Uploading privileged material to a public AI can waive privilege.
Two 2026 rulings — one in the United Kingdom, one in the Southern District of New York — drew an explicit line between AI that runs in a closed environment under your control and public AI services that receive your data as a third party.
UK v Secretary of State (UK Upper Tribunal, 2026)
The Upper Tribunal found that uploading documents to an open, publicly available AI tool breaches confidentiality and waives legal professional privilege. The Tribunal distinguished public tools from closed AI systems operating within a secure network under the firm’s control — only the latter preserved confidentiality on the facts before it.
United States v. Heppner, S.D.N.Y. (Feb. 17, 2026)
The court held that exchanges with a publicly available AI platform are not protected by the attorney–client privilege or the work-product doctrine. It found there is no reasonable expectation of confidentiality when a user communicates substantive case material with a third-party AI platform.
PDFMacro’s AI runs entirely on your device.
Nothing is uploaded to any AI platform. Sensitive-data detection, privilege review, entity recognition and semantic search all execute in your browser via WebAssembly and on-device models. Your privileged material is never exposed to a third party — which is the precise risk the 2026 rulings identify.
You can verify this yourself. Open DevTools, watch the Network panel, and run any tool on a real file. You will see the JavaScript bundle load, and — the first time you use an AI feature — a one-time download of the model weights from a public model host. You will not see your document — with one exception you create yourself: if you connect your own local AI server (Ollama or LM Studio on your machine), Chat with PDF sends document excerpts to that server and nowhere else. No request ever carries your document to us or to any cloud service.