Quick Answer: Are AI interviews legal in Germany, and what may they assess? They are legal, within limits that most candidates have never been told. A tool may record your answers and assess what you actually said against the role. It may not infer your emotions, motivation or honesty from your face, your voice or how fast you speak: that has been prohibited in the EU since 2 February 2025, for the employer as well as the vendor. You can also ask for a human to review a decision that was made automatically. Preparing for one is the same work as preparing for a human interview, with one difference: nothing in the room helps you.
The invitation says "interview". Nobody will be there.
What you have actually been invited to
Three different things get called an AI interview, and they are not equally serious.
A one-way video interview. Questions appear on screen, you record your answers, often with a time limit and no second take. A person may watch the recording later, or a system may score it first.
A chat screening. A bot asks about availability, work authorisation, salary expectations and a few role questions, then routes you on or out.
A scored interview. Your recorded answers are transcribed and rated against the requirements of the role, and the rating is passed to a recruiter.
The third is the one worth understanding, because the rules are clearer than most people assume.
The line the law draws
| What the tool may do | What it may not do |
|---|---|
| Record your answers and transcribe them | Infer your emotions or intentions from your face, voice or gestures |
| Rate what you said against the competencies of the role | Read stress, motivation or honesty out of your tone or speaking speed |
| Match keywords and experience to the posting | Derive sensitive traits, such as origin or beliefs, from biometric data |
| Pass a score and a recommendation to a recruiter | Be the only thing deciding your application, with no human involvement available |
The prohibition on emotion inference is not a future rule. It has applied across the EU since 2 February 2025 under Article 5 of the AI Act, it carries the highest penalty in the regulation, up to 35 million euros or 7 percent of worldwide turnover, and it binds both the company that built the tool and the company that uses it. Legal commentary in Germany treats the application process as part of the workplace context the article covers.
The right to human involvement is older. Under Article 22 of the GDPR, a decision based solely on automated processing that significantly affects you gives you the right to obtain human intervention, to state your view and to contest the outcome.
The tool is allowed to judge what you said. It is not allowed to judge how your face looked while you said it.
What changed this summer, and what did not
In July 2026 the EU postponed the heavy duties for high-risk systems. Recruitment tools fall into that category, and under the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, those obligations now apply from 2 December 2027 instead of August 2026.
Read that plainly: the protections many candidates assume already exist, such as documented human oversight and detailed information duties for the employer, are still ahead of us. The prohibitions did not move, and neither did the transparency duties that took effect on 2 August 2026. So the floor is real, but it is lower than the headlines suggest, and until the end of 2027 the practical safeguard you have is the question you ask.
What you can ask, without sounding difficult
You are allowed to ask, and asking is normal in a German process. Keep it short and factual, in writing, to the contact in the invitation.
- Whether a person reviews the recording before a decision is made
- What the tool assesses: the content of the answers, or anything else
- Whether the recording is kept, and for how long
- Whether an alternative format is possible if you have a disability or a technical limitation
A company with a clean process answers these in two lines. A company that cannot answer them has told you something useful about itself.
One honest caveat: refusing the format usually ends the application. Asking does not.
How to prepare, given nobody is in the room
The substance does not change. What changes is that there is no interviewer to react to you, no follow-up that tells you which part landed, and no chance to read a face and adjust.
That removes your safety net in three specific places.
Length. With a person you can tell when to stop. On a timer you cannot. Answers that run to ninety seconds and end on the result work; answers that wander do not.
Structure. Situation, task, action, result is worth more here than in a human interview, because the recording is often skimmed at speed or read as a transcript.
The unscripted moment. Many formats show the question once and start recording immediately. That is the same pressure a follow-up question creates, without a human asking it.
Preparing content is therefore the whole job. The competencies a role will assess are largely predictable from the posting, and for each one you either have a concrete example or you do not. Where you do not, the tool has nothing to score, however calm you look.
Practising out loud against questions you have not seen is the closest thing to the real conditions. TalentVP was built for that part: mock interviews that ask follow-ups and score what you said rather than how you looked, and an Interview Blueprint that maps your stories against one specific job so you know which questions you have nothing prepared for.
What to actually do
- Ask the two questions above in writing when you accept the invitation, especially whether a human reviews the recording.
- List the competencies in the posting and write one real example for each, in situation, task, action, result form.
- Time yourself at sixty to ninety seconds per answer and end on the result, not on an explanation.
- Record one answer and watch it back for filler, restarts and apologies. Fix the content first, the delivery second.
- Practise answers you have not rehearsed, because most formats give you one take.
- Check the setup once: camera at eye level, a quiet room, a wired connection if you have one, and a fallback if the browser fails.
- If a decision comes back without any human involvement, you may ask for a human review under Article 22 GDPR, in writing, with your reasons.
The short version
An AI interview is not a lie detector and is not allowed to behave like one.
It may weigh what you said. It may not weigh how you looked while saying it.
The protections that go further arrive in December 2027, so until then the practical move is to ask whether a person reviews the recording, and to prepare content that survives being read as a transcript.
Prepare the examples, not the expression.
Related guides
- German Job Interview: Why a B2 Certificate Isn't Enough, what actually breaks candidates in the interview.
- Why German Recruiters Reject CVs From Abroad, the filters that run before any interview.
- Probezeit in Germany: Why Employers Filter So Hard, the legal reason behind the caution.
- CV examples by role, how evidence reads on the page.
- ATS check, whether your document clears screening at all.
Sources
The legal provisions cited here are verifiable. How individual employers run these formats varies, and this article describes the rules rather than any one company's process.
- Regulation (EU) 2024/1689, the AI Act, Article 5 on prohibited practices, including emotion inference in the workplace, applicable since 2 February 2025.
- Regulation (EU) 2016/679, the GDPR, Article 22 on automated individual decision-making.
- The Digital Omnibus on AI, DLA Piper, on Regulation (EU) 2026/1744 and the deferral of the high-risk obligations for employment systems to 2 December 2027.
This article is not legal advice. For an individual case, the employer's own information and, where needed, a lawyer or a works council remain the right source.
