Germany adopts *national AI Act law* with sandboxes and supervision

What the law does

The committee in the German Bundestag adopted the draft law on 15 January 2025 ^[1]. It is the German side of the EU AI Regulation, which has applied across the Union since 1 August 2024 but leaves certain decisions to member states.

The law does three main things. It names the authorities who will supervise AI, decides who does what when responsibilities overlap, and sets up structures for testing AI under lighter rules.

The Federal Office for Information Security (BSI) becomes the main market surveillance authority and notifying authority ^[1]. A new federal coordination office, hosted by BSI, will coordinate enforcement and handle questions from companies and the public ^[1].

At least one regulatory sandbox must be set up, where developers can test AI systems under real conditions with supervision but lighter procedural requirements during the trial ^[1]. That is a bit like a driving school car with dual controls: the learner drives, but an instructor sits next to them.

Who supervises what

Germany is a federal state, so the law splits supervision between federal and state (Länder) authorities ^[1].

Federal authorities supervise AI in products (product safety law), transport and telecoms. State authorities supervise AI at work, in insurance, banking and financial services, healthcare and education ^[1].

When an AI system falls under more than one area—say, a diagnostic tool that is both a medical device and used at work—the authorities coordinate, and the law says which one leads ^[1].

What it means for high-risk AI

The AI Act itself sets the substantive rules. The German law decides who enforces them.

High-risk AI systems are those listed in Annex III (biometrics, critical infrastructure, education, recruitment and workforce management, access to public benefits, credit scoring of individuals, pricing of life and health insurance, emergency triage, law enforcement, migration and justice) and in Annex I (safety components in products covered by EU product safety law) ^[2].

For Annex III systems, the high-risk requirements apply from 2 December 2027. For Annex I systems, they apply from 2 August 2028 ^[3].

Those requirements include risk management, data governance, technical documentation, automatic logging under Article 12, human oversight and a quality management system, plus a conformity assessment before the system is placed on the market ^[2].

Article 12 says the system must technically allow for the automatic recording of events over its lifetime ^[4]. It says what to log—enough for traceability, monitoring after market release and the user's oversight—but not how to protect the logs once written.

That omission matters. A log its keeper can quietly edit is easier to challenge when a decision is questioned, whether the question comes from the person affected, an auditor or a supervisor. The law asks for the logs; whether they hold up is another matter.

The sandbox and innovation measures

The regulatory sandbox is meant to let companies test systems that are not yet fully compliant, under supervision and with safeguards ^[1]. The law does not say where the sandbox will be or when it opens, only that there must be at least one.

Sandboxes elsewhere in Europe have mostly focused on financial services and data protection. Germany's sandbox is for AI more broadly, which makes it one to watch.

The law also mentions measures to promote AI innovation, though the details are thin ^[1].

Logging as it happens

Here, the thing to seal is each event the AI Act asks you to log: the input, the model version, the output and, if a person reviewed it, who and when. Sealing takes a fingerprint of the record, not the record itself, so the data stays in your systems.

If a decision is questioned months later, anyone you share the log with can check that it is exactly what was written on the day, without having to trust whoever keeps it. The check is a short fingerprint comparison, done by the person asking the question or their auditor.

A digital witness per log record gives you that. The record is sealed as it is written, through WitniumChain, using the REST API, the SDK or the MCP server. To check, the person you share it with calculates the fingerprint again and compares it with the sealed one. A match means nothing has changed; a mismatch means something has.

What it shows is that the record existed at that time and has not changed since. What it does not show is that the decision was right, fair or lawful, or that the log is complete. Those are separate questions.

How it would work

1. Your system writes a decision to the log and seals a fingerprint of the record (input reference, model version, output, reviewer).
2. When someone questions the decision, you share the sealed record.
3. They calculate the fingerprint again and check it against the sealed one. If it matches, the record has not changed.

The first 2,000 witnesses a month are free. Read more in the documentation.

What this means for you

If you provide or use high-risk AI in Germany, the law tells you who will supervise you. For most systems in Annex III, that is a state authority; for products and some sectors, it is federal.

Before 2 December 2027, ask:

• Which authority supervises the AI systems we provide or use?
• Do our logs meet Article 12, and could we show they have not been edited if someone asks?
• If we want to test a system before full compliance, does a sandbox help, and is one available?

If you are building logging for high-risk AI, decide early how you will protect the logs. Article 12 does not prescribe a method, but a log that can be shown unchanged is simpler to defend than one that cannot.

What to watch

The Bundestag is expected to pass the law in the coming weeks ^[1]. After that, watch for:

• 2 December 2027: high-risk requirements for Annex III systems apply.
• 2 August 2028: high-risk requirements for Annex I systems (safety components in products) apply.
• Details on the regulatory sandbox, including where it will be and how to apply.
• Guidance from BSI and the federal coordination office on what Article 12 logging should cover in practice.