How AI Twin uses AI

• Classifies what you capture. When you save something, an AI model proposes what it is (a person, an event, a fact, a document) and how it should be filed. Nothing is saved to memory until you confirm it.
• Transcribes your voice. Voice notes are converted to text by a speech-recognition service so they can be reviewed like any other capture.
• Writes Briefs and chat replies. When you ask for a Brief or chat with your Twin, an AI model writes the answer from your confirmed memory, with citations to where each point came from.
• Suggests, never decides. Duplicate matches, conflict resolutions and form prefills are proposals. You confirm, edit or reject them.

To do this work, the content involved goes to the model and speech-to-text providers listed on our sub-processors page, each under a data processing agreement.

• It does not save anything silently. Every capture lands in Pending Review first.
• It does not act on your behalf in other services. AI Twin has no autonomous writeback.
• It does not profile your personality. Style captures how you like things written, never who you are.
• It does not use your memory to train AI models without your explicit consent.
• It does not recognise emotions, sort people into categories from biometric data, or make deepfakes.

We do not train AI models on your data. The model providers we use are on commercial terms under which customer content is not used to train their models. Their terms are cited, with the date we checked them, on our AI transparency page.

We check these terms and update the date shown when we do.

• Every chat reply carries an AI label that stays visible in every view, and every Brief is labelled as AI-generated.
• Exports, including conversation exports, carry the same disclosure line.
• Answers in Help that were written by AI say so.
• In the app, you can find this under Settings, then AI use.
• Your own words stay yours, sourced to you.

Every memory item carries its source, its sensitivity tier and an audit entry. Briefs cite the items they drew on, so you can check each point against where it came from. When you share memory with someone, the page they see states the sensitivity ceiling that was applied.

The demo on our homepage uses a small model in your browser to rank three sample memories. It does not write anything, and nothing you type into it leaves your device. Our trust page shows how to check this for yourself.

AI models make mistakes. They can misclassify, mishear a voice note, or compose an answer that overstates what your memory holds. That is why confirmation gates exist, why we never claim perfect recall, and why every Brief carries citations you can check. If something is wrong, it can be corrected, superseded or forgotten.

Article 50 also asks that AI-generated content carry a mark that software can detect, as well as a label people can read. That duty has applied since 2 August 2026, with a transition to 2 December 2026 for systems already on the EU market before that date. Machine-readable marking of what AI Twin writes is not yet in place and we do not claim it. Until it is, the labels described above are how you can tell. When it is settled, this page will say what is marked and how.

AI Twin Limited is a private company registered in Scotland, company number SC866947. Registered office: 272 Bath Street, Glasgow, G2 4JR. Registered with the Information Commissioner's Office, registration ZC044489.

The categories of provider behind the AI are listed at /sub-processors, with each one’s role and region. AI Twin is built to UK GDPR and EU AI Act standards.

The dated EU AI Act timeline is on our UK GDPR and AI memory page.

Questions or concerns: hello@ai-twin.co.uk. Our complaints procedure is at /complaints.