Loading page
About this tool
Scan pasted or exported AI conversations locally for personal data and secrets, then create a redacted transcript and counts-only report.
The AI Conversation Privacy Scanner reads a pasted chat log or an exported ChatGPT/Claude-style JSON file in your browser and flags personal data and credentials across 15 detector categories — names, emails, phone numbers, addresses, dates of birth, Aadhaar, Indian PAN, passport numbers, US SSN, payment cards, bank accounts, IBAN, IP addresses, MAC addresses, and API keys, tokens and passwords. It rebuilds the conversation as a redacted transcript you can copy or download, plus a counts-only JSON report that contains no conversation text. Nothing is uploaded — the parsing and the pattern matching both run in the tab.
Open AI Conversation Privacy Scanner on AltFTool — it loads instantly in your browser.
Paste the log into Chat text or export JSON, or use Open file to load a .json or .txt export of up to 2 MB; anything past 5,000 messages is cut off with a truncation warning.
Pick a Redaction style — Stable labels, Partial masking or Remove values — narrow the list under Choose detector categories, then press Scan locally.
Read Messages, Flagged, Signals and Format above the redacted transcript, then use Copy transcript, Download TXT for redacted-ai-conversation.txt, or Counts-only report for ai-conversation-privacy-report.json.
Parses ChatGPT-style mapping trees, messages arrays and role-labelled transcripts, ordering messages by timestamp and normalising user/assistant/system roles.
Stable labels keep the same placeholder for the same value, partial masking leaves a recognisable tail, and remove strips the value outright.
The downloadable JSON carries only per-category counts, message totals and flagged-message counts — the transcript stays in a separate file you control.
No. Both the file you open and the text you paste are parsed and scanned by JavaScript inside the page, and no request carries the conversation off the device. That also means the scan works with the network disconnected.
Up to a 2 MB file, 2,000,000 characters of source text, and 5,000 messages. Anything past those limits is cut off and the results panel shows a truncation warning, so a very large export should be split before scanning.
Stable labels replace each value with a consistent placeholder so you can still tell two mentions apart; partial masking keeps a fragment of the original so a human reviewer can recognise which record it was; remove deletes the value with no trace. Pick labels when the transcript still has to make sense, remove when it must not.
No — treat it as a first pass, not a guarantee. Pattern matching misses unusual identifiers, personal details implied by context, and anything inside images or attachments, so read the full redacted transcript yourself before sharing. If a conversation contains live credentials, rotate them rather than relying on redaction.