If an AI can read a medical scan better than any radiologist, why does it slow hospitals down?
This is the uncomfortable question at the center of a new analysis from MIT Technology Review, and the answer exposes a crisis in how the entire field measures AI progress. The short version: current benchmarks measure the wrong things. They reward synthetic task performance while systematically ignoring what actually matters—whether AI delivers value when embedded in real human teams and organizational workflows.
The Benchmark Illusion
For decades, the AI field has evaluated systems through a simple framework: can a machine outperform a human on a defined task? Chess. Math problems. Coding benchmarks. Essay quality scores. This approach generates clean numbers, sortable rankings, and compelling headlines. An AI that achieves 98% accuracy on a radiology task, reading scans faster and more precisely than any human expert, looks unambiguously superior.
But here's what those numbers miss. In hospital radiology units from California to London, researchers observed something the benchmarks never capture: staff required *extra* time to interpret AI outputs alongside hospital-specific reporting standards and national regulatory requirements. The AI technically outperformed radiologists on the benchmark. In practice, it added friction.
This is not an isolated failure. It is a structural problem. Researchers studying real-world AI deployment since 2022—in small businesses, healthcare systems, humanitarian organizations, and higher education across the UK, US, and Asia—consistently find the same pattern: benchmark performance and real-world value diverge.
Why Standardization Fooled Everyone
The appeal of current benchmarks is understandable. They are standardized, comparable, and objective. A score of 87% on MMLU means the same thing regardless of who runs the test. This transparency makes them invaluable for model selection and funding decisions. Organizations trust benchmark scores more than vendor claims because numbers feel scientific.
But this standardization creates a dangerous illusion of predictive validity. Benchmarks succeed as measurement tools precisely because they isolate variables—removing the messiness of human collaboration, organizational constraints, and extended time horizons. The moment an AI enters a real deployment environment, all those removed variables come flooding back.
AI is almost never used the way it is benchmarked. In production, it operates within workflows, alongside colleagues, subject to institutional norms and regulatory frameworks. Its true performance—or failure—only emerges over extended periods of use. Current benchmarks cannot see any of this.
What the Field Actually Needs
The proposed alternative, dubbed HAIC (Human-AI Context-Specific Evaluation), shifts the measurement paradigm entirely. Instead of asking "can AI do X better than a human?", HAIC asks "does AI improve outcomes when deployed within human teams over time?"
This is a harder question to answer. It requires longitudinal studies rather than one-shot tests. It demands evaluation within specific organizational contexts rather than controlled lab settings. It measures workflow integration, user trust, error recovery patterns, and value emergence—metrics that resist easy comparison across organizations.
But difficulty is not a reason to keep measuring the wrong things. The costs of benchmark-driven deployment decisions are already visible: organizations committing financial and technical resources to AI systems that underperform expectations, systemic risks overlooked because benchmarks never flagged them, and fundamental misalignment between what AI can do and what AI should do in human contexts.
The benchmark illusion will not correct itself. Until the field builds evaluation frameworks that capture real-world generalization—how AI actually functions within teams, workflows, and organizations—the scores will keep looking better than the reality.