ChatGPT can't tell a fact from a casual opinion

A statistical parrot with delusions of grandeur: deflating the AI hype

Generative AI doesn't think: it predicts the next word. It sounds like a barroom quip, but it's the conclusion anyone reaches who tightens the screws on ChatGPT by asking for a number without adjectives. The system answers. It answers fast, with flawless syntax and a confidence it doesn't have. That's where the problem starts.

What's behind an answer that sounds like the truth

There's no magic. There's token prediction, millions of examples and a layer of human labeling — outsourced, badly paid and opaque — that turns plausible mediocrities into paragraphs with the tone of a quarterly report. The economic consequence is uncomfortable: for two years we've been told these tools would replace half the workforce, and the result is that they confuse a fact with a barstool opinion. You ask for rigor and you get a sarracena lesson.

The cheapest test fits in one line: ask it for the public deficit figures without adjectives, without context and without filler sentences. That's where it all goes wrong. The failure doesn't come from a lack of data, but from an architecture that rewards the answer that pleases over the one that adds up. That sycophancy bias turns every flattering reply into an analysis nobody actually did.

Petrol or Chinese electric: a doubt that isn't about cars

The same pattern shows up when choosing a vehicle. A petrol sports car offers fun today, decent resale value and a countdown: in five years, driving through the center of any mid-sized city will be an administrative exception. A Xiaomi SU7 offers a tech toy today and total dependence tomorrow: it's a smartphone on wheels tied to someone else's cloud, with the battery as a deferred cost and an update that could leave two tons of steel turned into a paperweight.

The real question isn't the engine. It's where you plug it in and who repairs it when it breaks. Infrastructure and after-sales service, the two line items that never appear in the brochure.

The predictable outcome is that generative AI ends up being what it already is: a useful, noisy, oversold tool. The predictable outcome is that the electric car wins by regulation, not by love. Neither will happen at the pace their ads promised, though at these prices and with these margins, it doesn't pay to bet against them either.