Why the people building AI won’t touch it

Artificial intelligence (AI) and robotics aren’t inherently bad, but the people pushing this technology revolution are the richest people in the world. They suggest that Musk and Bezos, for example, aren’t staying up at night worrying about ordinary people. They aren’t worrying about working people. They’re not spending hundreds of billions of dollars to extend life expectancy, address global warming, shorten the workweek, and guarantee high-quality health care. They simply want more wealth and more power.

If a chef refused to eat their own cooking, a boat builder refused to board their own boat, or the chief of an aircraft manufacturer never flew in their own planes, you’d have second thoughts too.

You might be surprised to hear that training AI is actually a very human, manual and grubby business.

And almost without exception, the people at the coalface are the very people now telling their families, friends, and even their own children to stay away from the technology they help build.

According to the Guardian, the humans closest to the models have zero trust in them. They see the constant hallucinations, hidden biases, and outright dangerous outputs up close. Consequently, many have banned generative AI in their own homes and actively discourage others from using it.

The interviewees also noted that speed is crushing safety. Relentless pressure for fast turnarounds, vague instructions from AI industry bosses, and unrealistic deadlines meant quality is being sacrificed so companies can release new features and versions, maintain the perception of rapid progress and keep stock valuations rising.

According to the interviewees, AI hallucination rates are worsening, not improving. Indeed, an independent NewsGuard audit conducted between August 2024 and August 2025 revealed the top 10 models now almost never say “I don’t know” (down from 31 per cent last year to 0 per cent this year), while the rate of AI models confidently repeating dangerous falsehoods nearly doubled (from 18 per cent to 35 per cent).

Worryingly, sensitive topics, including medical questions, historical controversies, and hate-speech detection, are being handled by unqualified, low-paid contractors with no domain expertise, paid only cents per task, and racing against the clock. It stands to reason that if the data fed into these models is rushed, inconsistent, and frequently toxic, the people cleaning it must conclude the final product can never be truly trustworthy.

According to the same frontline workers, AI’s feedback loops are broken, and reported biases and errors remain unaddressed. The Guardian flagged an example where Gemini refused to answer basic questions about Palestinian history while happily giving long answers about Israel.