10
Loading video...
By Len Khodorkovsky
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
Companies are racing into AI faster than they can trust it. Deloitte reports that 74 percent of enterprises expect to be using AI agents—AI that takes action on its own rather than just answering questions—at least moderately by 2027, but only one in five has oversight in place for them. This is already causing problems. A recent industry survey found that half of companies rolled out an AI agent that passed internal testing but still failed once real customers began using it.
I learned why decades ago on a factory floor. For a few summers I inspected chassis at a sheet metal company where my dad, a mechanical engineer for more than 40 years, ran quality control. Most parts passed. Enough didn’t that I learned the blueprint and the part are two different objects, made by two different processes. The gap isn’t a failure of engineering. It’s a permanent feature of it.
My dad traced defects to a few recurring causes: bad raw material, a tolerance that looked fine on paper but failed on the machine, ordinary equipment wear, a worker’s fatigue, an engineer’s miscalculation. It was simply the everyday cost of manufacturing at scale with real materials and real people.
Read More on Opinion
AI breaks down in comparable ways. Bad training data is like bad raw material. A model that aces its benchmarks but fails in the real world is like a tolerance that only looks good on paper. And long AI conversations or long agent chains degrade the same way a worker’s attention does late in a shift. My dad shrugs at all of this. Every factory learned a long time ago that reality humbles perfect designs.
...
In the 1920s, Bell Labs engineer Walter Shewhart gave this idea its proper name: Every process contains natural, expected variation and specific, fixable breakdowns that signal something actually went wrong. My dad never expected zero defects. He expected a tolerable defect rate and checked parts against it. When enough of them failed the same way, we went back to the drawing board. The blueprint hadn’t changed; it told us what should happen. The parts told us what actually happened.
AI is now repeating that lesson. A model that tops a benchmark is the equivalent of an engineering drawing: proof the design can work under controlled conditions. But hospitals, banks and insurance companies operate on dependable results, not blueprints. Right now, the AI industry is only beginning to develop the quality engineering needed to turn blueprints into products institutions can trust.
Law is one flashpoint. A public database tracking court cases with AI-fabricated citations has grown from roughly 200 in mid-2025 to over 1,600 by mid-2026. Medicine is another. A peer-reviewed BMJ Open audit published this spring found nearly half of AI chatbot responses to common health questions were problematic, and one in five was flagged as potentially harmful.
On August 4, the U.K.’s AI Security Institute disclosed that an AI agent tried to trick a software engineer into approving dangerous code using fake online identities. Days later, a powerful Chinese AI model called Kimi K3, already accused by the U.S. of stealing technology, broke out of its own secure test environment and decided to cheat rather than do the work it was assigned. It follows similar cases at OpenAI and Anthropic in recent months.
The good news: humans caught both. The bad news: my dad’s generation could rely on one inspector catching one bad part. AI now produces mistakes, and escapes, at machine speed—and the systems built to test whether it can be trusted aren’t watching closely enough themselves.
A Google Research, Google DeepMind and MIT study published this spring found chains of AI agents with no one checking their work let mistakes pile up 17 times worse than a single AI working alone. Adding a dedicated verification agent reduced errors dramatically.
Even the CEOs of Anthropic, Google DeepMind and OpenAI agree the problem isn’t capability. This summer, all three separately called for independent verification before their own models ship.
That requires what manufacturing learned decades ago: verification built into the process itself, with clear accountability, auditability and continuous measurement.
Toyota built that idea into its factories decades ago. Any worker who spots a problem can pull a cord and stop the line on the spot. The idea, known as jidoka, holds that an imperfection caught immediately costs almost nothing, while the same one caught downstream costs a lot. In other words, don’t hope defects disappear. Design a system that catches them before the customer does.
Michal Kissos Hertzog, CEO of Poalim Tech, makes the same case from inside a bank rather than a factory: “In an era when models can do almost anything, true innovation is measured differently: not by the ability to impress, but by the ability to repeat the same action reliably on a large scale, without surprises or failures."
AI capability moves fast, on something close to a software release cycle: fast, cheap, competitive. Verification moves on a different clock because it requires organizations to redesign processes, assign accountability and build verification into everyday work. Manufacturing became transformative only after those two disciplines separated. AI is only beginning that transition.
Stanford’s 2026 AI Index, the field’s most cited annual report, reached a blunt verdict: AI’s capabilities are advancing quickly; our ability to measure and manage them is not.
My dad belonged to a generation of engineers that judged success by what left the loading dock, not by what looked good on the drafting table. They assumed defects were inevitable and built systems to catch them before customers ever saw them. AI’s biggest bottleneck isn’t intelligence anymore. It’s trust. And trust isn’t another model breakthrough. It’s the product of quality engineering.
Len Khodorkovsky is the senior adviser to the chairman of the Krach Institute for Tech Diplomacy at Purdue. He previously served as U.S. Deputy Assistant Secretary of State for Digital Strategy.
The views expressed in this article are the writer's own.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines