A source spoke to you on condition of staying in the shadows. AI can help you make the most of the interview, but five mistakes are enough to give them away. Here they are, with the fix for each.
A source talked to you for two hours last night, on one condition only: never to be recognised. This morning, you have forty pages of transcript, a document they handed you and a piece due on Thursday. The question asks itself: can I hand all of this to an AI without putting my source at risk?
Yes, provided you choose where the AI reads and you know what really identifies someone. Here are the most common mistakes.
In Switzerland, the Federal Constitution guarantees editorial secrecy (Art. 17), and Article 28a of the Swiss Criminal Code allows journalists to refuse to reveal their sources. Across Europe, the European Court of Human Rights treats the protection of sources as part of press freedom under Article 10 of the Convention, and the profession's codes of ethics make it a duty. But this protection belongs to the journalist. A third-party service holding your notes does not enjoy it, and a provider subject to foreign law can be compelled to hand over whatever it keeps.
The text goes out with the source's voice in it: their turns of phrase, their job, the names they mention. It is often kept in the service's history, beyond your control.
The fix: work with a confidential model. In IA Confidential, the local and specialised models are open models installed on servers in Switzerland, which pass nothing on to third parties. They are all you need to sort through an interview.
Here is the transcript of an interview with a person I call [SOURCE A]. Sort their statements into three columns: what they saw themselves, what they were told, what they assume. For each one, say what would be needed to corroborate it. Do not add any fact that is not in the text.
"A manager in the purchasing department, who joined six months ago and was at the meeting on the 12th": there is no name any more, but there is only one person left. It is the combination of details that gives people away, rarely the name on its own. The guide on anonymisation and pseudonymisation explains this trap in detail.
The fix: in the Privacy panel, write your anonymisation instructions (each source's pseudonym, the investigation's code name, the company under scrutiny). They are applied before anything is sent to an external model, which you only choose for general questions. On top of that, the Confidentiality filter detects sensitive data and offers to answer with a confidential model instead.
A file passed on by a source contains more than its text: an author's name, a date, formatting specific to one department, sometimes a GPS location in a photo.
The fix: attach the document rather than pasting it. It is read in your browser, nothing is stored on our servers' disks, and images are stripped of their metadata, including the GPS location. If the document is still too identifying for an external model, the assistant flags it: it is then processed by a confidential model, or only your question is sent.
The morning news brief and the sensitive investigation in the same thread: one day, an extract from the wrong conversation ends up somewhere else.
The fix: one space per investigation, of the "Research" type. Its conversations are kept apart, and you can export the space to a file to archive it with your notes.
AI can be wrong and make things up with confidence: a date, a job title, a quote the source never said. When the assistant searches the web, the search goes out without identifying you and the results are dated and sourced, but every fact still has to be checked at the source. Responsibility for what is published remains yours.
Before publication, ask the AI to read your text the way someone hunting for the source would:
Here is the passage of my article that relies on an inside source. Point out every detail that could allow their employer to recognise them: job title, length of service, the date of a meeting, an expression quoted word for word. For each one, suggest a more neutral wording that keeps the information.
With a confidential model, the transcript, the attached documents and the answers are processed on servers in Switzerland, and nothing is kept on our side: your history lives in your browser. If anyone asked us for your conversations, we would have nothing to hand over.
Towards an external model, nothing is sent without your decision, once the Confidentiality filter has run: an anonymised version that you review, pseudonymised extracts from a document, or the message as it is if you accept that.
On a shared newsroom computer, the history can be read by anyone using the same session: work in a personal session and log out when you leave.
Before you open a real case file, read the data you should never paste into an AI and what happens to your conversations. Then try it out on a fictitious interview: open the chat.
