PhantomBench: Benchmarking the Non-existential Threat of Language Models

A new benchmark called PhantomBench reveals that large language models routinely fail to recognize when they are being asked about non-existent concepts, with average hallucination rates reaching 86.7% in some evaluations, according to research submitted to arXiv on June 9, 2026 [1][2]. The benchmark, described as the first large-scale tool of its kind, comprises more than 60,000 non-existent terms and entities derived from real concepts across diverse domains [1][2]. Researchers used PhantomBench to evaluate 21 language models of various types and sizes [1][2]. The results showed that models frequently generated factually ungrounded responses rather than abstaining, a behavior the authors describe as a serious risk given that users tend to blindly rely on model outputs [2]. The problem was especially pronounced when the input question presumed the existence of the fabricated concept [1][2]. Hallucination in language models has been a persistent challenge in the field. The PhantomBench paper notes that despite progress in understanding why models produce ungrounded content, it has remained unclear how reliably these systems can recognize the boundaries of their own knowledge [2]. The benchmark addresses this gap by testing whether models will decline to answer when confronted with queries about things that do not exist. The findings indicate that even frontier models — the most advanced systems available — surprisingly fail to abstain on non-existent concepts [1][2]. The research also positions PhantomBench as a proxy for studying model behavior on rare concepts, where models are more prone to hallucinate [2]. The authors provide a pipeline for constructing the benchmark, which they say enables scalable generation of non-existent concepts tailored to the specific needs of other researchers and practitioners [1][2]. This design choice could allow the benchmark to be adapted for domain-specific evaluations, such as in medicine or law, where the consequences of fabricated information can be especially harmful [2]. The work arrives amid broader scrutiny of artificial intelligence reliability. While the PhantomBench paper does not reference specific regulatory frameworks, the United Nations Sustainable Development Goals highlight the importance of trustworthy technology in areas such as health, education, and climate action, all of which increasingly rely on automated information systems [6]. The benchmark's focus on knowledge-limitation awareness aligns with calls for greater accountability in AI deployment, though the paper itself remains a technical contribution to the field of computational linguistics [1][2].