What Makes an AI Model "Know" You? A New Study Says It's Not Your Résumé

Ask a large language model who you are, and the answer may hinge less on your career achievements than on whether you shipped something with a memorable name attached to it. That is the core finding of a new paper, "The Model Knows Your Project, Not You: Measuring Recognition in LLMs with NameRank," submitted to arXiv on July 14, 2026 and filed under the Artificial Intelligence (cs.AI) category.

The researchers set out to answer a deceptively simple question: what actually determines whether a language model recognizes a given person or entity? Their method, called NameRank, tested 36 different models against 4,685 entities spread across 54 cohorts, likely grouping subjects by field, prominence, or entity type. Each entity was probed with a single open-ended question, and the resulting model output was scored by independent judges issuing binary verdicts against curated gold-standard answers. That process produced a recognition score for every entity, normalized to a range between 0 and 1, allowing direct comparison across models and categories.

Named Artifacts Beat Titles

The headline result challenges an intuitive assumption about how these systems build internal representations of people. Rather than tracking with formal credentials, job titles, or institutional prestige, recognition tracked most strongly with named, indexable artifacts: things like papers, tools, products, or projects that carry a distinct, searchable label. In other words, an LLM is far more likely to "know" someone because they built or authored something with a catchy, unique identifier than because they hold an impressive title or affiliation.

Citations play a meaningful but partial role in this picture. According to the paper, citation counts explain roughly one-third of the variance in whether a model recognizes a given researcher. That leaves a substantial share of recognition unaccounted for by citation metrics alone, reinforcing the idea that named artifacts, rather than academic impact metrics in isolation, are doing much of the explanatory work.

Why This Matters for AI Evaluation

The implications extend beyond academic curiosity. As LLMs increasingly get used for tasks like literature review, expert identification, candidate screening, or even informal fact-checking about individuals, understanding the mechanics of "recognition bias" becomes practically important. If a model's sense of who is credible or notable is shaped primarily by whether that person's work happens to have a distinctive, indexable name, rather than by the substance or rigor of their contributions, that has downstream consequences for fairness and accuracy in any application relying on a model's implicit judgments about people.

The paper does not report any contradictory findings or methodological disputes flagged during review. Its central contribution is the NameRank benchmark itself: a reusable framework for quantifying recognition patterns across model families, offering a way for researchers to audit how LLMs form and reveal these implicit associations at scale.