AI’s Grand Promises Keep Running Into Human Reality
AI’s Grand Promises Keep Running Into Human Reality
The latest AI race is producing two very different kinds of spectacle: agents advertised as tireless digital helpers, and models generating a flood of scientific claims. In both cases, the sales pitch is speed. The practical constraint is still human judgment.
On the consumer side, companies are discovering that an agent can search for a service but stumble when it must actually finish the job. Marissa Mayer, whose startup Dazzle AI routes some requests to human contractors, said frontier models can offer strong suggestions but require checking before the company can stand behind the result. A user asking it to manage a roof replacement, broken window and glass tabletop saw the system confuse window glass with tabletop glass — “trying to figure out if he could make the tabletop fit in the window.”
That is why the human handoff is becoming part of the product, not merely an emergency patch. Duckbill has built an API for agents to delegate tasks such as phone calls and scheduling to contracted workers. Its chief executive, Meghan Joyce, says the company breaks work into “atomic bite-sized pieces” that people can complete consistently. Wajo AI’s Shivani Poddar makes the same argument more bluntly: “AI plus humans, collaborative AI, is what you need to actually get the job done.” But the arrangement raises privacy questions when contractors handle personal requests; Meta has said human-assisted testing is “core to the product development process,” while saying it does not currently use people to complete Muse requests.
The tension is also playing out in research. After OpenAI released hundreds of mathematical results in recent months, researchers complained of sparse context, weak attribution and too little verification. Sam Altman acknowledged the disruption but argued that withholding major findings would be worse: “Humanity deserves fundamental knowledge, and we should not be sitting on a major fundamental science discovery.”
Yet OpenAI has withheld details of at least one claimed breakthrough on expert advice, even as Altman signaled the company has more results in reserve: “We do. Of course.” The contradiction is telling. Whether an AI is booking a repair or proposing a theorem, the hard part is no longer just generating an answer — it is making that answer dependable, intelligible and safe to use.
Write a comment