Politicians Are Now listening. Brace Yourself
The reaction of government to disruptive technology follows a foreseeable pattern. First is the stage of muddled observation – what is this thing? Then there is the event that makes headlines and political pressure. Then there is legislation, often written by individuals who grasped the technology a half-year ago better than they do today, and almost always using structures designed to do something different.
The generation of AI videos is in the middle or third stage
The accidents are adding up. Misinformation campaign Synthetic video in elections. Intimate imagery that is not consensual and created and shared in large amounts. Mimicking of real figures and their utterances. Every episode brings about political urgency. The law is arriving, bit by bit, jurisdiction by jurisdiction, more or less quickly, and more or less sensibly, than in any other place you might choose to look.
The uncensored category of the ai video generators is going to bear the brunt of the regulation more directly than commercial tools used by mainstream. Not that it is necessarily more harmful, which is a disputed fact, but because it presents less resistance to the uses that produce the hottest political heat. Both will be pointed to by regulators who can indicate particular harms and particular tools.
It is not optional to know where regulation is going, what it is likely to achieve, and where it will cause any pain to legitimate creative work, should you be establishing any type of serious workflow around these tools.
The Three-Track Regulatory Approach
The regulation of AI video generation is not occurring as a coherent and single act of legislation. It is growing on at least three parallel tracks that intersect in such a way as to produce both clarity and confusion.
The track that is the most developed is harm-specific legislation. Non-consent intimate image laws, deepfakes of minors, and election-related synthetic media laws have been enacted or are under consideration in the United States on a state level, in the European Union, in the United Kingdom, and in several major Pacific jurisdictions. These laws are precise, practical and to a great extent justifiable. They focus on types of injuries that are sufficiently clear to be enforced. They do not need regulators to know diffusion models.
The second track is platform liability frameworks. They operate by imposing legal liability on the platforms hosting or facilitating AI-generated content instead of the tools or individuals generating such content. The incentive system that this would generate is substantial – the liability exposure of the AI-generated content that they host will lead to aggressive restraint, since the cost of not restraining aggressively is legal liability and the cost of restraining aggressively is frustration by creators, and neither of those is visible in a legal case.
The most general AI governance structures are the most general track. These seek to control AI systems as a type, impose transparency and audit responsibility and risk categorization on AI tools depending on their harm potential. The most developed is the EU AI Act. These standards present compliance overhead to AI tool developers and have an indirect impact on what capabilities are incorporated in publicly available products.
What Local Model Operators Do Face
This is the regulatory aspect that the local model eliminates all restrictions framing usually sweeps under the carpet.
Operating an uncensored model locally eliminates platform enforcement. It does not eliminate legal responsibility of outputs. Whatever you create, particularly what you share, is liable to the laws of the place you and your audience are. The permissiveness of the tool does not establish a legal exception, it simply eliminates the platform as an intermediary that could have been able to intercept or filter problematic outputs prior to you making them.
This is where this bites most in the harm-specific legislation track. The non-consent intimate imagery laws are not interested in whether you were on a commercial platform or on an open-weight model locally operated. They are concerned about the content and how it is distributed. Criminal sanctions applicable to creating and distributing synthetic intimate images of actual individuals are applicable regardless of whether the creation occurred on a server in Silicon Valley or on a GPU in your spare bedroom.
It is not a fright gimmick. It is an account of present legal fact in more and more jurisdictions, and the legislative trend is evidently towards greater coverage, rather than less.
The Disclosure Regulation That Really Is Coming
Among regulatory trends that are most likely to impact creative processes in the near future, the compulsory disclosure of AI-generated content is the one that is progressing at the quickest pace and with the most legislative traction.
A number of jurisdictions already mandate the labeling of AI-generated content in political ads. The demands are growing. EU, US and other proposals would broaden disclosure requirements to commercial content, entertainment that is disseminated on large platforms, and news or documentary content that is presented as factual.
The disclosure question is especially acute in case of creators who apply the ai photo to video workflow, i.e. animate real photographs into fake video. The original picture is actual. The video is artificial. What is proper disclosure? That is not yet settled, but the regulatory trend is that disclosure requirements will be targeted to address this case, and non-disclosure will have significant implications.
One of the pieces is platform-level disclosure. Creation-level disclosure requirements are evolving independently. The practical implication: mark your synthetic material, in a clear and consistent manner, whether any existing rule mandates this in your jurisdiction. The regulations are approaching. It is cheaper to get ahead of them than to react to them.
In Any Case Creative Freedom Arguments Work
Not every suggested AI video regulation is a good regulation. Part of it would seriously hobble legitimate creative work and the creative freedom case has a real basis in certain regulatory settings even though it may have a lower basis in others.
The most prevalent issue with draft AI video legislation is overbroad harm definitions. When the laws put forward describe the types of content prohibited in a way that would include horror fiction, war documentary, historical reconstruction and satirical content as well as the truly harmful uses of such laws that the proponents are attempting to combat, creators and civil liberties activists have valid reasons to protest. The quality of AI-specific legislation drafting has been inconsistent, and certain proposals have contained over broad definitions that failed to pass the test of scrutiny they warranted.
Most democratic legal frameworks have artistic expression protections which offer some avenue of protecting creative work that is not covered by the categories that AI video regulation is mostly addressing. There has been a general, but not consistent, recognition by courts that creative and journalistic work of hard content has different grounds than exploitative content of the same kinds. This case law is important to the interpretation and application of AI video regulation.
Those arguments will best shield the creators who can show artistic intent, creative context and responsible treatment – clear labels, proper distribution channels, sensitive treatment of the moral aspects of their work. The argument of it being art is one that is reinforced when all around the art is in favor of the argument.
The Jurisdictional Patchwork Problem is Here to Stay
The regulatory fragmentation between jurisdictions is at worst increasing, not decreasing.
This implies that creators that distribute AI video content abroad will find themselves subject to the strictest jurisdiction that may apply to them, rather than their own jurisdiction. A creator located in a jurisdiction that has few rules regulating AI video distribution that disseminates such content to audiences in a jurisdiction with stricter rules is bound by the rules of the target jurisdictions in a manner that could be practically impossible to enforce, but theoretically could be applicable.
The answer to that is not to waste all your time in comparative law. It is to work in a manner that is conservative enough to be justifiable within the major regulatory frameworks that you are practically exposed to, and which largely depend on the location of your audience and where the platforms on which you are working are integrated.
The Open Technical Question Regulators Haven’t Responded
Most AI video regulation proposals have a technical issue underlying them: how can you reliably detect AI-generated video?
Existing detection tools contain significant error rates. They are most effective in fighting the products of a previous, poorer generation and lag behind as models become better. The most technically promising, but necessitating industry-wide use as verification mechanisms, are provenance systems cryptographic signing of authentic footage at capture.
The premise upon which a regulation based on the assumption of reliable AI video detection is possible is not true now. This does not render the regulation wrong in its objectives. It is true that it will not be perfect enforcement, and that much of the compliance cost will be imposed on good-faith actors who label their content in a transparent way and bad actors who do not label their content evade detection.
The asymmetry, in which it is simpler to impose disclosure requirements on compliant creators and more difficult to impose disclosure requirements on bad actors, is a real structural issue in AI content regulation that has not been solved. It is better to speak the truth, even in favor of the overall tendency to make disclosure and make platforms responsible about the content they host and facilitate.
The regulation is forthcoming. Part of it will be good. Part of it will be out of calibration. It will all influence the development of AI video tools both censored and uncensored, their usage and what can still be offered to creators who are already making serious work in the field.