Yonathan Arbel

Let's build safe AI!
Law prof @ Alabama
Contracts, Defamation, Legal NLP, & AI Safety

[Yonathan Arbel @ProfArbel]

Enjoyed speaking with the wonderful Rosemary Church on CNN last night about AI governance and what it really means to win the AI race.

Building safe, reliable AI is part of what winning requires.

[Yonathan Arbel @ProfArbel]

@aryehazan Although they didn't. This is an embarrassing report (going by what was summarized in the article I didn't read the full report)

[Yonathan Arbel @ProfArbel]

@Jabaluck It may be that the best way to learn about world affairs doesn't involve John Oliver doing smug impersonations

[Yonathan Arbel @ProfArbel]

This is a simulated placeholder for a banger tweet and should be replaced by actual banger tweet.

[Gautam Kamath @thegautamkamath]

How do you tell if authors fabricated their research results? Sometimes: they just tell you. Read, e.g., this caption carefully.

Another banger of a post by my friend Nihar B. Shah on his blog, where he found at least 20+ such #ICLR2027 submissions:
https://t.co/TBKmnMjTiy

[Yonathan Arbel @ProfArbel]

@aredachem Contestability of AI outputs (and its limits!) is an important part of the future legal landscape. But human overrides aren't always better.

[Yonathan Arbel @ProfArbel]

Back from CELS presenting Generative Gap Filling to a crowded room.*

It was interesting to see how AI slowly comes to the legal world. Obviously the slides were majority Claude, and you could see signs that some of the analyses relied on agents. But I still think lawyers are two years behind the frontier. Still coming to terms with how to use AI and whether AI is reliable. Nobody is surprised anymore when you say that LLMs are doing a good job and they stop asking low information questions about stochastic parrots. But it's also clear that there is a field of low hanging fruit that new scholars will pick aggressively to great results

* It was a smaller room and the other speakers were doing cool work, but forget that.

* If you have read GGF then you are even behinder on the frontier. Go read it!

[Yonathan Arbel @ProfArbel]

https://t.co/E5RZEQBnu5

On judicial economy and the case for integration
https://t.co/38iM9cShSX

[Yonathan Arbel @ProfArbel]

Yesterday, @thehoffprof asked how we can manage law review article selection in times of the slopalypse

Today, a new study offers some evidence for fighting AI flooding with AI. In head-to-head comparisons, AI-generated reviews beat reviews by paid human experts. AI judges overwhelmingly preferred them. Three former journal editors also preferred them in 26 of 30 judgments.

That doesn't mean that AI should select our future scholarship. The downsides are obvious. The study doesn't really test selection decisions, much less questions of scholarly impact. But it makes the case for using AI-assisted triage to direct scarce human attention toward interesting, rigorous, and creative articles

Complain about it, by all means. But the flood is coming either way. If you don't have a better AND realistic solution, the complaint won't get us very far.

[Yonathan Arbel @ProfArbel]

Legal personhood to AI is coming. The fights are misplaced, because it's not a metaphysical question we get to debate in philosophy journals, but a question with a strict deadline.

The real, pressing question is how do we design AI personhood in a way that ensures our prosperity and welfare.

And please try to think it through without falling on comforting platitudes like "clankers don't have feelings, check mate"

[Peter N. Salib @petersalib]

Corporations are (legal) persons, even though we don’t think they are conscious or matter morally! It’s pure governance.

We can do the same with AIs.

https://t.co/WfbmRGODfA

https://t.co/5c5SMXGyBZ

[Yonathan Arbel @ProfArbel]

@IrvingLinks In both cases the question is why. Is there a rational basis or something more psychological.

[Yonathan Arbel @ProfArbel]

At a conference on AI & legal profession. I'm sorry folks, but you are over-indexed on hallucinations. You aren't calibrated on the magnitude of the problem and it's possible that one reason is psychological. It's comforting to think that the robots will never be as accurate & diligent as we are.

*I am NOT saying that AI doesn't hallucinate or never makes mistakes; indeed, I believe that it does.

[Yonathan Arbel @ProfArbel]

@HoffProf @apublicgood Anders, you have been sitting on a solution and didn't share it with your new colleague? Is it because I didn't finish my onboarding training modules?

[Yonathan Arbel @ProfArbel]

A rare opportunity to disagree with The Hoff.

So first, he is right on all the main points, including flooding and how ugly AI slop is, &c.

But I don't love the remedy (and I suspect he isn't too excited about it either?). Pangram is measuring the wrong thing and is also gamable (less so than I thought in the past!). Still the semi lazy can just have Claude generate an article and then rewrite it using meat tokens, and the AI signature disappears without resolving any of the underlying problems. The truly lazy will use AI humanizers, but I haven't benchmarked those.

What to do? Use AI to identify works worthy of human attention! AI triage is the only robust way forward IMHO, and also, I suspect that it's being done covertly anyways.

A lot more to say but I wanted to outline this quickly

[Dave Hoffman @HoffProf]

I've been thinking about how law reviews (and academics) should respond to AI writing, and this 👇 is my basic starting point. Pangram is a badly designed purity stick.

[Yonathan Arbel @ProfArbel]

@Roq44 @arpitrage No, but you are able to read more into my tweet than my mom when I tell her that I'm fine

[Yonathan Arbel @ProfArbel]

@ThatOtherTom1 @be_subtle :) but for real, try using the advanced models.

[Yonathan Arbel @ProfArbel]

@be_subtle Why isn't my nerf gun taking down a mammoth?

[Yonathan Arbel @ProfArbel]

@chrishampsonlaw It's a @petersalib neologism for the historical way of producing text, i.e., via human writers.

[Yonathan Arbel @ProfArbel]

@leastfixedpoint I just wish there was a section in the Olympics for motivated bad reading. 🏅

[Yonathan Arbel @ProfArbel]

@DanSoch If anything I'm surprised it's so low!

[Yonathan Arbel @ProfArbel]

@arpitrage Yes! the product gets better and users get better, and a lot of the conversation is based on free products and naive users

[Yonathan Arbel @ProfArbel]

@MackenZ_arnold I always ask grandma to pass a captcha before I tell her how I am doing

[Yonathan Arbel @ProfArbel]

@SequoiaMaple We gave our best design ideas here

https://t.co/CxLwlFUUWy

In short the answer is no in principle. We don't want a halt right that is only available when clear conditions apply ( which is a major flaw of the ant bill)