When
actor Bryan Cranston successfully pushed OpenAI to tighten consent controls
inside Sora 2, its AI video generation tool, the outcome was covered as a
good-news story about a famous performer winning a reasonable argument. What
it actually represented was something more structurally significant: the
first time a creative worker had shifted an AI company’s platform governance
position through organised professional advocacy rather than legislation,
regulation or litigation alone.
That distinction matters for everyone whose likeness, voice or
creative output is at risk of being replicated by AI systems that are
becoming simultaneously more capable and more accessible. The deepfake
problem has become a mainstream policy challenge faster than anyone in the
entertainment industry anticipated, and the Cranston moment established a
template that other professional groups are already examining for their own
situations.
What Cranston Actually Won
Cranston’s specific concern was that Sora 2’s consent
architecture, as initially designed, allowed users to generate video content
using realistic likenesses of identifiable individuals without first
obtaining consent from those individuals. The platform operated on an opt-out
model: harm could occur unless specific individuals had been proactively
added to a block list, which placed the burden of protection on the person at
risk rather than the person seeking to use a likeness.
Cranston raised the issue through SAG-AFTRA, Hollywood’s principal
performers union, which had been tracking AI-generated likeness concerns
systematically since the 2023 strike. OpenAI’s response was to shift Sora 2’s
default consent architecture toward a model that requires affirmative consent
signals before recognisable individual likenesses can be generated. Trade
publications Variety and Deadline confirmed the architecture change. Business
Insider reported that OpenAI simultaneously threw its support behind the
proposed NO FAKES Act, the federal legislation that would create a private
right of action for individuals whose AI-generated likenesses are used
without their consent.
The consent architecture shift is the detail that carries the most
long-term significance. Platform defaults determine the majority of user
behaviour at scale. An opt-out system allows harm to occur and then requires
remediation after the fact. An opt-in system prevents the harm before it
occurs. The difference between those two states, operating across a platform
serving millions of users, is not marginal. It is the difference between a
governance framework that functions and one that requires constant
enforcement against a volume of violations that no moderation team can keep
pace with.
The Research That Explains Why the Stakes Are
High
The urgency behind Hollywood’s response to deepfake governance is
not only about income protection or reputation management for individual
performers, though both are real. It is also driven by research showing that
deepfake content causes durable belief change even when audiences are
explicitly told they are watching fabricated material before they watch
it.
A study published in January 2026 by researchers Simon Clark and
Stephan Lewandowsky at the University of Bristol, published
in a peer-reviewed psychology journal, found across three
pre-registered experiments involving 673 participants that most people relied
on the content of a deepfake video even after being explicitly warned it was
fake. Participants updated their beliefs about the people depicted based on
the video content even when they reported believing the warning. The effect
was present and significant even for individuals who said they knew the
material was fabricated.
That finding has profound policy implications that are not yet
reflected in most current legislative approaches. If warnings do not
neutralise the influence of deepfake content, then regulatory frameworks that
focus primarily on labelling, disclosure and transparency are addressing a
problem that may be structurally unsolvable through disclosure mechanisms
alone. Prevention at the point of creation, platform consent architecture
requirements and criminal liability for production are the policy tools that
the research evidence actually supports. Labelling is necessary but not
sufficient.
Where the Law Stands as of October 2025
Legislative progress on deepfakes in the United States has
accelerated substantially in 2025. The TAKE IT DOWN Act, which makes it a
federal offence to knowingly publish non-consensual intimate imagery
including AI-generated synthetic images, passed the Senate and House with
strong bipartisan support and was signed into law in spring 2025. It requires
platforms to remove flagged content within defined timelines and gives
affected individuals a statutory basis for demanding removal that did not
previously exist at the federal level.
The NO FAKES Act, which addresses a broader category of
AI-generated likeness use including entertainment, commercial and political
applications, was still under active legislative consideration as of October
2025. SAG-AFTRA, major talent agencies and a coalition of entertainment
industry bodies have been among its most consistent advocates, arguing that
federal legislation is the only mechanism capable of creating a consistent,
enforceable consent framework across state jurisdictions that currently vary
considerably in scope and enforcement capability.
In Europe, the EU
AI Act, which requires visible labelling on AI-generated content
depicting identifiable individuals, was in active national implementation as
of autumn 2025. Enforcement responsibility sits with national competent
authorities, which means the practical experience of those requirements will
vary across member states for the foreseeable future. For a broader picture
of what AI governance frameworks are doing in practice across organisations
in 2025, our analysis of what
AI governance platforms are actually doing covers the gap between
governance aspiration and operational reality.
The Volume Problem No Policy Is Keeping Up With
Policy processes move at the speed of legislative chambers.
Deepfake technology moves at the speed of model releases. The gap between the
two is not narrowing, and the actual scale of the problem is essential
context for evaluating whether current governance frameworks are
proportionate to the challenge.
According to data from cybersecurity firm DeepStrike, the number
of deepfake videos accessible online grew from approximately 500,000 in 2023
to around 8 million by 2025, representing an annual growth rate approaching
900 percent. Voice cloning has crossed what researchers have called an
indistinguishable threshold: a few seconds of audio are now sufficient to
generate a synthetic voice complete with natural intonation, breathing
patterns and emotional inflection. Major retailers have reported receiving
more than 1,000 AI-generated scam calls per day using cloned executive
voices.
The entertainment industry’s concern about performer likeness
exists alongside but is distinct from the broader fraud and misinformation
threat. Both are serious. Both are worsening faster than most governance
frameworks are designed to address. What the Cranston moment established is
that targeted professional advocacy from an organised constituency can move
platform consent architecture on a faster timeline than legislative processes
typically allow. Other professional communities facing equivalent risks are
aware of this and are watching closely.
What the Precedent Means for Other Professions
The Cranston moment was specific in its context but general in its
mechanism. A professional union identified a concrete governance gap in a
platform used at scale by its members’ potential adversaries. It made a
specific, technically coherent demand about consent architecture rather than
a general objection to AI. The platform’s commercial and reputational
incentives aligned with compliance. And the change was made.
That mechanism is available to professional communities well
beyond Hollywood. Journalists whose likenesses, reporting styles and on-air
personas can be replicated. Politicians whose voices and faces are among the
most targeted for political deepfake campaigns. Educators whose recorded
lectures are being used without consent to train commercial AI products.
Executives whose voices are being used in fraudulent financial scam calls.
Each of these groups has the professional organisation structure, the
affected membership constituency and the specific governance demand necessary
to run the same play.
For context on how AI is intersecting with the legal and law
enforcement dimensions of deepfake harms, our analysis of whether
AI is transforming law enforcement and why the evidence is
troubling covers the investigative and evidential challenges that
deepfake proliferation creates for authorities. And for a wider view on who
controls AI capabilities and on what terms, our piece on the
open-source AI dilemma between freedom and governance addresses the
commercial and political dynamics that shape platform decisions like the one
Cranston influenced.
The deepfake governance story is not resolved, and the Cranston
precedent does not resolve it. What it demonstrates is that individuals and
professional communities are not limited to waiting for legislation. Platform
architecture can move faster than law when the right pressure is applied to
the right lever by an organised constituency with a credible case. That
knowledge is now part of the toolkit.
About the Author
By Stuart Kerr, Technology Correspondent, LiveAIWire. Stuart
covers artificial intelligence