Combat AI hallucinations Cover your blind spots

AI models learn from huge amounts of text, in a process called training, and the conversations people have with an AI can become part of that text. So a company that makes AI models, such as OpenAI, Anthropic or Google, has a use for what people type.

Safety First AI isn’t an AI model company. We build tools that help you use AI safely, and training AI models isn’t part of our business. We don’t train Konfida, the private model we run on our own hardware, either. So we have no reason to collect your questions, your documents or anything else you type to use as training data.

When you ask an AI for help, your text goes to the company that runs the AI, such as OpenAI, Anthropic or Google. In Document Analysis and Report Writing, Safety First AI replaces the personal details in your text before that happens, so the AI company never sees them. Personal details, also called PII (personally identifiable information), include names, dates, addresses, phone numbers, email addresses and organizations.

When you paste your text, a detection service that we run ourselves finds those details, without keeping a copy of your text. Each detail is then swapped for a placeholder, such as “Person 1” or “Date 1”. Before anything is sent, you check every placeholder and can add or remove any. Only the version with placeholders goes to the AI.

In Report Writing, the AI’s edits come back with the same placeholders, because the AI never saw the real details. When you’re done, one click puts the real names and details back in place of the placeholders.

In Document Analysis and Report Writing, only your latest document is saved, and it’s saved in your web browser, on your own device, not on our servers. Earlier documents aren’t kept.

Multi-AI Chat doesn’t save your chats anywhere unless you choose to. An unsaved chat is gone when you close the browser tab. If you choose to save a chat, it’s stored on our servers, so you can come back to it later.