Julian McAuley

"There is a difference between fixing bias and hallucinations at large and fixing these issues in specific scenarios and applications," McAuley told UC San Diego Today in February 2026, describing a study on which he was senior author.

The paper, led by Abeer Alessa and presented at IJCNLP-AACL 2025, found that language model summaries changed the sentiment of the product reviews they condensed in 26.5% of cases. Choosing examples where the framing shifted sharply, often from negative to positive, the team asked 70 people to read either original reviews or chatbot summaries for products such as headsets, headlamps, and radios. Readers of the summaries said they would buy in 84% of cases, against 52% for readers of the original reviews. The researchers tested 18 ways to reduce the bias, and none worked across every model and task.

In August 2026, McAuley and colleagues including Sheldon Yu, Tong Yu, and Sungchul Kim released Agent2UCB, a demonstration system for generative engine optimization accepted at ICDM 2026. It treats GEO as the practice of refining content so that AI search systems such as Google AI Overviews are more likely to cite or summarize it.

For each piece of content, the system tries nine rewriting strategies and uses a bandit algorithm, which blends a language model's initial guesses with live feedback, to settle on the one that works best. A separate check scores readability, topical coverage, and EEAT-style credibility so that gains in AI visibility do not come at the cost of traditional search quality. On the GEO-Bench benchmark, the authors report consistent visibility gains with SEO quality intact.

McAuley built his early reputation on what customers write. As a postdoctoral researcher at Stanford University from 2011 to 2014, working with Jure Leskovec, he analyzed 42 million online product reviews from 10 million users, including every Amazon review posted through early 2013. "Our approach more accurately predicts product ratings by harnessing the information provided in the text of reviews," he said of the model, which won the $5,000 grand prize in the first Yelp Dataset Challenge.

A related 2013 paper on how reviewers develop expertise drew coverage as far afield as the Beer Street Journal. "We studied over 10 million reviews where users have 'acquired tastes' for products such as beers, wines, gourmet foods and even movies," McAuley explained in a 2014 UC San Diego announcement of his hiring.

McAuley studied software engineering and pure mathematics at the University of New South Wales in Sydney, Australia, and earned a PhD in computer science at the Australian National University in Canberra in 2011, with a thesis on graphical models for computer vision. He interned at Xerox Research Centre Europe and in Google's documents group before moving to Stanford. He joined UC San Diego's Department of Computer Science and Engineering as an assistant professor in 2014 and is now a full professor there.