End of shift, three reports to write, and the same question every time: am I allowed to put this into an AI? The same task done twice, first without any care, then without anything from the case getting out.
It is nearly midnight. The shift is ending, the patrol has dealt with a neighbour dispute, a minor collision and an arrest for theft one after the other, and there are still three reports to write. The colleague at the next desk says it out loud: "It would be quicker if I pasted my notes into an AI." Then, after a pause: "But are we even allowed to?"
The short answer: not the way he was about to. The useful answer comes down to two tasks, each done twice: first the way they are too often done, then the way they should be done.
In Switzerland, official secrecy is protected by Article 320 of the Swiss Criminal Code, and the Swiss Criminal Procedure Code requires members of the criminal justice authorities to keep proceedings confidential. Most other countries have comparable rules: investigations are confidential, and those who take part in them are bound by official or professional secrecy, with criminal penalties for the most serious breaches. Everywhere, data processed for law enforcement purposes is subject to particularly strict data protection rules.
A report or a written statement brings together everything that secrecy protects: the identity of suspects, victims and witnesses, addresses, number plates, case numbers, and sometimes the line of inquiry being followed. On top of that comes something specific to the job: the statement records what the officer observed, heard or did. It is signed. Nobody else can write it.
Finally, your force, your command or your department has its own rules on IT tools. They take precedence. If they rule out all AI tools, the right course is to raise the need through the proper channels, not to work around the instruction.
The officer opens a consumer AI and pastes in his raw notes: the suspect's name and date of birth, the address of the home, the vehicle's registration, what the victim said, the case number. He asks for a finished report. In a few seconds, the entire case is with a provider that neither the department nor the people involved ever chose, and which often keeps the conversation history. The text he gets back reads smoothly: the AI has filled the gaps with plausible phrasing, and some sentences describe things that nobody observed.
What takes time is rarely remembering the facts: it is putting them in order, leaving nothing out, keeping a clean timeline. That work does not need a single detail from the case. So the officer asks for an outline based on a situation described in general terms, and refers to people by placeholders: [PERSON 1], [VEHICLE], [LOCATION], [TIME].
Suggest an outline for a report on an arrest after a theft from a shop, without any names: how we were called out, observations on arrival, people present referred to by placeholders, action taken, items recovered, follow-up. Under each section, list the questions I should ask myself so that I don't leave anything out, and point out what counts as an observation and what counts as a statement.
He does this in IA Confidential, on a confidential model. By default, the assistant answers with open models installed on servers in Switzerland and Europe, outside the reach of the CLOUD Act, which keep nothing of your conversations. He then takes the outline into the case management system, where he writes the facts himself, with the real names. The AI never knew which case it was.
The instruction seems harmless. The AI does as it is told: it corrects the spelling, but it also rephrases. "He appears to have been drinking" becomes "he was intoxicated". A statement quoted word for word is summarised in other words. "At around 10 pm" becomes "at 10 pm". Each change is minor, and each one shifts what the officer attests to under his signature. In court, the defence lawyer will compare.
The answer is to ask for a check, not a rewrite. The officer pastes a passage in which he has replaced the identifying details with placeholders, and asks the AI to point out, without touching the text, anything that raises a question.
Read this passage without rewriting it. Flag any vague times or places, breaks in the timeline, sentences that mix observation with interpretation, words that are too vague and spelling mistakes. Never change a statement in quotation marks, and keep the placeholders in square brackets exactly as they are.
He gets a list, not a new text. He decides, point by point, what he corrects in his own statement. The same method works for preparing an interview: a set of open questions, grouped by theme, with nothing leading, built on a typical situation.
With every new conversation, you have to tell the AI again that it must not invent anything, must keep the placeholders, must not rephrase statements. One tired evening, you forget an instruction, and that is exactly the evening it would have mattered.
The officer creates a "Reports" space and, in the Assistant tab of the space's page, writes his rules once and for all: flag rather than fill in, never change a quotation, keep the placeholders in square brackets, keep observation and statement apart, a plain, dated style. These instructions take precedence over his general preferences and apply to every conversation in the space.
In the Privacy panel, he also adds his masking rules: always mask registration numbers, case numbers, street names. They apply before anything is sent to an anonymised model, on top of his own vigilance.
AI can get things wrong, and it can make things up with total confidence: a section that does not exist in your procedure, an entry presented as mandatory when it is not, a questionable legal classification. It knows neither your internal instructions nor your case management system. An outline is checked against your department's templates, and a review comment is weighed before it is followed. What is written in a statement, the officer alone attests to.
With a confidential model, your requests and the answers are processed on servers in Switzerland and Europe, and nothing is kept on our side: the history of your conversations and spaces stays in your browser. We could not produce it if asked to, because we do not have it.
Anonymised models are the user's choice, never an automatic switch. If you select one, the Confidentiality filter checks the message first and, if it finds sensitive data, offers to answer with a confidential model, with your message unchanged, or to send an anonymised version that you review. Nothing goes out without your decision.
On a shared computer, the history can be read by anyone using the same browser session: open a personal session and close it at the end of your shift. And if you work with outlines and placeholders, what circulates cannot be used to recognise a case in any event.
Other uses of AI in a public service are covered in our guide for public servants and citizens' data. For the pitfalls common to every profession bound by secrecy, see AI and professional secrecy: five mistakes to avoid, and the list of data you should never paste into an AI.
To try it on a report outline, without any details from a case: open the chat.
Try it on your own case No account needed · First month free
Open
