Does Synthesia train on your data?
Synthesia's policy answers the question for your face and voice — a closed list of purposes, then destruction. For your scripts and finished videos it defers to an agreement your employer signed.
No opt-out found in this policy
"We and our vendors will store and use the Biometric Data described above only until it is no longer needed for the achievement of the above listed purposes."
- 1 Oct 2026
- 5 Oct 2026
- Your likeness (Biometric Data)
- global
- Fetched from the company
- 20E470A4…43C4
How we read it: Regraded 28 Aug 2026, and the correction is ours. The fine-tuning bullet quoted here is one of five enumerated purposes for Biometric Data, and the sentence after the list limits all of them: the data is stored and used "only until it is no longer needed for the achievement of the above listed purposes", then permanently destroyed. General model training is not among the five. This page previously read that bullet on its own and called the question unanswered. It was answered, one sentence further down. Training happens on the Sample you deliberately record, to build the avatar you asked for, and stops when that purpose does.
Biometric Data is processed for, only until, permanently destroy, scripts, output, videos you create
What the document does offer
- the five listed purposes: generating your Avatar, fine-tuning the model for it, verifying identity, evaluating authenticity, and security and fraud prevention
- "After the relevant processing has been completed or where instructed by our Customer, we permanently destroy your Biometric Data"
- "We require our vendor(s) to similarly delete or destroy Biometric Data"
- the scripts you type and the videos you make are covered by a different clause, and by deferral rather than by an answer — see the second tier
Why none of it applies: The closed list and the destruction commitment answer the question for Biometric Data — the face and voice, which is the sensitive part and the reason anyone asks. They do not reach the scripts entered or the finished videos, which the document hands to a separate agreement. B rather than A because this is a description of what Synthesia does, revocable by editing the page it is written on, not a statement that it cannot do otherwise.
The policy names this data and then hands it to a document the reader can't see. It defines Customer Data as the "audio, video, text, and scripts used to prompt Avatars to speak and perform", the files uploaded and the videos produced. Then it says a separate Agreement may govern its processing, that the Customer who signed that Agreement "controls its instance of the Services and any associated Customer Data", and that privacy questions should go to that Customer. For most users the Customer is their employer. So the answer here is a deferral, to terms a user has no right to read and didn't sign. Synthesia told us in writing on 28 Aug 2026 that none of this data reaches pre-training. That's a good answer, and it isn't in any document a user could hold Synthesia to.
- Synthesia privacy policy, captured 2026-08-14, re-confirmed unchanged 2026-08-28, re-read in the revision captured 2026-10-01 (https://www.synthesia.io/privacy-policy)
- Customer Data, separate agreement, Authorized User, train, scripts
- 0 of 0 raw hits
- 2 Oct 2026
No training language attaches to Customer Data anywhere in the capture. The only statement about it is the deferral quoted above. Recorded as a documented deferral rather than silence, because the two are different findings and this page called it silence for a fortnight.
What Synthesia said when we showed them this
Clarified“Synthesia does not develop large language models or general purpose AI models. [...] We pre-train these AI components on performances by paid actors we commission, on performances that are publicly available, and on performances we license. We do not pre-train them on customer data. [...] Beyond that consented fine-tuning step, Synthesia does not use any other customer data, including the inputs to the product or the videos produced with it, to pre-train its AI models.”
What each plan can refuse
The same questions, asked of every plan we could identify. Every cell comes from a quoted clause or a documented search - see the evidence above.
- Nothing to switch off
- Not stated
- 17 / 100
- Not stated
- Unknown
- Not established
- Not stated
- —
- —
- Not established
How to opt out of Synthesia AI training
Plan by plan, from what Synthesia's own policy says about refusing. Where it publishes the steps, they're listed here with the limits they carry. Where there's nothing to switch off, or nothing graded, this says so instead of inventing a menu path.
Your likeness (Biometric Data)
Nothing to switch off: training here is off until an opt-in is given.
Scripts, uploads and finished videos (Customer Data)
Not graded, so no route is pointed at. The policy names this data and then hands it to a document the reader can't see. It defines Customer Data as the "audio, video, text, and scripts used to prompt Avatars to speak and perform", the files uploaded and the videos produced. Then it says a separate Agreement may govern its processing, that the Customer who signed that Agreement "controls its instance of the Services and any associated Customer Data", and that privacy questions should go to that Customer. For most users the Customer is their employer. So the answer here is a deferral, to terms a user has no right to read and didn't sign. Synthesia told us in writing on 28 Aug 2026 that none of this data reaches pre-training. That's a good answer, and it isn't in any document a user could hold Synthesia to.
What this policy has done since we started watching
1 change detected since 14 Aug 2026. Each one is a diff between two archived copies, and each links to the passage that moved.
Everything we read
Vendor response
Their reply is quoted in full at the top of this page.
We wrote back on 2 Oct 2026 with a question that's still open. It's listed on the register.
More design tools on file
All 23 in design →Show this verdict on your site
Synthesia is graded B, and a company that's earned that can say so with the verdict itself. The image below links back to this page and shows the live grade. If the grade changes, the badge changes the same day, wherever it's embedded. It's free and it isn't for sale.
