A representative survey by TU Darmstadt, conducted with YouGov among more than 2,000 people in Germany, has found a striking link between AI literacy and job anxiety. The better people say they understand artificial intelligence, the more likely they are to expect it to take over their work. This counter-intuitive result challenges the common assumption that familiarity with AI reduces fear.
Key facts at a glance
- 43% of people with very strong AI knowledge expect AI could soon take over their work.
- That share is higher than among people with less understanding of AI.
- Only 15% of respondents have had any AI training.
- The survey was carried out by TU Darmstadt and YouGov, covering more than 2,000 people in Germany.
- Concern about job security rose across every occupational group compared with last year.
- Younger workers are more anxious than older workers.
- Hallucinations and data protection top the list of worries, ahead of superintelligence.
- Article 4 of the EU AI Act has applied since 2 February 2025, requiring AI literacy measures for staff.
The knowledge-anxiety paradox
Peter Buxmann, who led the study, called the pattern remarkable and worrying. Those who understand AI best are also the most aware of its disruptive potential. Instead of feeling more in control, they see more clearly how many tasks can be automated or augmented by AI systems. This does not mean they think machines will replace them entirely tomorrow, but they expect significant changes to their roles in the near future.
The fact that expertise increases fear rather than reducing it may seem surprising. Yet experts often have a front-row seat to the rapid improvement of large language models, image generators, and other AI tools. They have tested these systems and know what they can do. They also know the direction of development. For many, the question is not whether AI will affect their job, but when and how deeply.
Output-based exposure
The survey found that the industry in which someone works matters less than the type of output they produce. People whose jobs involve text, analysis, reports, code, or slides are more exposed. These are exactly the kinds of white-collar tasks where generative AI has shown the most capability. A marketing writer, a financial analyst, a software developer, or a management consultant may all face similar levels of exposure even though they work in completely different sectors.
This finding aligns with broader research on AI and the labour market. Routine cognitive tasks are increasingly within the reach of AI models. The ability to draft documents, summarise meeting notes, write code, and design presentations is no longer a uniquely human skill. Workers who produce such outputs are therefore more likely to see their tasks partially or fully automated.
The worry is not limited to any particular profession. Concern about job security rose across every occupational group compared with the previous year. That suggests the threat is perceived as widespread, not confined to a few vulnerable niches. It also indicates that workers are making connections between AI developments and their own livelihoods, even if they have not yet experienced direct displacement.
What worries workers most
The second surprise in the survey is the nature of the worries themselves. Hallucinations and data protection top the list, well ahead of any concerns about machines becoming superintelligent. Only 40% of respondents think superintelligence is even likely. This stands in contrast to the dramatic headlines that often dominate discussions about AI risk.
Hallucinations, where an AI model confidently generates false or nonsensical information, are a concrete and immediate problem for people who use AI at work. If an AI assistant produces a plausible-sounding report with invented numbers or fake citations, the user must spend time verifying everything. That can reduce the productivity gains that AI is supposed to bring. It also creates a new kind of workplace stress, because errors may be hard to spot.
Data protection is another practical concern. Many workers are unsure what happens to the information they enter into AI tools. Are their prompts being stored on external servers? Could sensitive company data be used to train other models? These questions are especially acute in Germany, where privacy culture is strong and legal consequences for data breaches can be severe. Employers who introduce AI without clear policies on data handling are likely to face resistance and distrust.
Younger workers more anxious
The survey also found that younger workers are more anxious about job security than older ones. That inverts the usual expectation that older employees, who may have less time left in the labour market, would be more concerned. One possible explanation is that younger workers have more of their career ahead of them, and therefore more to lose. They are also more likely to have grown up with digital technology and to have a clearer picture of how fast AI is evolving.
Younger workers may also be competing with AI for entry-level positions. Many routine tasks that used to be assigned to junior employees, such as drafting reports, preparing presentations, or writing initial drafts of documents, can now be done by AI tools. If companies choose to automate these tasks, they may hire fewer junior staff. That would make it harder for young people to gain experience and build their careers.
The rising concern across all age groups and occupational categories suggests that the anxiety is not irrational panic. It is a response to observable changes in the workplace. People see AI tools being adopted by their employers and in their professional networks. They read about companies reorganising teams around AI. They hear about layoffs attributed to automation. In this context, worry is a rational reaction.
The training gap and the AI Act
One number should trouble employers more than any other. Only 15% of respondents have had any AI training at all. Europe has had a rule about that for eighteen months. Article 4 of the AI Act has applied since 2 February 2025. It obliges providers and deployers of AI systems to take measures supporting AI literacy among the staff who operate them. The measures must be weighed against the relevant staff's training and the context of use.
Article 4 does not set a particular level of literacy and it binds only organisations actually using such systems. So 15% is not a compliance rate. It is a readiness figure instead. Eighteen months of a literacy duty, and most German workers have had nothing. This gap is not just a legal issue. It is a practical problem for companies that want to use AI productively and responsibly.
Workers who do not understand how AI systems work are more likely to misuse them, trust them blindly, or reject them outright. They are also more likely to be afraid. Training can help in two ways. It can give people practical skills to use AI tools effectively. And it can give them a realistic sense of what AI can and cannot do, reducing both panic and overconfidence.
The lack of training is particularly worrying because the AI Act's literacy obligation has been in force for eighteen months. The fact that most workers still report no training suggests that many employers are either unaware of the obligation, unsure how to implement it, or waiting for guidance that may never come. It also suggests that policy alone does not change workplace practice.
What employers can learn from the survey
The survey offers several lessons for employers. First, introducing AI without training creates fear. If employees understand what the tool will do and what it will not do, they are less likely to imagine worst-case scenarios. Second, training should be practical and role-specific. A marketing team needs different training from a finance team or a software engineering team. Generic presentations about AI will not build real literacy.
Third, employers should be honest about the limits of current AI systems. They should acknowledge hallucinations and data protection risks, rather than pretending that the tools are perfect. This honesty is more likely to build trust than reassurance. Workers who see their employers sugar-coating AI's shortcomings will only become more anxious.
Fourth, companies should think about how AI changes career paths. If junior tasks are automated, how will young employees learn? If reports are drafted by AI, what will human analysts do? These questions require deliberate responses. The goal should be to use AI to augment human work, not to quietly hollow out the early stages of professional development.
The pattern found by TU Darmstadt is not inevitable. Knowledge can lead to fear, but it can also lead to preparedness. The difference lies in what people do with that knowledge. Training, transparency, and honest dialogue can turn anxiety into competence. Without those measures, the fear is likely to keep growing.
In the end, the survey is a reminder that AI policy and AI practice are two different worlds. The AI Act may require literacy, but literacy will not happen by itself. Employers, workers, and policymakers all have a role to play in closing the gap between the rules and the reality. The eighteen-month record so far is not encouraging. But it also sets a clear benchmark for improvement.