Will AI Lawyers Make Swords or Shields?
There is good reason to think that lawyers may be one of the most automatable professions under the current trajectory of AI development. At its core, the practice of law is reasoning about the set of texts—statutes, regulations, court cases—comprising the law as they apply to particular fact patterns. Language models excel at exactly this sort of textual reasoning task. Today’s AI systems can already achieve passing scores on the LSAT,1 law school exams,2 and the bar exam.3 They continues to make factual mistakes about the precise content of the law (and thereby get lawyers in trouble),4 but I expect that today’s frontier AI systems that can perform research tasks are much less likely to hallucinate legal authorities, and tomorrow’s systems will be better still.5 As they say, this is the worst that AI Lawyers will ever be.
When I first thought of the possibility of AI Lawyers, my initial reaction was that AI lawyers will be overall good for the cause of justice. Armed with cheap AI Lawyers, defendants with strong defenses will be less likely to be coerced into unfavorable settlements because it costs them much less to take their cases to trial.
As I thought about it more, though, I realized I was failing to consider how AI Lawyers might affect prosecution dynamics. Lower legal costs for prosecution will likely mean that there will be less of a need for prosecutorial discretion. This would imply that the government might be more aggressive in pursuing more meritorious (or possibly-meritorious) cases that it currently lacks the resources to pursue. AI Prosecutors would also have less of an incentive to settle with defendants, since it would be cheaper to take cases to trial.
I don’t want to take a strong stance here on whether more meritorious prosecutions and harsher sentences are, on the current margin, good, assuming they are pursued in good faith.
But I worry that AI Prosecutors would also remove a major bottleneck to selective prosecutions, since those draw resources away from legitimate prosecutions in the current regime. If “Show me the man and I’ll show you the crime” is right—if, with enough investigative zeal, law enforcement could probably construct a meritorious criminal case against almost anyone6—then the AI Lawyers could be a lot more sinister than I initially hoped. AI Prosecutors (and AI Investigators) could make it easier for the state to single out political rivals for intense scrutiny until a meritorious case is found, then pursue that meritorious case to the fullest extent of the law. It may also be easier for AI Prosecutors to hide the fact that their prosecutions are selective, or come up with plausible legal arguments for why their prosecutions are not impermissibly selective. (And note that selective prosecution arguments are already very unlikely to work.) Defendants would still have AI Defense Lawyers, of course, but if the defendant is provably guilty, then the prosecution can still secure a conviction.
As usual, my preferred policy solution would not be to impede the development and rollout of AI lawyers in gross. Hopefully we can preserve something like the current equilibrium, under which abuses of prosecutorial power are limited by resource constraints, through a combination of tools. This should include not just cheap AI Defense Lawyers, but also a more wholesale rethinking of the criminal code, sentencing guidelines, and the prosecutorial system. But I worry that the transition period will be one in which prosecutors have many more resources than they used to, but the same underlying criminal penalties (which were themselves often calculated to be overly punitive so as to have some deterrent effect in an environment of imperfect prosecution). This seems worrisome.
See OpenAI, GPT-4 Technical Report 5 (Mar. 4, 2024) (unpublished manuscript), https://arxiv.org/pdf/2303.08774.
See Andrew Blair-Stanek et al., GPT Gets Its First Law School B+’s (2024) (unpublished manuscript), https://ssrn.com/abstract=4717411.
See Eric Martínez, Re-evaluating GPT-4’s Bar Exam Performance, A.I. & L. (2024), https://doi.org/10.1007/s10506-024-09396-9.
See, e.g., Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023).
Cf. Varun Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (May 30, 2024) (unpublished manuscript), https://arxiv.org/pdf/2405.20362 (using “retrieval-augmented generation” to improve the reliability of LLMs’ legal analysis)
I do not know whether this is true. The famous meme that the average person commits “three felonies a day” is not actually supported by the book bearing that title. See David Henderson, Three Felonies a Day?, Econlib (Jan. 5, 2019), https://www.econlib.org/three-felonies-a-day/ [https://perma.cc/3XVS-SVZM].