AI Can't Pass 'Humanity's Last Exam', and That Tells Us More Than Any Benchmark Score

Why Traditional Benchmarks Miss the Mark

The deeper problem exposed by Humanity's Last Exam is not simply that AI struggles with hard questions. It is that the entire benchmark-driven development model is structurally flawed. AI developers routinely use benchmarks as training targets, optimising their models to achieve high scores. This might look like progress on a leaderboard, but it frequently produces systems that are better at specific test formats rather than demonstrating adaptable, genuine intelligence.

Since the exam's online publication in early 2025, AI scores have climbed steadily. Rather than celebrating this, researchers have pointed out that the improvements likely reflect models becoming more adept at the particular question styles featured in the exam, not any meaningful leap in underlying reasoning. The distinction matters enormously for anyone making procurement or deployment decisions in a European organisation.

Professor Neil Lawrence of the University of Cambridge, who has written extensively on the gap between benchmark performance and real-world AI capability, has argued consistently that optimising for narrow metrics produces systems that excel at the test while failing on the genuinely novel problems that matter in practice. His concern is well founded here. When a model learns to navigate the surface features of a benchmark, it is not acquiring understanding; it is acquiring test-taking skill.

The European AI Office, established under the EU AI Act framework and operational since 2024, has similarly cautioned against over-reliance on benchmark scores when assessing AI systems for high-risk deployments. Its technical guidance emphasises that conformity assessments must include real-world performance evidence, not merely standardised test results. That regulatory pressure is, gradually, pushing European vendors and buyers towards more honest evaluation frameworks.

The Shift Towards Real-World Assessment

Recognising the limitations of benchmark-driven development, parts of the industry are beginning to move in a more useful direction. OpenAI has introduced GDPval, a metric designed to assess the real-world usefulness of AI by evaluating performance on practical professional tasks: drafting project documents, conducting data analyses, producing deliverables common in workplace environments. It is an imperfect instrument, but it represents a more honest attempt to measure what actually matters to organisations deploying these tools.

The comparison is instructive:

• Traditional benchmarks focus on academic knowledge and offer limited practical application.
• GDPval metrics target professional tasks and carry direct workplace utility.
• Domain-specific tests address industry requirements and are highly relevant for specialists.
• User-defined criteria assess personal workflows and deliver maximum practical value.

For European organisations, particularly those operating under the EU AI Act's requirements for transparency and human oversight in high-risk contexts, this shift is not merely academic. Choosing an AI tool because it scored well on a graduate physics benchmark and then deploying it in, say, a medical diagnostic or legal research context is not just strategically naive; it may create compliance exposure.

A Practical Adoption Strategy for European Organisations

Across Europe, the most forward-thinking organisations are already taking a more pragmatic approach to AI evaluation. Rather than chasing benchmark scores, they are focusing on practical applications that deliver measurable, auditable business value. The priorities tend to cluster around several themes:

• Regulatory compliance and data governance, particularly under the General Data Protection Regulation and the EU AI Act.
• Hybrid infrastructure models that balance performance with sovereignty requirements, an area where French AI lab Mistral AI has positioned itself as a credible alternative to US hyperscalers for European enterprises that need to keep data within EU jurisdiction.
• Domain-specific applications tailored to sector requirements, whether in healthcare, financial services, or manufacturing.
• Integration with existing workflows rather than wholesale replacement of human roles.
• Return-on-investment measurement grounded in operational outcomes, not theoretical capability.

This approach reflects a mature understanding that AI's impact varies significantly across contexts. A model that performs brilliantly at coding assistance may be mediocre at summarising legal contracts. Benchmark scores collapse that variation into a single number, and single numbers lie.

How to Evaluate AI for Your Specific Needs

The practical implication for any European organisation currently selecting or reviewing AI tools is straightforward: define what you genuinely need AI to accomplish, then test different models against those specific criteria using representative samples of your own data and tasks.

Accuracy within your domain, integration capability with existing systems, cost-effectiveness over a realistic deployment horizon, and alignment with your organisation's governance and compliance requirements should all carry more weight than a position on a public leaderboard. The question is never whether an AI can solve a graduate-level problem in ancient Sumerian; the question is whether it can reliably support the tasks that determine outcomes in your organisation.

ETH Zurich's AI Centre, one of Europe's leading applied research institutions, has published evaluation frameworks specifically designed to help organisations move beyond generic benchmarks towards task-specific assessment. Their work reinforces the point: rigorous, domain-grounded evaluation is both feasible and necessary, and European research infrastructure is well placed to support it.

The broader lesson from Humanity's Last Exam is not that AI is failing. It is that the way we measure AI has been failing us. Correcting that is not a minor methodological adjustment; it is a prerequisite for making genuinely informed decisions about one of the most consequential technologies in current deployment. European organisations, backed by a regulatory framework that demands exactly this kind of rigour, are better positioned than most to lead that correction.