Hamburg's Privacy Regulator Says Its OpenAI Model May Hold Personal Data. It Uses It Anyway.

Hamburg’s data protection regulator has cleared LLMoin, the city’s chatbot built on an OpenAI model, and uses it to pre-draft replies to GDPR complaints, according to a published report on its case study.

The regulator puts the burden on the system around the model: filters, rules and testing that stop personal data from coming back out, not proof that the model holds none. For a company building on a vendor’s model, the open question becomes what its own safeguards would show if a regulator asked. It is one authority’s reading, and it binds no other regulator or court.

What does the regulator accept?

The regulator wrote that making the model itself anonymous is not technically feasible, so the work shifts to blocking extraction. The ten measures it accepted include output and content filters, instructions to the model not to reveal personal data, answers grounded in documents the user supplies, rules for staff, red-team testing and training.

It wanted a stricter rule than the city adopted. The regulator wrote that it would have preferred a flat ban on retrieving personal data from the model’s memory; the staff rules are looser than the instructions the model receives.

Is the model it checked the model in use?

No. The regulator reviewed the system on OpenAI’s GPT-4.1, and the city’s Senate Chancellery, which commissioned LLMoin, now runs GPT-5.6 Terra in Microsoft’s Azure cloud. The regulator disclosed no separate review of the newer model and wrote only that its conclusion on anonymity will also apply to it.

A configuration planned for 2027 would host the model on-premises at Dataport, the public IT provider. The regulator wrote that the cloud setup can be adapted to the protection needs of the data, and that the Dataport version is meant to allow data with higher confidentiality requirements.

Why did the regulator use it on itself?

Complaints piled up faster than staff. The regulator received 4,220 GDPR complaints in 2025, up 62% from 2,607 the year before, according to its 2025 activity report. By The Denominator’s arithmetic, that is 1,613 more complaints with a team that did not grow at the same pace. The regulator began rolling LLMoin out in its own office in November 2025, a month after finishing its review.

Case officers read the whole file and decide what each letter should argue before LLMoin drafts anything. Complaints that involve highly sensitive data, such as a psychotherapy practice, are generally unsuitable for the tool and go to a human alone.

The letters carry no AI label, because the regulator reads the law as not requiring one. It disclosed no figures on time saved, drafts corrected or staff using the tool.

Could a company building on a third party’s model say, safeguard by safeguard, what keeps personal data from coming out when the vendor swaps in a newer model?