Most redaction software asks for trust at exactly the wrong moment: before it has hidden anything, it wants your document on its server. Veil’s pitch is that it doesn’t. It’s a desktop app for macOS, Windows and Linux that finds sensitive information, removes it, and also lets you redline, fill and sign. The detection model and text recognition run on the computer.
I read through Veil’s source code to see whether that pitch holds up, and looked at the developer’s own benchmarks to see how well the detection actually works. The short version: the privacy story is real, and the detection is good but not good enough to trust blindly.
Detection: impressive for its size, imperfect in practice
Veil’s detector is tiny, about 0.6 million parameters, small enough to run on a single processor core. On synthetic test sets it outscores much larger open models like NVIDIA’s GLiNER-PII and Microsoft’s Presidio.
The more telling numbers come from real text it never trained on. On RedactionBench, a set of 200 hand-annotated documents, the default engine found about 89% of the details that had to be hidden, and about 63% of what it redacted was something that needed redacting. In plain terms, it misses roughly one sensitive item in nine, and it blacks out a fair amount that didn’t need it. That still beat every other system the developer tested on that benchmark, OpenAI’s privacy filter included. On legal documents specifically, though, OpenAI’s filter scored higher.
Over-redaction is the easier problem: Veil shows everything it found in a review panel, grouped by type, and you can switch off a category or remove one box. Missed items are the harder one. Veil makes review easy, but you still have to do it.
Redaction: done the careful way
Here Veil is better than many tools. When it exports a PDF, every page with a redaction is rendered to an image with the boxes burned in, so none of the original text, fonts or hidden layers survive. The new file carries no metadata, bookmarks, scripts or attachments from the original. Images are re-encoded, which drops location data. An optional certificate records what was removed by category and page, without the removed text.
The trade-off is that redacted pages become images. Veil puts invisible text back for the parts that weren’t redacted, so the file stays searchable, but it’s no longer the original PDF underneath.
The AI assistant is the surprise
Veil can connect to your own ChatGPT plan or an OpenAI, Anthropic or OpenRouter key and answer questions about the open document. By default the model sees the redacted text, with placeholders like [PERSON_1]. In the code, if detection fails for any reason, the text sent is replaced with [REDACTED] rather than sent as is. That’s the right default, and it isn’t one I expected to find. You can allow the full text for one document; that switch resets when the document closes.
Everything else
There’s a lot more than redaction: Word-style tracked changes and comments that export correctly to Word and PDF, version comparison, form filling that finds blanks on flat scans, profiles, drawn and digital signatures, and batch redaction of whole folders. It’s ambitious for a 0.1 release. There are rough edges: you can’t password-protect and digitally sign the same export yet, signature checks don’t validate certificate authorities, and text recognition handles Latin script only, so names in Cyrillic, Chinese or Hebrew scans may be missed.
Verdict
If you redact documents for work and don’t want to upload them anywhere, Veil is worth the 14-day trial and, at $129 or $249 once, easy to justify against a subscription. Use the Accurate engine for anything that matters, read every page before you send it, and treat it as a very fast first pass rather than a guarantee. That’s true of every automatic redactor; Veil just makes that pass faster and keeps everything on your computer.