Guides
14 guides on the question the general AI-safety literature keeps leaving open: what to do when the input itself is confidential. Each one is self-contained and you can have a look, but please check out the FAQ page as well.
Law-firm guidance ends with “use a secure, enterprise-grade AI tool.” This is the framework for what that actually requires, and we look at it clause by clause.
Read the guide →The most common answer to the AI-confidentiality question is “we have an enterprise licence.” It answers the wrong contract, and you should know that before compliance requirements catch up.
Read the guide →Health privacy has a de-identification standard. NDA-bound information has none, and treating the two as interchangeable is how confidential data ends up in AI models.
Read the guide →Every guide says de-identify before you prompt. Almost none say what it costs your output. Here is the measured answer, and we look at it method by method.
Read the guide →A worked definition of the technique behind Untraceable, and how it differs from every other way of hiding data from an AI.
Read the guide →Every AI rule written so far assumes a person typing into a chatbox. Agentic AI is bounded by scope, not intent, and what it can reach on a working machine is not one confidential file, but years of accumulated data.
Read the guide →Every guide tells you whether you must disclose. Almost none tell you what to write, or that the agencies asking you to declare publish their own AI use. In this guide we use the agencies' own words to build the most defensible template available.
Read the guide →No American regulator requires you to declare that AI helped write a submission, but your NDA already does the work. And the federal energy in 2026 is aimed at state laws, not at authors. Regulation is moving quickly right now, so be prepared for what is next and avoid being caught behind the curve.
Read the guide →HTA agencies and journals increasingly require you to declare AI use, and the rule is often not in the guidance but in the submission form. A rule-by-rule account of what we found, and what it implies for pharma.
Read the guide →Executives name governance and compliance as their top AI agent concerns. Their staff, meanwhile, are pasting confidential documents into tools nobody approved. That gap does not close with a stricter policy, it closes with the right tool.
Read the guide →Text produced with frontier AI is becoming machine-detectable, and the parties who can run that check are the ones you answer to. The mark is not the problem. Work and watermarks that cannot be explained will be.
Read the guide →The field’s guidance says use AI carefully, keep humans in the loop, and be ready to defend it in front of an HTA. Here is the piece it leaves open: what to do when the input itself is confidential.
Read the guide →Most “AI compliance” writing is US-centric. Canadian consultants operate under PIPEDA and, in Quebec, Law 25, both of which bear directly on sending client data to an AI model.
Read the guide →A one-page, permissioned policy, and it is free to copy and adapt. Built around the question that actually matters: which data can go into which tool.
Read the guide →