With AI image generation tools like Stable Diffusion and DreamBooth, it is easy to make life-wrecking deepfakes with a few photos of a person from social media
and the same process can apply to real people with just a few real photos pulled from social media: https://arstechnica.com/... https://twitter.com/... Susan Bell / @susanbellair : Worrisome read. We're gonna get to a point where nothing can be confirmed as “real” unless we're physically present. https://twitter.com/...
Context & Ripple Effects
The deepfake problem has been compounding for years: researchers flagged an AI service that turned a single uploaded face into nonconsensual deepfake porn back in 2021, and creators were already publishing YouTube tutorials as the tooling got cheaper in 2020. What Stable Diffusion plus DreamBooth changes is that fine-tuning a model on a specific person now takes only a few scraped social media photos on commodity hardware — no dedicated service or specialist skill required.
That lands on top of an older warning from coverage of the liar's dividend: once fakes are common enough, bad actors gain cover to dismiss real evidence as fabricated. Susan Bell's reaction in the article — that nothing will be confirmable as real without physical presence — is the endpoint of exactly that dynamic, now accelerated by open-weight image models rather than closed services.
First-order effects
- Private individuals become directly targetable: anyone with a handful of public photos can be dropped into fabricated imagery, extending the harm pattern first documented with face-swap porn services to anyone with a social media footprint.
- Stability AI and Google (DreamBooth) face immediate reputational and policy exposure, since their own released tools are the ones lowering the cost of targeting real people.
Second-order effects
- Platforms hosting model weights and outputs are pushed toward stricter likeness policies and takedown workflows, while demand rises for detection tooling of the kind DARPA was researching when deepfake videos first went mainstream.
- The same capability gap pressures social networks to treat unverified imagery of people as presumptively synthetic, changing how ordinary users share and authenticate photos.
Third-order effects
- If fabrication stays this cheap, provenance at capture time — cryptographic attestation of real images — shifts from nice-to-have to core infrastructure, because post-hoc detection alone cannot keep pace with open-source generation.
- The pattern points toward formal likeness governance: legal and platform-level rights over one's own face and voice, analogous to how copyright fights forced training-data consent onto model makers.
The trend: Consumer-grade generative models are collapsing the cost of fabricating a person's likeness, forcing authenticity infrastructure — provenance, detection, and likeness rights — to be built after the fact rather than before deployment.