Hanna Wallach

This is happening now!!!

Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge The measurement tasks involved in evaluating generative AI (GenAI) systems lack sufficient scientific rigor, leading to what has been described as "a tangle of sloppy tests [and] apples-to-oranges com...

1) (Tomorrow!) Wed 7/16, 11am-1:30 pm PT poster for "Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge" (E. Exhibition Hall A-B, E-503)

Work led by @hannawallach.bsky.social + @azjacobs.bsky.social

arxiv.org/abs/2502.00561

Oh whoops! You are indeed correct -- it starts at 11am PT!

If you're at @icmlconf.bsky.social this week, come check out our poster on "Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge" presented by the amazing @afedercooper.bsky.social from 11:30am--1:30pm PDT on Weds!!! icml.cc/virtual/2025...

I also want to note that this paper has been in progress for many, many years, so we're super excited it's finally being published. It's also one of the most genuinely interdisciplinary projects I've ever worked on, which has made it particularly challenging and rewarding!!! ❤️

Check out the camera-ready version of our ACL Findings paper ("Taxonomizing Representational Harms using Speech Act Theory") to learn more!!! arxiv.org/pdf/2504.00928

Why does this matter? You can't mitigate what you can't measure, and our framework and taxonomy help researchers and practitioners design better ways to measure and mitigate representational harms caused by generative language systems.

Using this theoretical grounding, we provide new definitions for stereotyping, demeaning, and erasure, and break them down into a detailed taxonomy of system behaviors. By doing this, we unify many of the different ways representational harms have been previously defined.

We bring some much-needed clarity by turning to speech act theory—a theory of meaning from linguistics that allows us to distinguish between a system output’s purpose and its real-world impacts.

These are often called “representational harms,” and while they’re easy for people to recognize when they see them, definitions of these harms are commonly under-specified, leading to conceptual confusion. This makes them hard to measure and even harder to mitigate.

Generative language systems are everywhere, and many of them stereotype, demean, or erase particular social groups.

Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge The measurement tasks involved in evaluating generative AI (GenAI) systems lack sufficient scientific rigor, leading to what has been described as "a tangle of sloppy tests [and] apples-to-oranges com...

Check out the camera-ready version of our ICML position paper ("Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge") to learn more!!! arxiv.org/abs/2502.00561

(6/6)

Real talk: GenAI systems aren't toys. Bad evaluations don't just waste people's time---they can cause real-world harms. It's time to level up, ditch the apples-to-oranges comparisons, and start doing measurement like we mean it.

(5/6)

We propose a framework that cuts through the chaos: first, get crystal clear on what you're measuring and why (no more vague hand-waving); then, figure out how to measure it; and, throughout the process, interrogate validity like your reputation depends on it---because, honestly, it should.

(4/6)

Here's our hot take: evaluating GenAI systems isn't just some techie puzzle---it's a social science measurement challenge.

(3/6)

But there's a dirty little secret: the ways we evaluate GenAI systems are often sloppy, vague, and quite frankly... not up to the task.

(2/6)

Alright, people, let's be honest: GenAI systems are everywhere, and figuring out whether they're any good is a total mess. Should we use them? Where? How? Do they need a total overhaul?

(1/6)

I'm so excited this paper is finally online!!! 🎉 We had so much fun working on this with @emmharv.bsky.social!!! Thread below summarizing our contributions...

Please spread the word to anyone who you think might be interested! We will begin reviewing applications on June 2.

This program is open to candidates who will have completed their bachelor's degree (or equiv.) by Summer 2025 (inc. those who graduated previously and have been working or doing a master's degree) and who want to advance their research skills before applying to PhD programs.

Exciting news: The Fairness, Accountability, Transparency and Ethics (FATE) group at Microsoft Research NYC is hiring a predoctoral fellow!!! 🎉

www.microsoft.com/en-us/resear...

Exciting news!!! This just got into @icmlconf.bsky.social as a position paper!!! 🎉 More updates to come as we work on the camera-ready version!!!

Thank you for posting! Very timely as the paper just got accepted to ICML's position paper track!

Reading - Evaluating Evaluations for GenAI from
@hannawallach.bsky.social madesai.bsky.social afedercooper.bsky.social et al-This work dovetails with our work at
@worldprivacyforum.bsky.social on measuring AI governance tools from governments, through privacy/ policy lens arxiv.org/pdf/2502.00561

At the #HEAL workshop, I'll present "Systematizing During Measurement Enables Broader Stakeholder Participation" on the ways we can further structure LLM evaluations and open them for deliberation. A project led by @hannawallach.bsky.social

2. Also Saturday, @amabalayn.bsky.social will represent our piece arguing that systematization during measurement enables broad stakeholder participation in AI evaluation.

This came out of a huge group collaboration led by @hannawallach.bsky.social: bsky.app/profile/hann...

heal-workshop.github.io

📣 New paper! The field of AI research is increasingly realising that benchmarks are very limited in what they can tell us about AI system performance and safety. We argue and lay out a roadmap toward a *science of AI evaluation*: arxiv.org/abs/2503.05336 🧵

Screenshot of 'SHADES: Towards a Multilingual Assessment of Stereotypes in Large Language Models.' SHADES is in multiple grey colors (shades).

⚫⚪ It's coming...SHADES. ⚪⚫
The first ever resource of multilingual, multicultural, and multigeographical stereotypes, built to support nuanced LLM evaluation and bias mitigation. We have been working on this around the world for almost **4 years** and I am thrilled to share it with you all soon.

Remember this @neuripsconf.bsky.social workshop paper? We spent the past month writing a newer, better, longer version!!! You can find it online here: arxiv.org/abs/2502.00561

Thank you!!!! 😍