Personal data in ChatGPT and other LLMs: FAQ

Which mode should I use before sending material to an LLM?

Already have a document and want to restore details later? Choose mode 1, labels and a key. Want permanent removal, including from scans, or to retain a PDF layout? Choose mode 2, redaction without a key. Asking AI to write a new document from an instruction? Choose mode 3, a prompt with fictional details and a key for restoring the real details in the AI response.

How does mode 1, document pseudonymization, work?

You already have a document with real details and want to work on it with AI. Load a TXT, MD, RTF, DOCX, ODT or text PDF, review the selections, then download a document with labels such as [PSN-1] and a key file. Editable files keep their format; a PDF becomes TXT or DOCX without the page layout. Later, load the AI response and your key to restore the real details.

How does mode 2, document and scan redaction, work?

You want to remove selected details permanently. Load a supported text document or a PDF, including a scan, then review or add selections. TXT, MD, RTF, DOCX and ODT produce the same format with redaction markers. Every PDF produces a new PDF with covered areas and the visual page layout retained. There is no key and the removed details cannot be restored. The remaining content may still identify someone.

How does mode 3, a prompt with fictional details, work?

You are starting a new letter, claim or other document. Type or paste your prompt with real details into gallsec, review the selections and mark anything missed. You receive a prompt with fictional details and a key. Send only that prompt to AI. Then save the AI response in a supported file, load it with your key in gallsec, review the matches and download the response with the real details restored.

How do I prepare personal data before using ChatGPT, Claude or Gemini?

Include only what the task needs. Use gallsec to replace selected details with labels, review the whole text and store the key separately. Send only the reviewed version to your chosen AI tool. gallsec does not connect to those services or decide whether their use is permitted in your case.

Can I upload a contract, claim or case file to AI?

Using gallsec does not itself authorize disclosure. Check organizational rules, confidentiality duties and the chosen service's terms. Where AI use is permitted, labels help prepare an existing document; prompt mode helps draft a new one without sharing selected real details.

What about medical records, CVs and employee information?

Names and identifiers are only part of the issue. A rare diagnosis, work history, location or combination of dates can identify someone. Automated detection does not assess the entire medical or employment context. Review every page and manually mark details that need removing.

How can a browser application work offline?

Your browser runs code on your computer. Once the application and dictionaries are downloaded, gallsec reads files, detects details and generates output locally, without a document-processing server. Loading a file in this editor is not an internet upload. Wait for the ready message before disconnecting.

How can I check whether files leave my computer?

Open DevTools and its Network tab. Wait for readiness and any optional PDF preparation to finish, then clear the request list. Load and process a document. You can also disconnect and repeat the full workflow. The page's CSP includes connect-src 'none'. A separate offline preparation mechanism downloads public resources and receives no documents or keys.

Why must anything be downloaded first?

The initial download contains application code, dictionaries and file parsers. PDF support is an optional package you can prepare before disconnecting. The cache stores these public resources so the tool can start again offline. The download size is shown in the interface. No language model is downloaded.

Will refreshing the page preserve my work?

No. Your document, prompt and mappings are held in the current tab's memory. Refreshing or closing it may lose them. The cache retains the tool, not your session. Download the result and key before leaving. Browsers may also evict cached resources, especially if persistent storage was denied.

How does pseudonymization actually reduce exposure?

If you send [PSN-1] instead of a name, the recipient does not receive that name at that place in the document. The key stays with you. This reduces the direct identifiers given to a service, but remaining context may still identify someone. It reduces data exposure rather than eliminating all risk.

Do real incidents explain why limiting data matters?

In March 2023, OpenAI reported a bug exposing some users' chat titles to others. The ICO also described PSNI employee data exposed in a hidden spreadsheet tab. One concerns an online service, the other unseen file content. Neither proves every AI use causes a leak. Both illustrate reasons to review what you share, including hidden document elements.

OpenAI: March 20 ChatGPT outage (2023-03-24) · ICO: PSNI spreadsheet disclosure (2024-05-23)

Does pseudonymization or redaction make a document fully anonymous?

No. Pseudonymization is reversible using additional information, here the key file. The data remains personal data. Redaction removes selected passages without a key, but remaining details may still identify someone. The same applies to the third mode, prompt pseudonymization.

EU 2016/679, Art. 4(5), Recital 26 (2016-04-27)

What is in the key, and should I send it with my prompt?

The key contains real details and their mappings to labels or fictional values. Do not send it to an LLM or to the recipient of the prepared version. Download it and protect it separately. Password encryption is enabled by default; losing the key or password prevents restoration through the tool.

Can I restore details in a completely new document written by AI?

Yes, with label mode or prompt mode. Load the model's response and the matching key, even in another browser or a fresh session. The filename can differ. gallsec shows matches, missing entries and decisions needed. It cannot recover information the model omitted or changed beyond recognition.

What if AI translates a label or changes a fictional name's grammatical form?

Labels support tolerant matching and selected translations in eleven languages. Fictional-data keys contain observed variants and selected inferred name forms. Uncertain restorations need confirmation; an unrecorded grammatical form, a new surname or a first name alone may remain unmatched.

Can I black out a selectable-text PDF without changing the layout?

Yes, in redaction mode. gallsec rasterizes the pages, burns in coverage and writes a new image-based PDF. The visual layout remains, while the original text layer disappears. Scans use the same mechanism. Label mode instead exports PDF text to TXT or DOCX without preserving the layout.

Will a scan be automatically read with OCR?

Yes. Redaction and document-label modes run on-device OCR (Tesseract in the browser) on pages with fewer than 20 text-layer words. OCR is in this language version's PDF package (one model, English on this site). You can turn it off before processing. Recognition can err and is not guaranteed; review every scan. Prompt and restore modes do not use OCR.

Can personal information remain in metadata or tracked changes?

Yes. gallsec therefore processes supported document parts and removes author metadata. It warns about changes, comments and embedded items. It cannot read text inside DOCX or ODT images. RTF containing changes requires prior cleanup. TXT and MD are easier to inspect, although protected Markdown code blocks remain unchanged.

What if the tool missed personal information?

Select a passage in the preview and add your own pseudonymization or redaction. You can mark project names, amounts, dates or whole descriptions. Use rectangles on PDFs. Detection is a starting point; downloading requires your active confirmation that you reviewed the material.

Can I work with documents from different countries?

Each language version of the site uses only that language’s detection pack. This English homepage covers the United States / United Kingdom (SSN / NINO / NHS and US/UK name dictionaries); other English-speaking countries will follow. Other packs: Polish /pl, French /fr, German /de, Spanish /es, Italian /it, Chinese /zh, Hindi /hi, Portuguese /pt, Bengali /bn, Japanese /ja. Street names are empty in this pack (no national street gazetteer). Review every page before sharing.

Is gallsec free for law firms and businesses too?

Yes. All three modes are free for everyone, including commercial users. There are no accounts, subscriptions or document-count quotas. Technical file and device-memory limits remain: the current version accepts files up to 64 MB and places additional limits on unpacked office archives.

Does gallsec guarantee GDPR compliance or complete detection?

No. Kantante provides a tool to assist the user. It does not guarantee all personal data will be found or take responsibility for the completeness and correctness of the user's pseudonymization. The material, authorization to share it and final result still require human assessment.

Does offline processing also protect against browser extensions?

No. gallsec restricts transmission by the application, not everything on the computer. Extensions with page access, malicious software or clipboard synchronization may operate independently. Use a trusted device. Once you copy the result to another service, that service's processing rules apply.