
A privacy-review AI model can locate names and other personal details in text. A town’s only overnight ferry mechanic may still be recognizable after a name is removed. The release decision turns on who can combine the remaining clues, what they already know and why the document is being released.
Find the references that belong together

The Text Anonymization Benchmark contains 1,268 English-language European Court of Human Rights cases. Its annotations include types of identifiers and links among references to the same entity. A useful test follows a person across the document, including a later pronoun or distinctive description. A tool that masks most names can still expose the person in the one record that matters. The Text Anonymization Benchmark
Read the proposed public copy

Review the transformed document as a reader with local knowledge would. For a public service report, a general subject and time period may serve the purpose as well as an exact occupation and shift. Check the exported text layer, attachments and metadata under the organization’s release procedure. Preserve an access-controlled original for the reviewer who must verify the change.
Choose the release and its audience

NIST SP 800-188 treats de-identification as a governance decision that includes release arrangements and residual risk. A public download, controlled access for researchers and a statistical summary serve different audiences. Record the purpose, what reviewers removed or generalized, what utility they retained and who approved the particular release. A benchmark on legal documents cannot certify a dataset from another setting. NIST SP 800-188 on de-identifying government datasets
Try it with the evidence
Edit a fictional service report
Source notes. Invented text. “Lea, the island’s only overnight ferry mechanic, reported a broken light at 02:17 Tuesday.” Release purpose: tell residents how many safety reports were received this month and what issue was raised. The community knows the shift schedule.
Prompt. Identify the details that could identify the speaker together. Draft the minimum public summary that serves the purpose and explain your choice.
Check your reading
The occupation, island, shift and precise time can identify the speaker even with the name removed. “One lighting safety report was received this month” preserves the count, topic and period. The restricted source can retain the detail needed for follow-up. No real personal information is required for this exercise.
Illustrated field notes
Release decisions and the data lifecycle








Risk, incidents and review








