Friday, 25th April - 11am: Mohit Iyyer Title: Is detecting AI-generated text a fool's errand? Abstract: In this talk, I'll share how my thinking about detecting AI-generated text has shifted over the past two years. Back in 2023, I was skeptical about classifier-based detection because our experiments showed they were easy to trick with simple paraphrasing. At the time, watermarking seemed like a better bet—it was tougher to evade and gave more accurate results. But watermarking relies on cooperation from LLM providers like OpenAI, and they don’t really have any incentive to play along. To get around this, we developed something called PostMark in 2024, a watermarking method that can work with any black-box LLM without needing provider cooperation. Unfortunately, PostMark's usefulness in the real world still depended on government involvement to set up a trusted third-party authority, which didn't seem likely. By this point, my students and I had grown disillusioned with the topic, viewing it as both technically limited and practically irrelevant. But then, a few months ago, we stumbled upon something surprising: people who frequently use ChatGPT were actually great at spotting AI-generated text, outperforming every automated detector we tested. Even better, these human experts could clearly explain what tipped them off—things like overly formulaic sentences, specific "AI-ish" words (like "delve" or "crucial"), and a general lack of detail. Another unexpected twist was discovering that Pangram, a classifier, performed nearly as well as these human experts, which completely flipped my earlier views on classifiers. I'll wrap up by talking about our ongoing work trying to train LLMs that can fool Pangram, and I'll connect that to broader questions about creativity, memorization, and what makes human writing distinct from AI-generated text. (Was this abstract written by an LLM? Find out at the talk!) Bio: Mohit Iyyer is an associate professor in computer science at the University of Maryland, College Park, with a primary research interest in language generation. He is the recipient of multiple best paper awards at NLP conferences, and he also received the 2022 Samsung AI Researcher of the Year award. He obtained his PhD in computer science from UMD in 2017, spent the following year as a researcher at the Allen Institute for Artificial Intelligence, and was then faculty at UMass Amherst until rejoining UMD in January 2025. Apr 25 2025 11.00 - 12.00 Friday, 25th April - 11am: Mohit Iyyer This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. G.03 IF and Teams
Friday, 25th April - 11am: Mohit Iyyer Title: Is detecting AI-generated text a fool's errand? Abstract: In this talk, I'll share how my thinking about detecting AI-generated text has shifted over the past two years. Back in 2023, I was skeptical about classifier-based detection because our experiments showed they were easy to trick with simple paraphrasing. At the time, watermarking seemed like a better bet—it was tougher to evade and gave more accurate results. But watermarking relies on cooperation from LLM providers like OpenAI, and they don’t really have any incentive to play along. To get around this, we developed something called PostMark in 2024, a watermarking method that can work with any black-box LLM without needing provider cooperation. Unfortunately, PostMark's usefulness in the real world still depended on government involvement to set up a trusted third-party authority, which didn't seem likely. By this point, my students and I had grown disillusioned with the topic, viewing it as both technically limited and practically irrelevant. But then, a few months ago, we stumbled upon something surprising: people who frequently use ChatGPT were actually great at spotting AI-generated text, outperforming every automated detector we tested. Even better, these human experts could clearly explain what tipped them off—things like overly formulaic sentences, specific "AI-ish" words (like "delve" or "crucial"), and a general lack of detail. Another unexpected twist was discovering that Pangram, a classifier, performed nearly as well as these human experts, which completely flipped my earlier views on classifiers. I'll wrap up by talking about our ongoing work trying to train LLMs that can fool Pangram, and I'll connect that to broader questions about creativity, memorization, and what makes human writing distinct from AI-generated text. (Was this abstract written by an LLM? Find out at the talk!) Bio: Mohit Iyyer is an associate professor in computer science at the University of Maryland, College Park, with a primary research interest in language generation. He is the recipient of multiple best paper awards at NLP conferences, and he also received the 2022 Samsung AI Researcher of the Year award. He obtained his PhD in computer science from UMD in 2017, spent the following year as a researcher at the Allen Institute for Artificial Intelligence, and was then faculty at UMass Amherst until rejoining UMD in January 2025. Apr 25 2025 11.00 - 12.00 Friday, 25th April - 11am: Mohit Iyyer This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. G.03 IF and Teams
Apr 25 2025 11.00 - 12.00 Friday, 25th April - 11am: Mohit Iyyer This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk.
