AI chat log anonymizer
Real conversation logs are gold for debugging a flaky assistant or studying how users phrase things — but they’re full of names, emails, order numbers, and other details you can’t paste into a public bug tracker or research dataset. This tool anonymizes a transcript in your browser, replacing identifiable information with consistent pseudonyms so the conversation stays readable and useful without exposing anyone.
How it works
Paste the chat log and toggle which categories to anonymize: emails, phone
numbers, URLs, IP addresses, card numbers, account or ticket references, and
person names. Each category uses consistent pseudonyms — the same email always
becomes [EMAIL_1], the same person always becomes [PERSON_2] — so the
back-and-forth still makes sense to whoever reads it. Live counts show how much
was replaced, and everything runs locally so the raw, sensitive transcript never
leaves your machine.
When to use this
Bug reports and issue trackers. A model behavior issue often only reproduces with the exact phrasing a user wrote. Anonymizing the log lets you paste the real conversation into a public issue or forward it to a vendor’s support team without exposing the user’s identity.
AI evaluation datasets. Building an evaluation set from real user conversations is one of the fastest ways to get realistic test data, but it typically requires PII removal before the data can be used internally or shared between teams.
Research and usability studies. Transcripts of human-AI interactions are valuable for studying how people phrase requests, where they get confused, and what clarifications they need. Anonymized logs can be shared with researchers or written up in papers.
Red-teaming and safety audits. When auditing an AI system for harmful outputs, reviewers often need to examine actual conversations that triggered concerning responses. Anonymization lets those conversations be reviewed broadly without exposing the individuals who triggered them.
What consistent pseudonyms preserve
The key design choice is consistency within a transcript: the same raw value always maps to the same placeholder throughout the document. This matters because:
- A debugging reader can still tell that “the person who sent EMAIL_1 is the same one who later complained about ORDER_7” — even without knowing who that person is.
- A dataset analyst can track a single anonymized user’s journey through a multi-turn conversation.
- The logical structure of an escalation (“user threatened to cancel, agent offered a credit”) remains intact.
Tips and notes
- Local only. The raw log is never uploaded — paste freely.
- Consistency aids debugging. Because identical values map to identical placeholders, a reader can still trace the conversation logic.
- Name detection is heuristic. It catches most full names but review the output for single names, nicknames, or unusual formatting it may miss.
- Datasets need more care. For public release, combinations of non-obvious details (zip code + age + job title) can still re-identify someone — review carefully beyond the automated pass.