You can find our deepfakes paper ('Deepfake Defences') with the full research findings on our website

Platforms regulated under the #OnlineSafetyAct have duties to tackle the sharing of certain types of deepfake. This includes any deepfake that amounts to illegal content (e.g. deepfake fraud ads) or content that is harmful to children (e.g. deepfakes intended to bully a target).

We identify four routes by which DFs can be addressed. This includes prevention (e.g. through the use of prompt/output filters in models), embedding (e.g. via watermarking), detection (e.g. via ML classifiers), and enforcement (e.g. by blocking problem accounts).

Tackling deepfakes requires action from all players in the tech supply chain - from the developers creating GenAI models, to the platforms that host this technology, through to the user-facing services that act as spaces for deepfake content to be shared and amplified.

A ‘deepfake economy’ is driving the sharing of this content. This economy includes easy-to-use tools like nudify apps, websites that are dedicated to hosting this type of content (as well as DF manuals), and professional deepfake creators who sell their services to others.

Deepfakes can cause harm in myriad ways. We set out a typology of deepfakes that demean (e.g. non-consensual sexual imagery), defraud (e.g. fake adverts promoting scam investment schemes), and disinform (e.g. political deepfakes that intend to disrupt elections).

The advent of #GenAI is changing the deepfakes landscape. Not only does it allow for the creation of more convincing content, it is lowering the barrier to entry and allowing new types of creators to get involved (including children who are creating deepfakes of one another).

Last month @Ofcom published a new paper on #deepfakes and what we're doing to address them under the new #onlinesafety regime. A short thread of the main messages 🧵.

Today @Ofcom publishes two papers on #generativeAI - one looking at the proliferation of #deepfakes and what to do about them. The other looking at the merits of red teaming as a way of stress testing GenAI models. Have a gander (they're pretty good imho).

RT @Ofcom: ⚠️ Two in five people have seen at least one deepfake in the last six months. But only one in ten are confident in their abilit….

RT @soundboy: “We recently provided Claude 3.5 Sonnet to the UK’s Artificial Intelligence Safety Institute (UK AISI) for pre-deployment saf….

RT @soundboy: 1/ Today the UK's AI Safety Institute is open sourcing our safety evaluations platform. We call it "Inspect": .

Excellent primer from @witnessorg on the promise and pitfalls of deepfake detection tools

RT @Ofcom: The deadline for our consultation on protecting people from illegal harms is just 3 weeks away. 📅. Our proposals cover how servi….

RT @charliebholtz: David Attenborough is now narrating my life. Here's a GPT-4-vision + @elevenlabsio python script so you can star in your….

RT @EastEndTrades: RT: 20 DAYS TO GO! It's make-or-break time! Please help us #TakeHackneyArches into community ownership so young people l….

RT @esme_harrington: Wonderful to speak with @BenedictDel about @Ofcom preparing to become the online safety regulator & conduct algorithmi….

Brilliant reporting from 404 Media on how open source GenAI models are being used to create non-consensual pornography en masse

Great piece from Rana Foroohar on bottom-up, worker-led AI controls (the new US Writers Guild deal being a case in point)

RT @matthewclifford: I’ve had lots of questions about the AI Safety Summit in November. Today we share more detail on our focus and objecti….