AI has develop into excellent at passing the checks we set for it.
Stanford College’s 2026 AI Index Report captures the issue neatly. Whereas AI fashions can grasp summary logic, they usually struggles with fundamental spatial reasoning duties. As an illustration, a number one AI mannequin might win gold on the Worldwide Mathematical Olympiad, but it might accurately learn an analogue clock solely half the time.
That’s the paradox of AI. Distinctive in a single area. Unreliable in one other.
That unevenness issues as a result of AI isn’t marketed as conditional, solely as succesful.
We’ve seen the identical sample play out on a number of the greatest phases in tech. As an illustration, Tesla’s Optimus robots had been introduced as a glimpse of autonomous robotics, however after their look at a 2024 occasion it was reported they relied on human intervention for some capabilities. Meta’s AI glasses failed twice throughout dwell demos, with the corporate later pointing to technical points.
The main points differ, however the sample is constant: AI can look succesful in managed circumstances after which behave very in another way when it meets the messiness of the true world.
The identical hole is exhibiting up throughout industries. MIT’s Venture NANDA Gen AI study discovered that 95% of organizations are getting zero return, with most programs caught with out measurable P&L affect.
The query value asking is less complicated than it sounds: is your AI being measured in your actuality, or another person’s?
THE AGE OF BENCHMAXXING
Each aggressive business finally learns to optimize for its scorecard. AI is not any totally different. Now there’s even a reputation for it: benchmaxxing.
Benchmarks exist for good causes. They create a standard language. They make comparability attainable. The issue emerges when the benchmark turns into the goal slightly than the measurement. As soon as that occurs, optimizing for a take a look at and genuinely bettering functionality develop into two totally different actions.
Benchmarks themselves aren’t as stable as they appear. A 2025 study discovered that giving builders even restricted entry to check information might enhance leaderboard scores by as much as 112%. In the meantime, a February 2026 paper discovered almost half of broadly used benchmarks have hit saturation, which means high fashions rating so equally that the checks can not distinguish between them.
In my nook of the business, I hear the identical claims a number of instances a day. World’s greatest. Quickest. Most correct. Distributors are discovering ever extra artistic methods to outdo one another on no matter metric is at present in trend. What’s much less seen is what these comparisons omit: the fashions that didn’t make the reduce, the take a look at circumstances that weren’t disclosed, the consumer populations that had been by no means included within the first place.
A benchmark can inform you how a mannequin performs on a take a look at. What it can not inform you is how that mannequin performs beneath your circumstances, together with your customers, on the issues that really matter to your small business.
THE SHOWROOM AND THE ROAD
Demos are designed to showcase strengths. Which means, by definition, eradicating the circumstances that create issues in manufacturing: managed environments, predictable inputs, rigorously chosen use instances, customers who behave precisely as anticipated. No person demos the sting instances.
The hole this creates is effectively documented and nearly universally underestimated. Latest evaluation confirms it: most groups uncover the onerous method, after a prototype that dazzled stakeholders begins silently degrading in manufacturing. In my house—voice AI—the metrics chosen for demos are a part of the identical drawback. Distributors lead with the numbers they’re assured successful on. The measures that may reveal weaknesses, how a system performs beneath strain, with tough inputs, at scale, have a tendency to not make it onto the slide.
Demos should not dishonest. However they’re incomplete by design, and patrons hardly ever have sufficient data to know the place the demonstration ends and the precise product begins.
WHO WAS THIS BUILT FOR?
Benchmark saturation isn’t the one blind spot; illustration is the opposite.
Each analysis framework makes decisions about who it contains. These decisions decide whose expertise is measured, who it’s optimized for, and who’s quietly handled as an edge case.
It is a large sticking level in voice AI. A system may carry out effectively in a clear take a look at set and nonetheless battle with the best way individuals really communicate: regional accents, dialects, code-switching, overlapping speech, background noise, interruptions, mumbling, laughter, emotion, technical language, older audio system, second-language audio system.
The chance is {that a} purchaser sees a single accuracy quantity and assumes it represents everybody they serve. It hardly ever does. A mannequin skilled and examined on slender circumstances will look sturdy for the individuals most represented in that information. It could carry out very in another way for everybody else.
That hole doesn’t all the time present up as an apparent failure. It reveals up as extra corrections, extra friction, extra abandonment, extra human evaluate, and worse outcomes for the customers least seen within the analysis.
WHAT TO DO ABOUT IT
Benchmarks are helpful alerts. Deal with them as a place to begin, not a conclusion.
Earlier than any procurement determination, ask distributors to display efficiency in your particular use instances, together with your precise consumer inhabitants, beneath circumstances that resemble manufacturing slightly than a demo script. If they can not present it, that hole within the proof is itself the reply.
Check towards edge instances intentionally. The customers most definitely to be underserved by a system are hardly ever those centered within the demo. Embody them in analysis from the beginning.
Then preserve measuring after deployment. AI efficiency is just not a hard and fast level. It degrades, drifts, and surprises.
The organizations extracting actual worth from AI should not those that ran one of the best procurement course of. They’re those that handled go-live as the start of analysis, not the top of it.
The query is not whether or not AI can cross the take a look at. It’s whether or not the take a look at resembles actuality.
Katy Wigdahl is CEO of Speechmatics.
