The privacy conversation around AI and personal images spent years focused on still photos. Deepfakes of faces. Non-consensual intimate images generated from selfies. Facial recognition tagging in family albums. These were real and serious harms — but they operated in a relatively bounded domain.
That boundary has moved.
In February 2026, sixty-one data protection authorities from jurisdictions including the European Union, Canada, Australia, and the United Kingdom issued a joint statement. The subject: AI systems now capable of generating highly realistic videos depicting identifiable individuals without their knowledge or consent. The authorities described this as an urgent concern requiring coordinated regulatory response.
Statements from sixty-one regulators don’t happen without something significant happening. Here’s what changed, and why it matters for how you store and protect your personal visual archive.
The Capability Threshold That Shifted
For years after deepfake video tools first appeared, the output was detectable. Faces flickered at frame boundaries. Heads moved unnaturally. Skin texture degraded under movement. Detection software could identify synthetic video with reasonable accuracy, and the quality gap between synthetic and authentic footage was visible to a careful human eye.
That’s no longer consistently true.
Systems like Google’s Veo, OpenAI’s Sora, and Runway’s Gen-3 now produce temporally coherent synthetic video approaching broadcast quality under many conditions. Facial consistency across frames has improved dramatically. Motion fidelity — the way a generated face moves while speaking or turning — has crossed a threshold where synthetic content becomes difficult to distinguish from authentic footage under ordinary viewing conditions.
The practical implication: video generation no longer requires video as input. These systems can produce realistic synthetic video from still photographs. Anyone with a collection of photos of you — from your social media profiles, a public website, or any platform that makes images accessible — has potential raw material for generating synthetic video of you.
The gap between “someone has your photos” and “someone can generate realistic video of you” has closed.
Why Sixty-One Authorities Acted Together
Coordinated international regulatory statements at this scale are uncommon. They typically follow events that authorities across jurisdictions agree are urgent enough to require simultaneous public response.
The February 2026 joint statement identifies several categories of harm that drove the response:
Non-Consensual Intimate Imagery
NCII using AI-generated video has become a growing crisis affecting predominantly women and girls. The February statement noted particular concern about vulnerable groups, including minors. Unlike earlier NCII that required stolen authentic footage, video AI can now generate convincing material from publicly available photos — photos shared on social media, class directories, public profiles, or any platform without strict access controls.
The key change is scale and accessibility. Previously, creating convincing synthetic intimate video required significant technical expertise. Current video AI tools have democratised the capability to the point where technical skill is no longer a meaningful barrier.
Identity-Based Manipulation
Synthetic video can show an identifiable person committing crimes, making inflammatory statements, holding political positions they don’t hold, or behaving in ways calculated to damage their reputation, career, or relationships — fabricated entirely from photos they shared years before these capabilities existed.
This isn’t a hypothetical concern. The authorities referenced documented cases of manipulation in political and professional contexts. As generation quality improves, the difficulty of challenging synthetic video in legal and social contexts increases.
Fraud and Social Engineering
Video-call-based impersonation has become substantially more convincing when backed by high-quality synthetic video. Earlier social engineering relied on voice alone; adding a realistic synthetic face dramatically increases the conviction that a caller is who they claim to be.
Consent as a Structural Problem
At the core of the February statement is a point regulators describe as a “prior consent” problem. Most of the photos now being used to train video generation models were posted before these capabilities existed. The people in those photos made a reasonable decision, at a reasonable time, to share visual content — without any ability to foresee that decision would effectively contribute training data to systems capable of generating synthetic video of them years later.
This consent problem cannot be retroactively solved for data already collected. It can only be addressed going forward.
How Video AI Uses Your Photos
Understanding the mechanism clarifies both the risk and the mitigation.
Video generation models are trained on large datasets comprising video and image content, much of which is scraped from the publicly accessible web. Photos you’ve posted to social media, a public website, a photo sharing platform, or any service that doesn’t restrict access are candidates for inclusion in training data.
Once a model has trained on images that include your face, it develops an implicit representation of your appearance. This representation is used when the model generates content — either in general (where your face might appear incidentally) or specifically (when someone uses a photo of you as a conditioning input to a model that supports it).
Image-conditioned generation — where you provide a source photo and ask the model to animate it, make the subject move, or place the subject in a different context — has become a standard feature of consumer video AI tools. The model takes your still photo and generates video of that face doing whatever was requested.
Stored vs. Posted: A Critical Distinction
The risk described above is almost entirely a function of what you’ve made publicly accessible. Photos stored privately — in an encrypted cloud backup, on your personal device, or in any service without public link sharing — don’t feed web-scraped training datasets. They don’t appear in conditioning inputs unless someone gains unauthorised access to your storage.
This distinction matters practically. The threat isn’t primarily from someone hacking into your private photo backup. It’s from the photos you’ve already made publicly accessible — social media posts, publicly visible profiles, images in image search indexes. Private storage is not what creates this exposure.
What private storage does is prevent the further growth of that surface area. Every photo you store privately rather than posting publicly is a photo that isn’t adding to your machine-readable visual profile.
The Home Video Problem
Beyond synthetic generation from still photos, there’s a separate question about the home videos you already have stored.
Home video is among the most intimate personal data most people create. Birthday parties. Family holidays. Children’s first steps. Private moments between people who trust each other. Most people store this content carelessly — in iCloud, Google Photos, on social media platforms, or on devices that are eventually lost, sold, or recycled.
The platforms that store video have their own terms about how that content can be used. Google Photos’ terms permit use of content to “develop and improve” services. Social media platforms typically claim broad content licences. AI model fine-tuning — where specific footage is used to improve a model’s ability to replicate someone’s appearance, voice, or movement — is an emerging use case that few platforms explicitly prohibit.
Whether home video is being used for AI training purposes isn’t always transparent from reading current privacy policies. Many of these policies were written before the current generation of video AI existed and haven’t been fully updated to reflect new technical capabilities.
The practical question for anyone storing home video is simple: does the platform you’re using have an explicit, specific commitment not to use your content for AI training, and is that commitment legally binding?
What Regulation Currently Covers
The regulatory framework available to people whose video likeness is used without consent is limited and jurisdiction-dependent.
The EU AI Act classifies certain uses of biometric data in AI systems as high-risk or prohibited, but was drafted before the current generation of video AI reached its current capabilities. Enforcement guidance in this area is still being developed.
In the United States, Illinois’ Biometric Information Privacy Act regulates collection of biometric identifiers and has been applied to facial recognition. Texas and Washington have similar laws. Whether and how these apply to video generation training data is being tested in litigation.
The UK’s ICO, one of the signatories to the February joint statement, has published guidance on AI training data but acknowledges the guidance is catching up with capability developments rather than preceding them.
The practical implication: the legal protections available to individuals are not yet reliable or comprehensive. Regulatory frameworks are developing, but they lag the technology by years. The most effective protection today is reducing what’s accessible rather than relying on legal remedies after the fact.
Practical Steps
Audit What You’ve Made Public
Systematically review the photos and video you’ve posted publicly. Social media accounts, public websites, image sharing platforms, photo albums with publicly accessible links. You can’t undo what’s already been scraped, but you can stop adding to the accessible collection.
Most social media platforms allow posts to be restricted to followers or made private after the fact. This doesn’t remove content from already-scraped training datasets, but it stops contributing to ongoing collection.
Use Platform AI Opt-Outs Where Available
Some platforms offer opt-outs from using your content to train AI models. These are inconsistently implemented and of uncertain effectiveness — platforms can use data already collected before an opt-out is set. But they’re worth enabling.
Check your settings on any platform where you’ve stored or posted significant collections of personal photos or video. These opt-out settings often reset when platforms update their privacy policies, so checking periodically is worthwhile.
Choose Storage With Explicit AI Training Commitments
Not all storage providers are equivalent on this point. Look for explicit, specific language in the provider’s privacy policy stating they will never use stored content to train AI models — either their own or third-party systems.
Vague language about “improving our services” doesn’t provide meaningful protection. Specific prohibitions on AI training of user content, with explicit commitments to GDPR and CCPA compliance, offer more reliable protection.
Keep Your Most Sensitive Visual Archive Private
Home video, intimate photos, and images of children warrant a higher level of care than public-facing content. These categories benefit most from storage that is explicitly private, that limits employee access, that doesn’t share content with third parties, and that clearly commits to permanent erasure when you close your account.
The Compounding Nature of This Risk
Here’s why the February 2026 joint statement carries weight: the problem compounds over time.
Video AI capabilities will continue improving. Footage that’s difficult to convincingly animate today will be straightforward in two or three years. A photo collection that provides limited conditioning data for current models provides richer data for future models.
Every decision about what to post publicly and where to store private visual content should account for what future AI capabilities might do with that content — not just what current capabilities can do. The authorities who signed the February statement are signalling that the trajectory is concerning, not just the current state.
Your personal archive of photos and videos represents captured life across years or decades. The decision of where it lives and who can access it — and whether it feeds AI training pipelines — is worth thinking about in those terms.