AI in the risk assessment

Must AI-related strain be included in the psychological risk assessment?

Note: this applies to German law. Yes. § 5 (3) no. 6 ArbSchG is technology-neutral: if introducing AI creates work intensification, insecurity or overload, it belongs in the existing psychological risk assessment — as an extension, not a new separate duty.

AI-related risk factors and where they attach in the risk assessment

Risk factor	How it shows up at work	Where it attaches in the risk assessment
Work intensity	Same hours, higher expectations: whatever AI speeds up is immediately filled with new tasks	Work content and intensity — volume, deadline pressure, interruptions
Task design	AI takes over the demanding parts; what remains is checking and correcting	Work content — completeness of the task, scope for action
Job insecurity	An open question whether one's role still exists in two years; rumours instead of communication	Work organisation and leadership — information, predictability, perspective
Qualification pressure	Constant new tools without time to learn; learning happens outside working hours	Work equipment and qualification — fit between demands and competence
Opaque systems	Employees cannot follow how results were produced, yet carry responsibility for them	Work equipment — comprehensibility, feedback, usability
Social decoupling	Coordination moves from the colleague to the chat window; informal exchange breaks down	Social relationships — support, collaboration

What the German data shows — and what it does not

The most robust German data source on this topic is the BAuA survey DiWaBe 2.0: around 9,800 employees subject to social security contributions, surveyed in 2024, published in 2025. The result: 62.1 % of employees use AI applications at work, around 30.4 % intensively — and mostly informally, that is, without any company rules.

The central finding for the risk assessment is a double one: more intensive AI use goes together with higher work autonomy — and at the same time with higher work intensity. AI therefore shifts strain rather than simply reducing it. Equally important is what the study does not show: the BAuA found no statistically significant correlation between AI use and employee health. Declaring AI a health hazard as such goes beyond the German evidence.

Internationally the figures sound more dramatic. The Global Wellness Institute trend report for 2026 cites a global survey of around 37,000 workers in which worry about one's own job security due to AI rose from 28 % (2024) to 40 % (2026), with 62 % considering their leadership blind to the emotional impact. Those are global values, not German ones — usable as a direction, not as a number for your own workplace. The German complement comes from the AOK absence report 2025: 42 % of employees already encounter AI at their own workplace, but fewer than 40 % of employees in AI-using companies have been trained. This gap — deployment without qualification — is the strain driver you can actually measure.

The legal basis: § 5 ArbSchG is technology-neutral

Germany has no AI-specific occupational safety duty — and does not need one. § 5 (3) no. 6 ArbSchG has explicitly named 'psychological strain at work' as a hazard to be assessed since the end of 2013, without limiting where that strain comes from. Whether it originates in shift schedules, leadership behaviour or the roll-out of an AI assistant makes no difference to the duty.

The decisive point is the duty to keep the assessment current: under § 3 (1) sentence 2 ArbSchG the employer must review the effectiveness of measures and adapt them to changing circumstances. A company-wide roll-out of an AI tool that changes task design, pace and collaboration is exactly such a change. A risk assessment describing the company before the AI roll-out is simply no longer current.

Often overlooked as a flanking duty: Art. 4 of the AI Act has, since 2 February 2025, also obliged deployers of AI systems to ensure a sufficient level of AI literacy among their staff. Training is therefore not only the most effective lever against overload but already required in AI-using companies. And because AI systems capable of monitoring performance or behaviour regularly trigger mandatory co-determination under § 87 (1) no. 6 BetrVG, the works council is at the table for the roll-out anyway — the more practical route is to attach the strain survey to that same process.

Six dimensions along which AI shifts strain

DGUV forum (issue 1/2026) sorts the impact of AI along six dimensions of work: tasks, work organisation, working time, social relationships, work equipment and work environment. In each of them AI can relieve or burden — depending on how the introduction is designed. That is the decisive message: the strain sits not in the technology but in its design.

In practice this means: do not ask 'is AI a burden for you?', but work through the six dimensions one by one. Task design shows the pattern especially clearly — when the system takes over the demanding parts and turns the person into a checker, scope for action falls even though the volume of work stays the same. That is a classic strain factor of work psychology, not an AI-specific phenomenon.

Equally typical is opacity. Employees are expected to take responsibility for results they cannot follow. DGUV forum explicitly recommends explainable systems ('Explainable AI') so that internal processing steps remain comprehensible. Deciding this during tool selection reduces strain at the root; afterwards it can only be managed.

How to extend the survey in practice — without a second questionnaire

The most common mistake is turning AI-related strain into a special topic with its own survey. That creates survey fatigue, produces a second data set alongside the risk assessment and makes results incomparable. The right way round is the opposite: AI-related questions are integrated as an additional block into the existing survey — with the same validated instruments already in use (COPSOQ, psyGB, or whichever method the company has settled on).

A workable additional block needs only a few items covering exactly the factors from the table above: has your workload changed since the tool was introduced? Can you follow how the system reaches its results? Did you have enough time and guidance to learn how to use it? Do you know how your tasks are meant to change over the next twelve months? Four questions of this kind provide more steering knowledge than a separate AI report.

For evaluation the usual rules apply unchanged: results only in aggregate from groups of five people upwards, no conclusions about individuals, documentation as part of the risk assessment. And timing matters: one measurement before the roll-out and one six to twelve months afterwards shows the shift — a single measurement during the roll-out only shows a snapshot taken in an exceptional state.

Three levels on which measures work

DGUV forum distinguishes three intervention levels that translate directly into a measures framework. Technical: choose systems whose results are explainable, and deliberately limit the degree of automation instead of reducing people to pure checking instances. Organisational: run the AI introduction as a change process — involving employees in the planning, not only in the training. Personal: build AI literacy so that opportunities and limits can be assessed realistically.

The order is not arbitrary. Changing conditions comes before changing behaviour: a resilience workshop against work intensification that originates in tool design treats the symptom. Anyone who cannot or will not touch the cause should at least say so in the measures plan — that is more honest and keeps the assessment clean.

Two formats have proven to be low-threshold entry points in practice: a regular open AI office hour where questions can be asked without judgement, and transparent communication about which tasks new tools are actually meant to change. Both cost little and address exactly the two factors that come through most strongly in surveys: insecurity and lack of guidance.

Related measures & topics

Key takeaways

§ 5 (3) no. 6 ArbSchG is technology-neutral — AI-related strain belongs in the existing risk assessment, not in a separate document.
BAuA DiWaBe 2.0 (2024, ~9,800 employees): 62.1 % use AI, 30.4 % intensively; more intensive use means more autonomy AND more work intensity.
No statistically significant correlation between AI use and health (BAuA) — what matters for strain is design, not technology.
AOK absence report 2025: 42 % encounter AI at work, but under 40 % of those affected have been trained — the measurable gap.
Check six dimensions (DGUV forum 1/2026): tasks, organisation, time, social relationships, work equipment, environment.
Integrate as an additional block in the ongoing survey (COPSOQ/psyGB), measure before and 6–12 months after the roll-out, evaluate only from groups of five upwards.

Frequently asked questions

Do we need a separate risk assessment for AI?+

No. § 5 ArbSchG does not know technology-specific partial assessments. AI-related strain is integrated into the existing psychological risk assessment — as an additional focus of the survey, not as a separate document. A stand-alone 'AI risk assessment' creates duplicate work and makes results harder to compare.

When must the risk assessment be updated after an AI roll-out?+

As soon as working conditions change substantially — § 3 (1) ArbSchG requires adaptation to changing circumstances. A company-wide roll-out that changes task design, pace or collaboration is such a change. A single employee privately using a chatbot is not.

Does AI demonstrably make employees ill?+

The German evidence does not support that claim. The BAuA survey DiWaBe 2.0 (2024, around 9,800 employees) found no statistically significant correlation between AI use and health — but it did find one between more intensive use and higher work intensity. What matters for strain is work design, not the technology as such.

Must the works council be involved?+

As a rule, yes. AI systems suitable for monitoring the performance or behaviour of employees trigger mandatory co-determination under § 87 (1) no. 6 BetrVG — regardless of whether there is any intention to monitor. Since the risk assessment is subject to co-determination anyway, a joint approach is preferable to two separate procedures.

Which questions belong in the survey specifically?+

Four blocks are enough to start: change in workload since the introduction, comprehensibility of the system's results, sufficient time and guidance to learn, and clarity about one's own role over the next twelve months. These items cover work intensity, transparency, qualification pressure and job insecurity — and fit into any established method.

Make AI-related strain measurable

EasyBGM digitises the psychological risk assessment (COPSOQ and psyGB) — including additional blocks, evaluation only from groups of five people upwards, and a documented measures plan.

See EasyBGM Check your BGM maturity level

Sources

Last updated: 2026-08-18. Not legal or tax advice — have your specific case reviewed by a professional.

Psych Risk Assessment for SMEsDoes a small or mid-sized business (SME) have to conduct a risk assessment for psychological workload?GBPsych: Repetition and DeadlinesHow Often Must the GBPsych Be Repeated? Deadlines and Triggers at a Glance AI & Health DataMay AI analyze my employees' health data — what do the GDPR and the AI Act allow?GWI Trends 2026Which workplace wellbeing trends for 2026 matter for occupational health management in the German Mittelstand?