We've just announced IQ AI.
Julian Michael is an American researcher in artificial intelligence, specializing in AI safety, evaluation, and alignment. He is currently a researcher at Meta, where he works on AI alignment within the company's Superintelligence unit. [1] [2]
Julian Michael attended the University of Texas at Austin from 2011 to 2015, where he earned an undergraduate degree in computer science. He then pursued graduate studies at the University of Washington in the Computer Science & Engineering department, completing his Ph.D. in 2022. His doctoral thesis was titled "Building Blocks for Data-Driven Theories of Language Understanding," and his advisor was Luke Zettlemoyer. Following the completion of his doctorate, Michael worked as a postdoctoral researcher at New York University's Center for Data Science from 2022 to 2024, under the advisement of Samuel R. Bowman. [3] [4]
Michael's career has spanned roles in both academic research and the technology industry. After his postdoctoral fellowship at New York University, he joined Scale AI to lead its Safety, Evaluations, and Alignment Lab (SEAL). The lab's mission was to conduct research focused on safeguarding the behavior of AI systems and ensuring they amplify human agency.
In mid-2025, Michael announced his departure from Scale AI to join Meta. This move was part of a larger transition that saw Scale AI's co-founder and CEO, Alexandr Wang, leave to head a new Superintelligence unit at Meta following a significant investment by Meta in Scale AI. Michael, along with other key talent from Scale AI such as Head of Research Summer Yue, joined Wang's new team at Meta to continue working on AI safety and alignment.
Michael's research is primarily centered on AI alignment, the formal semantics of natural language, and the empirical methods used to understand intelligent systems. His work often involves creating new datasets, benchmarks, and methodologies for evaluating and training AI models.
A significant focus of Michael's work is on AI alignment, particularly scalable oversight, which refers to methods for supervising AI systems that are more capable than humans. He has explored the use of debate as a paradigm for both training and evaluating AI. The goal is to create a process where two AI systems debate a topic, and a human judge can determine the correct answer more easily by observing the debate than by solving the problem directly. This approach aims to ensure that AI systems help users find truth rather than simply generating persuasive-sounding arguments. His work in this area includes human experiments to validate debate as a truth-seeking process.
His work in this domain also touches on issues like deceptive alignment, where a model might appear aligned during training but behave differently after deployment. He has also contributed to research on mitigating "jailbreaks" in large language models and studying how models can be taught to identify and verbalize instances of reward hacking.
In the field of Natural Language Processing (NLP), Michael has concentrated on using machine learning and data-driven approaches to advance the scientific understanding of language, particularly in syntax and semantics. His PhD thesis laid out a paradigm for a "scalable, data-driven theory" of language, which argues for using empirical methods to build and test linguistic theories. A paper summarizing this work won a Best Paper award at The Big Picture Workshop.
To build the foundations for this approach, he has developed novel methods for crowdsourcing complex linguistic annotations. A key contribution is his work on Question-Answer Semantic Role Labeling (QA-SRL), a framework that represents the predicate-argument structure of sentences through question-answer pairs. This method makes it easier for non-experts to provide detailed semantic annotations, enabling the creation of large-scale datasets. His research has also explored inducing semantic roles from text without relying on syntactic parsers.
Michael has made significant contributions to the evaluation of AI models. He was involved in the creation of the diagnostic set for the General Language Understanding Evaluation (GLUE) benchmark, which provides a fine-grained analysis of model performance across a range of linguistic phenomena.
More recently, he was part of the team that developed GPQA, a "Graduate-Level Google-Proof Q&A Benchmark." This benchmark consists of challenging multiple-choice questions in biology, physics, and chemistry written by domain experts. The questions are designed to be difficult for even advanced AI models to answer correctly using standard search engine queries, thus testing their reasoning capabilities more rigorously. He has also worked on explicitly incorporating ambiguity into task design, as seen in the AmbigQA benchmark, which challenges models to produce multiple plausible answers for ambiguous questions. [13] [14] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [16]
In addition to his formal publications, Michael has written essays and blog posts on various topics related to AI and science. These include a detailed review of the OpenPhil "Biological Anchors" report on forecasting transformative AI timelines, an analysis of the form-versus-meaning debate surrounding language models, and philosophical essays on the semantics of imperative sentences and whether inflation theory in cosmology qualifies as science. [15]