The rapid advancement of artificial intelligence is being rigorously tested against the benchmark of human knowledge, with current models demonstrably falling short of expert-level performance on complex academic questions. This assessment, encapsulated in a project known as “Humanity’s Last Exam,” is prompting a reassessment of the current trajectory of AI development and raising questions about the nature of intelligence itself.
Launched in early 2025, Humanity’s Last Exam (HLE) comprises 2,500 questions spanning a vast range of disciplines, from mathematics and the humanities to the natural sciences. The benchmark is designed to be a challenging, closed-ended academic assessment, intended to represent the pinnacle of human understanding. According to the project’s creators at the Center for AI Safety and Scale AI, HLE isn’t simply about testing AI’s ability to recall information, but rather its capacity for nuanced reasoning and problem-solving.
Recent results indicate that even the most sophisticated AI models are struggling to achieve high scores on the exam. While the project acknowledges that success on HLE wouldn’t automatically equate to artificial general intelligence (AGI) or autonomous research capabilities, the findings highlight significant limitations in current AI systems. The exam’s difficulty lies not in requiring specialized knowledge, but in demanding a broad understanding and the ability to apply that understanding to novel situations.
The emergence of HLE coincides with a broader debate about the impact of digital technology on human cognition and well-being. A panel discussion held at Harvard University in October 2025, titled “How is digital technology shaping the human soul?,” explored these themes, with experts from computer science and the humanities weighing in on the potential consequences of increasing reliance on AI. The discussion, part of the Public Culture Project, centered on whether technology ultimately enhances or diminishes human fulfillment.
Nataliya Kos’myna, a research scientist with the MIT Media Lab, presented findings from a study involving 54 students in the Greater Boston area. The study monitored brain activity via electroencephalography while students wrote essays on topics such as “Is there true happiness?” Students were divided into three groups: one allowed to use ChatGPT, another with access to the internet and Google, and a third relying solely on their own intellect. The results were striking. The group utilizing ChatGPT exhibited “much less brain activity” compared to the other two groups. The essays produced by the ChatGPT group were remarkably similar, largely focusing on career choices as the primary determinant of happiness.
Kos’myna’s research suggests that readily available AI tools, while capable of generating text, may stifle original thought and critical analysis. This observation echoes concerns raised by other experts about the potential for AI to promote superficiality and conformity, rather than fostering genuine intellectual exploration. The Harvard panel questioned whether humanity is becoming “tech people,” and whether the tools designed to improve life are, in fact, hindering the pursuit of happiness and fulfillment.
The implications of these findings extend beyond the academic realm. As AI systems become increasingly integrated into various aspects of life – from education and healthcare to finance and governance – understanding their limitations becomes crucial. Humanity’s Last Exam provides a clear measure of AI progress, offering a common reference point for scientists and policymakers to assess capabilities and potential risks. This, in turn, can inform the development of more responsible and effective governance measures.
The benchmark’s creators emphasize that HLE is not intended to be a definitive test of AI’s ultimate potential, but rather a tool for fostering informed discussion and guiding future research. The project’s dynamic fork version, released in August 2025, encourages contributions from the wider AI community, aiming to continuously refine and expand the exam’s scope and complexity. The ongoing development of HLE underscores the importance of a collaborative and critical approach to AI development, one that prioritizes human values and long-term well-being.
While AI continues to demonstrate impressive capabilities in specific domains, the challenge of replicating the breadth, depth, and adaptability of human intelligence remains significant. The results of Humanity’s Last Exam serve as a reminder that true intelligence encompasses more than just information processing; it requires creativity, critical thinking, and a nuanced understanding of the world – qualities that, for now, continue to distinguish human minds.
The conversation surrounding AI’s limitations is not about halting progress, but about directing it responsibly. As one Harvard researcher noted, humanity has a long history of creating tools to enhance life, but not always to enhance happiness. The current moment demands careful consideration of how these tools are shaping not only our lives, but also our very souls.
