Personal Information Protection for Cloud AI

Before your question and the matching passages from your documents go to cloud AI, Elephas finds many common personal details on your computer and replaces them with placeholders. When the answer comes back, Elephas puts the real words back in, on your computer. Protection works in seven languages: English, German, Spanish, French, Hindi, Japanese and Chinese.

​

Before you start

• The private-information models must be downloaded. They download during first-run setup and appear in Settings › Downloads as Private-information models. See First-run downloads.
• In a Team workspace, the Team’s AI policy decides whether redaction runs, not your personal setting. See Team AI policy.

​

Steps

​

Turn protection on or off

The steps are the same on Mac and Windows.

​

Always hide specific names

Automatic detection can miss things. To make sure a client name, project code name or other term is always hidden, add it to the Brain’s Protected Names: “Names and terms listed here are always replaced with placeholders before this Brain’s content is sent to cloud AI.” See Protected Names.

​

Check what was hidden

Under each answer protected this way, a row shows how many details were hidden, for example 3 details kept on this Mac · Review (this PC on Windows). Click it to see the list. See Review hidden details.

​

How it works

​

What Elephas hides

Elephas looks for these kinds of personal details. The names below are the kinds of details you may see in the hidden details list. Plus anything you add to a Brain’s Protected Names.

​

What Elephas does not hide

These are left as they are, so the AI can still answer questions about amounts and timelines:

• Ordinary dates, such as a contract date or a due date. Only dates of birth are hidden.
• Years and periods, such as “2025”, “Q1 2024”, “FY2025”, “monthly” or “today”.
• Money amounts, such as $1,250.00 or 3.400,50. Long ID-like numbers, such as account numbers, are still hidden.
• Gender.
• Job titles and honorifics, such as “Dr.” or “CEO”, so the AI can still understand who is who.

This means dates and amounts in your documents reach cloud AI. If that matters for a document, add the terms to Protected Names or use Local AI.

​

How it runs on your computer

Detection happens on your computer, before anything is sent. Elephas uses three methods together:

1. Protected Names you add to a Brain. Matched as whole words, ignoring case and accents.
2. Patterns for structured details such as email addresses, card numbers, IBANs and phone numbers.
3. An on-device model that recognises names, organisations, places and similar details in running text, in all seven supported languages. It runs locally and is downloaded once during setup.

Elephas checks your question, the chat history it sends, and every passage it sends. Each detail is replaced with a placeholder such as [PERSON-3fa2c1]. The same name always gets the same placeholder within a Brain, so the AI can still tell that two passages are about the same person. It does not see who that person is. The link between placeholders and real words is kept in memory only. It is never saved, synced or written to a log.

​

Languages

Elephas decides the language of each document on its own. A document that is mostly German with a few English quotes counts as German. Tamil documents can be searched, but not protected yet. See Search your documents in other languages. With protection on, if a cloud question would use a document in a language Elephas can’t protect, or a document with too little text to tell its language, Elephas sends nothing and shows: “Some of the selected sources are in a language Elephas can’t protect yet, so nothing was sent.” Use Local AI for those documents, or turn protection off if you accept sending them as they are.

​

After you update to 1.0.2

Elephas re-checks your existing documents in the background. As a rough guide, that takes about an hour per 1,000 five-page documents. Until a document is re-checked, documents in the newly supported languages are still refused for protected cloud questions. Nothing is sent unprotected while this runs.

​

Protected questions take longer

Detection takes about a second per page of text in English and German, and up to about two seconds for Chinese. Long chats take longer, because the chat history is checked too.

​

If the models are not ready, nothing is sent

If protection is on but the private-information models have not finished downloading, Elephas does not fall back to weaker protection. It stops and shows: “Elephas can’t protect personal information until it finishes downloading. Nothing was sent. Open Settings to finish setting up, or turn the protection off.” If protection fails for any other reason, Elephas shows: “Personal information couldn’t be protected for this answer, so nothing was sent.”

​

Limits you should know

• It detects the large majority of personal information. No automated detection is perfect. Some details will be missed. Unusual names, nicknames, and details that only identify someone in context (a job title plus a small town, for example) can slip through. Company names are the most likely to be missed when Elephas searches in several steps for one answer.
• It sometimes hides things that are not personal. That is the cautious side of the tradeoff. The real words are restored in your answer.
• The topic of your question is still sent. Placeholders hide who and which, not what. The AI still reads the rest of the passage.

For material where any miss is unacceptable, use Local AI. In Local AI mode, your question and your documents are not sent to any AI provider.

​

Troubleshooting

​

FAQ

​

Related articles