On the Ethics of Hollywood Deepfakes

Why has there been such a drop lately in TV and movie deepfake content? The author worked for movie deepfake pioneer Metaphysic.ai for over two and a half years, including 20 months as a full-time employee.

Opinion/Feature As a kid, I was captivated by stories of the foundation of Industrial Light and Magic (ILM), operating out of a basic warehouse in Van Nuys, and painstakingly building the new generation of motion-controlled visual effects methods that would go on to dominate most of the 1980s, up until the advent of CGI.

Blowing up bits of the Death Star in the courtyard of ILM at Van Nuys in the mid-1970s. Source

It’s an outstanding origin-story, and I recommend the many books and articles that retell how the effects for Star Wars (1977) were – after a great deal of toil, recrimination, and delay – eventually delivered.

The fledgling ILM would later move to San Francisco and go on to industrialize photochemical effects production. Even by the time of VFX work on The Empire Strikes Back in 1978/9, ILM, while still a pioneering and innovative effects house, was already replete with accounting and HR departments.

But in that period, and since, my imagination was captured by wondering what those first few ‘raw’ years at Van Nuys were like, cannibalizing model kits and 1960s optical methods alike, and practically inventing an industry that had once been the preserve of individual auteur VFX specialists.

Face of the Revolution

In the nearly three years that I worked at the now-dissolved Metaphysic.ai visual effects company, I think I came as close as I possibly could have come to the mid-70s Van Nuys experience.

In that time, Metaphysic produced AI-driven, ID-specific effects shots for some of the biggest franchises, including the Mad Max and Alien franchises, as well as producing the pioneering (if much-criticized) de-aging deepfake work for Robert Zemeckis’ Here (2024), and many other works:

There were more ‘dailies’ meetings (the showing of the latest footage produced, in crowded Zoom meetings) than any one person could reasonably attend and still get their own work done, as well as skunkworks-style huddles, where select groups would work together on mini-moonshots, seeking to advance the company’s state-of-the-art.

Many of the staff came to the company either directly from ILM, or else had worked for ILM in the past*; and even though it was clear, from anecdotes, that the Van Nuys spirit had long-since given way to mass-rationalization at ILM, and the various other major VFX houses that some of the staff had come from, the association remained electric for me.

George Lucas reviews effects shots with John Dykstra, Richard Edlund, Joe Johnston, and Rose Duignan during the production of Star Wars. Source

I don’t know if one could argue that Metaphysic would invent an AI equivalent to the DykstraFlex motion-control rig that came out of the Van Nuys sweatshop; but it put AI-driven VFX on the map, and was responsible for more professional-level AI face-replacement and altering than any other effects house or single project undertook in the period, or has undertaken since.

Slowdown

And I have to note that since the company was subsumed into DNEG’s Brahma AI sub-company, that steady stream of AI-based visual effects work seems to have slowed to a crawl.

Despite headlines such as Ben Affleck’s AI VFX Netflix partnership, and a commitment to AI from directing luminaries Soderbergh and Aronofsky, I haven’t seen a whole load of deepfake faces in movies and TV lately (excluding The Capture‘s simulation of deepfake tech), compared to ‘peak Metaphysic’.

Given the extent of the AI backlash across so many fields and sectors, I wonder if we aren’t all beginning to ‘read the room’ on AI-altered identities; and if AI-facilitated efforts such as the resurrection of Ian Holm, the youthification of Harrison Ford, and the revival of the Rat Pack to sell gin might soon elicit the now-classic meme ‘It was a different time’.

Negative Associations

In many ways, it’s a problem of association; but it’s not the same association problem that it was back in 2021, when Metaphysic formed.

At that time, deepfake porn had become a viral phenomenon, with all SFW and NSFW output alike coming from one of two open source software repos^† . In contrast to today, there was no ‘gatekept’ or censored way of producing photoreal AI content.

Then, the biggest deepfake porn site in the world (destined to be taken down by an investigation in 2025) was where the official instructions were for the biggest deepfake software repo, DeepFaceLab, itself the admitted foundation of Metaphysic’s original approaches:

The original user-guide for DeepFaceLab, at one time the world’s most popular deepfake software – hosted at the world’s biggest deepfake porn site. Since then, the software has been archived, and the site taken down. Source

Since software developers were clearly not shying away from the association between porn and deepfakes, this would inevitably affect any company trying to legitimately adopt this tech.

Consequently (perhaps not least because the company’s founder was included with the founder of DeepFaceLab in the authorship of an Arxiv position paper on deepfakes in 2020, and was subject to much criticism over these associations) Metaphysic had a dedicated ethics department, and the company was vocal and very public on this topic.

These days, with the extensive and growing tranche of global legislation criminalizing deepfake content, that ‘unfortunate’ association regarding AI-created identities has transferred from porn to poverty, as global work-forces directly threatened by different kinds of AI (usually Large Language Models, or LLMs) are beginning to demonize the term across the board.

The Villagers are Revolting

At a certain point, perhaps the middle of 2023, attending various company meetings, I began to feel, admittedly at second-hand, a growing wave of casual ad hoc resistance to the services the company was offering, coinciding with the earliest social upswells around AI as an existential threat to workers (though it has to be noted that Hollywood unions were in the vanguard in this regard).

AI appeared to be viewed in some quarters as ‘not worth the risk’. Additionally, in Metaphysic’s peak period of output, it never really had an undisputed hit after the shock value of putting (a rather shiftless) young Elvis Presley in America’s Got Talent. In fact, neither did any company that was outputting AI-generated content in mainstream movie and TV work.

There always seemed to be some reason to hate: the ever-present uncanny valley; the association with porn; and then, the association with a threatened AI-driven economic apocalypse.

AI’s ‘Tyrannosaur Moment’

It had seemed to me, for a long time, that recreating humans with 100% fidelity in a VFX context that had been failed by CGI for a long time, was a worthwhile, if not an inevitable goal.

I imagined that the evolution of AI in visual effects would culminate in a third ‘epochal change’ moment for the industry, to the delight of viewers. The first had been the first sight of the star-destroyer thundering overhead in Star Wars in 1977, and the second, the raging Tyrannosaur escaping its paddock in Jurassic Park, in 1993.

Each had led to radical shifts in the industry, and each had been presaged by early adopters pushing the limits of a nascent technology or method.

Thus it seemed to me that Metaphysic’s stream of releases, while not necessarily supplying audiences with that ‘tyrannosaur moment’, would inevitably lead to it; that one day, we would see something impossible in a movie; something fundamentally made by AI (rather than papered-over with CGI), that would knock our socks off.

Now…I don’t know. It isn’t necessarily that I don’t believe that moment of amazement will come, for AI in cinema; nor solely that that there are many critical issues for AI video to overcome before it can join the professionals at the table.

It’s more that I have begun to share, in a vague (and perhaps morally cowardly) way, that nascent and growing public feeling of negative association between AI in VFX, and the negative ramifications of AI in other sectors, in this period.

Besides this, I’m not currently convinced that the historic influence of unpaid and non-free datasets such as Common Crawl has really been accounted for yet, and I consider that many of the much-touted ‘AI licensing’ deals that have made headlines over the last 18 months or so represent some extent of IP-washing.

Therefore, the moment that a perfect and non-uncanny AI-regenerated Marilyn Monroe or Humphrey Bogart steps onto the screen, I think that twin feelings of awe and awkwardness are likely to arise in me, because AI wizardry’s baggage has begun to weigh me down.

Conclusion

The prior two technological breakthroughs came with no such caveats: the advent of ILM and motion control rigs effectively invented an industry that had hitherto been a series of sporadic commissions met by a sparse elite of ad hoc freelancers; and the advent of CGI massively increased the number of jobs in the VFX industry, and created value in adjacent software development, as well as various supporting services.

My interest in Computer Vision remains intact, along with my interest in the potential of AI in a wide number of possible audiovisual applications; and my enthusiasm for wider applications of machine learning, across a range of other sectors, is undimmed.

I only wish I could continue to look forward to that AI ‘wow’ moment at the movies.

As for the reason why there is a sudden paucity of AI-generated faces in movie and TV, compared to the most active years of Metaphysic – one possibility is that AI-aided tweaks are now applied by production houses with little or no publicity (compared to the blaring press releases of the 2022-23 period), given that public response to AI innovation in entertainment has become so hostile lately.

Secondly, to restate: after AGT, where the company made its debut, Metaphysic never really had an unalloyed hit in regards to its own contribution to any movie or TV show; and neither the Indiana Jones 4 deepfake nor any of the other work from other production houses was well-received enough, or lucrative enough to start a rush of copycats, or to give AI in VFX genuine credibility and investor-friendliness.

* This is easily verifiable by merely looking at publicly available LinkedIn profiles and IMDB credits.

^† DeepFaceLab, and the less porn-associated FaceSwap.

First published Monday, April 27, 2026