Many Voices · Avatar Impact Stories
Every presenter on this wall is synthetic. Every account they give is not. This page explains where the stories came from, who made them, what tools were used and when — and where the boundaries were drawn.
Survivor testimony is the most persuasive evidence there is, and it is also the most dangerous thing a survivor can give. Publishing a real face alongside an account of trafficking or forced labour can expose someone to retaliation from the very people they escaped.
The conventional workarounds all cost the thing that makes testimony land. A silhouette removes the face. Voice modulation removes the voice. An actor's dramatic reading replaces the person entirely. A synthetic presenter breaks that trade-off: the account stays first-person and human, while the identity behind it stays protected.
The project leverages generative AI to create realistic avatars that deliver authentic survivor testimonies. This technology allows for the sharing of crucial, attention-grabbing stories while protecting survivors' identities, demonstrating how AI can be ethically applied to address pressing social issues and create empathy without compromising safety or dignity. — Generative AI for Good
"Many Voices" was inspired by the Generative AI for Good programme and the work of Shiran Mlamdovsky Somech, whose advocacy for applying generative AI to social impact is the reason this project exists.
The avatars were generated in Fall 2025 using D-ID, which drives a photorealistic presenter from a still image and a voice track.
The date matters more than it normally would. Avatar generation is moving fast enough that a few months is a visible generation gap. What you are watching is a snapshot of what this tooling could do in late 2025 — not a demonstration of the current state of the art.
That shows up in specific places: lip-sync drifts on harder consonants, head and hand movement is limited, gaze is fixed, and emotional range is narrow compared with what newer models produce. Those artefacts are left in deliberately. They are part of the record of when this was made, and they help a viewer tell that the presenter is synthetic.
| Programme | Generative AI for Good, Fall 2025 cohort |
|---|---|
| Advisor | Shiran Mlamdovsky Somech |
| Avatar tooling | D-ID, as available in Fall 2025 |
| Stories | 21, running roughly 40 seconds to 2 minutes each |
| Subjects | Human trafficking, forced and child labour, domestic violence, forced conscription, cobalt mining |
The work was done by students in the Fall 2025 cohort, advised by Shiran at Generative AI for Good. They were taught not just the tooling but the craft and the ethics around it. Each story went through the same path.
Students sourced survivor accounts and other first-person material, and stayed faithful to them. The requirement was that the testimony be traceable back to a real source — a synthetic presenter is a way to protect a real account, not a licence to compose one.
An avatar holds attention for a much shorter window than a human speaker does. Testimony had to be tightened to that budget without flattening it — which mostly meant cutting context and keeping specifics, since it is the concrete detail that carries a first-person account. The finished stories run roughly 40 seconds to 2 minutes.
Presentation, language and delivery were matched to the community a story comes from rather than defaulting to a Western frame. The wall includes stories in and from multiple languages and regions. The presenter's appearance and voice were treated as part of the testimony's accuracy, not as a styling choice.
Students learned to generate and direct a presenter in D-ID, and — just as importantly — the practical limits of what one can carry convincingly. Longer scripts, rapid emotional shifts and complex names were where the Fall 2025 tooling showed its seams, so scripts were written with those limits in mind.
Each finished story is a matched pair: a short silent loop that sits on the idle wall, and the full video with audio that plays when a visitor taps it. Nothing on the wall plays sound until someone chooses it.
A note on the content. These accounts describe trafficking, forced and child labour, violence and conscription. They are presented plainly, without dramatisation. Please take care.
The wall is built for a room, not a browser tab. Tiles loop silently over a crowd backdrop — the visual argument being that these accounts are everywhere and mostly unheard. Nothing plays until someone chooses it, which makes engagement a deliberate act rather than something autoplayed at a passer-by. When a story ends, the wall returns to its idle state and waits.
The source, the build notes and the full technical write-up — including how the media is encoded and cached so stories start quickly on an ordinary connection — are in the repository.