By Gordon Hull
As I argued last time, authorship is a political function, and we should be applying that construction of it to understand whether AI should be considered an author. Here is a first reason for doing so: AI can’t really be “accountable.”
(a) Research accountability: The various journal editors all emphasize accountability. This seems fundamentally correct to me. First, it is unclear what it would mean to hold AI accountable. Suppose the AI fabricates some evidence, or cites a non-existent study, or otherwise commits something that, were a human to do it, would count as egregious research misconduct. For the human, we have some remedies that ought, at least in principle, to discourage such behavior. A person’s reputation can be ruined, their position at a lab or employer terminated, and so on. None of those incentives would make the slightest difference to the AI. The only remedy that seems obviously available is retracting the study. But there’s at least two reasons that’s not enough. First, as is frequently mentioned, retracted studies still get cited. A lot. Retraction Watch even keeps a list of the top-10 most cited papers that have been retracted. The top one right now is a NEJM paper published in 2013 and retracted in 2018; it had 1905 cites before retraction and 950 after. The second place paper is a little older, published in 1998 and retracted in 2010, and has been cited more times since its retraction than before. In other words, papers that are bad enough to be actually retracted cause ongoing harm; a retraction is not a sufficient remedy for research misconduct. If nothing else, whatever AI is going to find and cite it. And all of this is assuming something we know to be false, which is that all papers with false data (etc) get retracted. Second, it’s not clear how retraction disincentivizes an AI any more than any other penalty. In the meantime, there is at least one good argument in favor making humans accountable for the output of an AI: it incentivizes them to check its work.
In short, in the absence of analogues to accountability for humans, evidence that making an AI accountable achieves anything, or evidence that limiting accountability to humans doesn’t incentivize error-checking, it seems to me that the burden of proof ought to be on those who don’t think accountability is a good reason to deny authorship to AI.
A basic account of juridical personhood would seem to buttress this line of reasoning. I am not a scholar of the personhood literature, so this is going to be basic, but Locke, for example, in Essay II.27 identifies personhood with continuity of consciousness. If you didn’t have some sort of framing construct along those lines, accountability would be difficult. This is why Locke – who, remember, is arguing that personhood is a juridical concept and that trying to establish it metaphysically leads to total confusion – lands on difficult cases like amnesia. If someone chooses to get drunk and then does something stupid, Locke is in favor of assigning responsibilty for their actions, since they chose to get drunk. But if they have some sort of complete, non-recoverable break in their memory that isn’t their fault, he’s inclined to think that it makes sense to call the person after the break a different person from the one before it.
Against this, Ryan Jenkins and Patrick Lin (in the paper I mentioned last time) note that “authors are sometimes posthumously credited, even though they cannot presently be held accountable for what they said when alive, nor can they approve of a posthumous submission of a manuscript; yet it would clearly be hasty to forbid the submission or publication of posthumous works.” While obviously true, this strikes me as misplaced in its application to AI authorship. First, in the case of humans, posthumous publication is a clear exception to a general rule about how accountability works. In the case of AI, the difficulty in accountability is the rule.
Second, all authors are going to die, which means that there’s a limit built in to accountabillity in any case. But this applies to AI authors too: pretty much anything to do with computing is going to go offline at some point. Indeed, this is likely a more serious problem with AI, since the life cycle of AIs is very, very short. GPT-3, which was released last year, is going to be replaced… this year. Now, one could argue that all the versions of OpenAI products are the same “author,” since they’re the product of the same research team and will have trained on an accumulating corpus of text. There’s probably a Ship of Theseus problem buried in there somewhere, as research teams and algorithms change, as does what the model scrapes off the Internet. Let’s concede that, since all of us at least have constantly varying inputs, and those inputs can change the cognitive content in our heads, sometimes quite radically. The bigger issue is that there’s lots of different LLMs out there, and we have no reason to think that any one of them is going to be around for that long. Even the companies that produce the LLMs (who have lots of reasons why they asren’t responsible for the outputs in any case) aren’t always around all that long. And if they are, they change ownership: Microsoft just dropped a $10 billion investment into OpenAI. Is OpenAI the same now as before, from the point of view of accountability? It seems to me there’s at least a prima facie case that it isn’t, and that there’s probably some expensive lawyers at Microsoft who would say the same thing.
If we want accountability, in other words, we want to attach it to something at least a little bit stable. Human authors do that better than LLMs.
(b) Corporate Accountability? All of that said, I do think a different way of conceptualizing accountability might be helpful: if assigning authorship to AI somehow incentivizes accountabiliy on the part of those who create LLMs – the corporate entities behind them – then that might well be worth pursuing. I’m not sure how that would happen, at least in a legal sense (and the legal sense is the only one they’ll respond to), because there are so many plausible gaps between the engineering design and the output of the system, such as the training data and the fact that the system generates its own internal nodes. But if there were a way of somehow forcing the creators of LLMs to internalize some of the costs and harms they generate, that would be at least a prima facie reason to support that endeavor. Authorship would be one such strategy. Doing this would require at least two things: first, the incentive structure would have to somehow transfer liability. The model that comes to mind is the limit on consumer loss to $50 from fraudulent credit card transactions. The policy has both enabled consumers to feel safe using credit cards, and caused the credit card industry to try very hard to prevent fraud, and to absorb the damage when it does. There’s obviously a ton of disanalogies with LLMs but something like that could be a model.
The other thing that would have to be developed is a standard for due diligence and when it would be right to tag the creator of the LLM with the harms it caused. In the case of research authorship, we might argue that failing to give the LLM hard rules against fabricating a bibliography (as in Faust’s example, discussed last time), constitutes some form of legally accountable failure. The limit of this approach is that the complexity of using LLMs “in the wild” is going to generate a continuous stream of examples of questionable behavior (for many of them, people are going to argue about whether the behavior is really questionable). It will be relatively easy to develop post hoc rules to ban certain offensive behaviors, but it’s going to be very hard to articulate workable standards for when a model’s creators are working hard enough to prevent those harms from happening in the first place. The current model of privacy regulation is really discouraging – the soft compliance and best practices model has been completely corrupted into meaninglessness by corporations who’d rather not change their data collection practices, as Ari Waldman has conclusively demonstrated.
In short, I think accountability for research is not served by assigning LLMs author roles. Accountability for LLM creators might be, but there’s a lot of “devil in the details” work in the way. I’ll look at a completely different kind of question – social justice – next time
Recent Comments