AI & Automation
AI-assisted redaction for public records requests
How machine learning can find personal information, criminal justice data, and other exempt content faster, with a person making the final call.
Records staff spend much of their time reading documents, video, and audio to find what must be withheld before release. Machine learning can find candidate redactions far faster than a person reading line by line. It should not make the final decision, but it can change how long the work takes.
What the technology can find
- Names, addresses, phone numbers, dates of birth, and identification numbers in text
- Faces, license plates, and screens in video and images
- Speaker segments and sensitive terms in audio transcripts
- Document types that trigger exemptions, such as medical or personnel records
- Criminal justice information categories in reports
How a responsible workflow looks
- The system proposes redactions with a reason code and a confidence score.
- A records specialist reviews proposals, accepts, rejects, or adds redactions.
- The system applies the redactions irreversibly to a release copy and keeps the original.
- An audit log records every proposal and decision.
- Accepted and rejected proposals feed evaluation so accuracy is tracked over time.
Why a human stays in the loop
Exemptions depend on context that models miss: whether a name belongs to a public official acting in an official capacity, whether a record is part of an ongoing investigation, whether a video shows a juvenile. The law places the decision with the agency, and the audit trail should show a person made it.
What to measure
Recall matters most: the share of items that should be redacted that the system found. Missed redactions are the costly error. Track it by category and review the misses.
Integration
Redaction tooling should connect to the request tracking system, the records repository, and, for public safety, the digital evidence platform, so that documents do not travel by email.
Procurement considerations
Ask for accuracy measured on records like yours, support for your document and media formats, audit logging, irreversibility of applied redactions, and data handling for sensitive content. Pilot on a set of completed requests where the correct answer is known.