevention) · [SJWEH — Algorithmic management and psychosocial risks](https://www.sjweh.fi/article/4270) · [Frontiers in Public Health (2026) — Generative AI and psychological strain](https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2026.1798423/full) · [Frontiers in Public Health (2026) — AI and the digitalisation of work, review](https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2026.1857108/full) · [Brougham & Haar (2018) — STARA](https://doi.org/10.1017/jmo.2016.55). The same evidence is collected on our research page.