Ven's Declassified Posting Survival Guide
Malt Liquidity 137
I’ll preface this post by highlighting that I am not much of a fan of sci-fi nor sci-fi fanbases. I think the “predictive” reputation of famous sci-fi media is largely a function of survivor bias, and there’s a fairly unsettling trend of tech power brokers ignoring that a lot of these stories are cautionary tales of enjoyment and thus end up unironically thinking “why not build this?” Inherently, law is not rigorously logical (like math) and is in fact designed to back-propagate (e.g., cost of doing business regulation, retrofitting current progress to precedent and selling it — not an easy job by any means for an appeals judge!), and after all nobody ever innovated by following the rules. But this doesn’t mean all rationalist and skeptical considerations of should we do this should be disregarded (as much as I believe AGI/p(doom) discussion is mostly a pseudo-intellectual dorm room bong rotation discussion at the moment until it actually does invalidate an industry.)
With the advent of mass training data and AI being all the rage nowadays, it is almost certain that novel shifts in the way we think about copyright, fair use, and intellectual property are going to come in the next 3-5 years. Consequently, a single three-letter tweet could completely shift the trajectory of how things will play out:
For context, this was related to a new Chat-GPT voice system, where, of course, Altman wanted Scarlett Johansson to provide the baseline vocals.
Tangentially, it is slightly terrifying how much this sect of the industry enjoys “Her”, a movie about a dude with an AI girlfriend. It’s not particularly, er, glamorizing, but I won’t spoil it. Legally, this is shockingly not a novel question at all. The general tech world seems to not get the issue at all, which is, hilariously, essentially Contracts 101. Any first year law student would instantly recognize this as a very similar scenario to Midler v. Ford, a case that came, notably, from California’s circuit, where a similar OpenAI lawsuit could be filed.
Ford Motor created an ad campaign for the Mercury Sable that specifically was meant to inspire nostalgic sentiments through the use of famous songs from the 1970s sung by their original artists. When the original artists refused to accept, impersonators were used to sing the original songs for the commercials. Midler was asked to sing a famous song of hers for the commercial and refused. Subsequently, the company hired a voice-impersonator of Midler and carried on with using the song for the commercial, since it had been approved by the copyright-holder. Midler's image and likeness were not used in the commercial but many claimed the voice used sounded impeccably like Midler's…
The appellate court ruled that the voice of someone famous as a singer is distinctive to their person and image and therefore, as a part of their identity, it is unlawful to imitate their voice without express consent and approval. The appellate court reversed the district court's decision and ruled in favor of Midler, indicating her voice was protected against unauthorized use.
You can take a listen for yourself:
It is odd (and disconcerting) that a hypothetical first-year law student would probably have handled this situation better than OpenAI’s corporate counsel: namely, all you have to say is “hey, dumbass, don’t post something that clearly links our product’s similarity to the person we approached to be the basis of the product and failed to secure a contract with.” (My default style of phrasing is why I can’t work in big corporate.) The only job of corporate counsel is to limit litigation and liability as much as possible by maintaining plausible deniability, due to how the legal system works:
Intuitively, we can derive the primary purpose of modern law from this hypothetical: it’s about property rights first and foremost. Law primarily exists to settle property disputes so that people with resources do not mobilize them to war over said property. Criminal law, on the other hand, exists as a sort of societal insurance — for egregious misconduct, alright, one loses their privilege to participate in society (as long as these laws are enforced).
Thus, the standard of proof being lower for civil cases makes sense: time is of the essence and, ideally, minimal monetary and legal resources are expended to arrive at a settlement. You don’t want to settle disputes in a hamfisted, drawn-out, “beyond a reasonable doubt” manner, as it would kill liquidity (and, as Coase pointed out nearly a century ago, that’s precisely why matters end up settling between companies rather than through lawsuits.)
So if the paper trail comes out supporting Scarlett’s timeline, given Altman’s direct reference timed with that product’s release, it’s a sticky legal situation to talk yourself out of. This is essentially the reason why the compliance industry went through the roof post-Dodd Frank: good luck doing anything in finance without a ton of dense corporate legal lectures and lawyers clearing everything you do, whether it’s placing trades or sending iMessages. The tech industry is certainly more laissez-faire about compliance: “move fast, break things” and “funding secured” are more than enough evidence that the mentality of doing first and asking for permission later is highly valued. And Altman’s brand is notably talking and posting a lot. But I get the sense that he feels like he’s untouchable, like Elon, when OpenAI is very much involved in difficult, complex litigation with the New York Times and Sarah Silverman that will open the door to courts (such as the 9th circuit that ruled on Midler) writing new opinions and methods of interpretation, which might be further expounded on through statutes. This problem did not exist in the 1980s, of course.
The game here is one of reputation, not trial and error. To bend the new rules in your favor, you need to be a trustworthy individual (or, at the very least, not an outright bad actor.) If you are actively antagonizing other powerful interests (and, certainly, Hollywood is one of the most powerful entities in the US), when push comes to shove and it’s time to write the new rules, you’re going to get screwed over. (Regulatory capture is also a resultant effect from Coase.) It’s very hard to see how you could word rules under current precedent specifically regarding training data and machine learning methods without invalidating the legality of using regular, mathematically efficient algorithms. Then again, there is obviously something that feels different about, say, scraping the New Yorker or using an author’s books to train generative text to look similar (albeit distinguishable to the discerning reader — you won’t trick me on GPT-slop vs human writing) versus just doing normal pandas operations on datasets. So while this entire incident is not monetarily significant, any discovery of how this model was trained — especially if they did use some of Scarlett’s work in the process — could have serious sentiment repercussions as to how this gets handled down the line. In my view, this is a much bigger crisis than the “AI Safety team” debacle from a philosophical point of view.
So, what follows are my (totally not legal advice) rules for posting while under contract:
Don’t do it
If you do, don’t mention anything remotely specific about what you do for work
If you do talk about your work, only talk in general concepts and nothing related to proprietary work
If you do want to talk about your proprietary work, do not create an electronic, time-stamped trail that aligns with public releases of that work.
I’m sure I’ll have more rules to add, but we’ll start with those ones for now. On a side note, is anyone else amused at how much Altman looks like Burke from Aliens?