The AI video company is adding moderation infrastructure built for AI-generated content, doubling down on a screen-at-creation model it has run since 2017.
Synthesia decides whether a video is allowed before it exists. The London AI video company, which lets users generate avatar-led clips from a script, announced on 4 June that it is extending its trust and safety stack by partnering with Cinder, a moderation-infrastructure firm built for AI-generated content. The deal reinforces a model Synthesia has run since its early days: judge the request, not the finished file.
That ordering is the distinctive part. The default for online platforms has long been detection after the fact, hosting content and waiting for it to be flagged. Synthesia assesses every script against its policies at the point of generation, before the model renders a frame.
Cinder slots into that workflow as what the company describes as an in-house agent making a second pass on every model decision, gathering context and escalating to a human reviewer only when there is a genuine judgement call. Each reviewer action retrains the system, and Cinder’s classifiers support more than 100 languages out of the box.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The case for the partnership is volume. In 2025, Synthesia’s automated tooling reviewed more than 11.5 million pieces of content and removed 841,957 that violated its policies, with manual review handling a further 382,792 items and removing 70,272.
Automated review volume grew roughly 77% year on year, from 6.56 million items to 11.58 million, while content reaching human reviewers fell from 792,586 to 441,086. The pattern is deliberate: push the obvious cases to machines and reserve people for appeals and ambiguity. The firm processed 12,450 user appeals in 2025, reversing about 31% on review.
Synthesia frames safety as a commercial asset rather than a cost. Its customers skew heavily enterprise, including more than 60% of the Fortune 100 across financial services, healthcare, defence, and the public sector, buyers for whom a platform that can be turned into a deepfake engine is unusable. The company holds ISO 27001, ISO 42001, and ISO 27701 certifications, and publishes an annual responsible-creation report it calls Futuresafe.
A recent external test gave the pitch a concrete edge. When Belgian public broadcaster RTBF investigated AI video tools for a report titled IA, la fabrique à arnaques (AI, the scam factory), Synthesia blocked the creation of non-consensual deepfakes, political propaganda, racist content, and crypto investment scams, while several rival platforms generated all of them.
For a company that has built its enterprise positioning on refusing certain prompts, an independent broadcaster documenting the refusal is the kind of validation marketing cannot buy.
The Cinder tie-up is, in the company’s telling, about keeping that record intact as the product accelerates. As Synthesia ships new avatar types, languages, and surfaces, its trust and safety team wants moderation that moves at the pace of the roadmap rather than lagging it.
Whether the combined system holds up under the scale Synthesia is targeting is a question only its future Futuresafe reports will answer. The bet underneath the announcement is older than the partnership: that in AI video, the safest moment to say no is before anything is made.
