A new AI system redacts sensitive text and explains exactly why it blacked out each line.
Researchers describe an instruction-tuned language model that sanitizes documents using natural-language redaction rules instead of rigid pattern matching. A user defines what counts as sensitive, whether that is structured data like names and phone numbers or fuzzier categories like legal boilerplate, in plain English. A general-purpose LLM turns that description into detailed rules, which are then used to fine-tune a smaller model. That smaller model reads a document, reasons through the rules step by step, and redacts matching content while citing the specific rule behind each deletion.
Most sanitization tools only catch standardized PII and give no audit trail, a real problem in fields like legal discovery or medical records where regulators expect a documented reason for every redaction. By combining flexible, natural-language scope with a generated explanation for each cut, this approach targets the compliance gap that keyword matching has never solved. The team also built a reconstruction test to check whether redacted text can be pieced back together, reporting high reconstruction error and high redaction precision.
It is still a research prototype validated on its own benchmark rather than a shipped tool tested against messy real-world files, so the real test will be whether the smaller model stays fast and accurate enough for the redaction volumes compliance teams actually handle.