Jessica Hullman

New paper: People use LLMs as decision assistants, e.g. to estimate risk & recommend treatment.
To align w/our goals, we need to know:
When a model makes a bad call, why? Did it fail to arrive at correct beliefs, or not optimize its action under the utility? 🧵

It really is pointless and unnecessary to allow external reviewers to use AI as they want. There are obvious better alternatives.

By now, everyone should already be using AI to check their work before submitting. So what's the point of reviewers meat proxying to send more AI-generated feedback? Seemingly none.

What happens if we ban AI use in peer review? The main risk is to the imagined role of peer review as a basic soundness filter. But this is much better addressed by the venue having a centralized approach that uses validated AI tools to check that a paper is what it says it is than by individual reviewers feeding papers to LLMs using basic prompts. Such checks might include things like whether the results can be reproduced from the data/code, whether claims can be traced back to some provided evidence, whether there are TODOs left by an LLM to replace table results with non-fake data, etc.

Some of these can be used to determine if a paper should be desk rejected. Anything that is more controversial can be surfaced for reviewers.

This will shift what reviewers pay attention to, but this can be addressed by: 1) making clear to reviewers that their role is to identify what these checks miss and to weigh them, 2) being intelligent about how you select the set of LLM checks that are surfaced. See e.g., our work on learning LLM-implemented checklist that avoid adding more bias than they reduce variance. arxiv.org/pdf/2610.28666

What happened with ICLR being permissive with reviewers using AI freely is seeming pretty inexcusable.

Maybe we will finally learn from this ICLR sh*t-show that tolerating so much slop is not working. But I’m not very hopeful.

The irony is how fear of hurting authors’ feelings by desk rejecting their obv AI-generated paper ensures that all submitters have a terrible experience.

I wish everyone could experience firsthand what we've been seeing in AI/ML, at the scale we've been seeing it, before they weigh in on how obviously better science is going to be now.

Presenting this paper tomorrow (Monday) at @INFORMS2026 in SF - 11am session on Bridging Uncertainty Quantification and Decision Making.
Come say hi if you're around!