Some actors work for years to achieve any sort of success, but one putative thespian who has been doing the rounds recently has achieved a degree of international notoriety with just the line “three seasons and a podcast.” Or, at least, she has if you listen to some of the cheaper kind of YouTube video which uses an AI generated voice over to describe some AI generated video clips of an ersatz person.
Step back a little, though, look at what’s actually been published, and it becomes clear that the word “actor” is being confused with the phrase “a bunch of video clips.”
Okay, at some level, any film or television production can be described as a bunch of video clips. Still, the reason you might recognise the name Tilly Norwood is that her creator has managed to deftly position a very few clips of her – almost none with speech – as an existential threat to the existence of actors. What they actually are is evidence of another micro-advancement in video- and voice-generation technology.
What it’s not is an actor. Other published media shows the ersatz person in various situations: performing in ersatz sci-fi, period and dramatic movies, appearing on ersatz red carpets, and giving ersatz interviews. As so often with AI, what’s presented is a series of individually persuasive but unconnected clips. She talks more in the AI Video School piece, which is essentially a parody, but tells us a lot about Tilly’s range and how directable she is (to wit: not much, and not very). The amount of work it takes to have an AI generate the same person twice is already quite large, let alone showing that person in the same scene from two different angles.
Influences
So, Tilly’s appearances are so far limited mostly to montage. Constancy of environment is already a solvable problem – in fact, material was shown at this year’s NAB which showed an edit of several shots in which several very similar cars were driven down several very similar streets, certainly to the point where the casual observer might be persuaded to interpret them as an actual sequence. The AI prompts required to do this are extremely verbose and extremely verbose prompts require AIs with huge and expensive resources, but recent history suggests that things are likely to get easier.
The question, then, is not whether non-people are likely to become a plausible alternative to actual-people, at least in the minds of marketing executives (Tilly’s most famous appearance is a joke at the expense of the film and television commissioning process). The main issue seems to be based on the idea that the pretty brunette at the heart of this controversy is, in effect, a complex extrapolation of a number of preexisting people, in the same way that most artificial intelligences are trained on preexisting works made by actual intelligences.
This is not wrong. It’s a general principle of AI, to the point where the output of AI image generators has been called a collage of its training data. That’s probably too reductive a description of a new computing technique which has such unprecedented capabilities, but it’s also a particularly dangerous tendency to write off a process that is essentially very similar to human inspiration.
It’s hard to avoid the reality that training an AI to write, paint, speak, sing or act is analogous to the process of training a human to do things. Artists of all types are influenced by one another. Genre movies owe their existence to this kind of influence. Human beings talk freely about their influences.
Perhaps surprisingly, accusations that AI is producing work that’s unfairly close to prior art seem pretty rare. The criticism seems to be that anything produced by an AI is inherently unoriginal, which, again, seems a pretty dangerous flag to fly in a world where humans work in much the same way. Still, the fact that AI developers are careful about where they get their training data suggests that there’s some awareness of the concern.
The speed issue
If there’s a problem, it’s that AIs can be trained very quickly, and they can then create finished work very quickly. A human might take years to learn to act, sing or draw. An AI might take days, and it might be able to duplicate a previous artist well enough to put that artist out of business, especially since the AI can create an image or a document in seconds.
If there is an argument for changes in the rules, that is the justification – not that AI is doing anything new. It isn’t. New actors emerge all the time. The difference is in the value judgment that most people seem to agree with – that human beings stand to be vastly out-competed by AI in all of the world’s most interesting jobs, and that’s probably not okay.
That seems worth opposing, if only quite carefully, so we don’t accidentally end up opposing the development of medical AIs which do so well at identifying cancers. Whether the horse has long since bolted, however, is a much more difficult question. Perhaps worse, the idea of people being made to train their cheaper replacements is often scowled-at in regular employment. It is hard to argue that AI training should not be looked on with at least as much suspicion. But really, for the moment, as AI Video School showed so well, Tilly’s acting abilities are – to put it mildly – rather situationally-specific.