Habr: AI Made Forecasts Cheap, but Checking Them Still Takes Hours of Human Work

A report on a nonexistent vulnerability takes minutes and a few cents to write. A living maintainer of an open-source project such as curl has to spend hours checking it. The author of a recent Habr analysis takes this asymmetry from a neighboring article about code and applies it to a different commodity: forecasts about AI. His thesis is that the next line item in the AI bill will go not to code but to predictions about it.

Where the imbalance comes from

The starting point is an article arguing that code has a cost of writing and a cost of ownership, and that the two have drifted apart. The author of the new text says outright that he borrows the accounting from there. He separately notes the first author's honesty in handling sources, especially the Veracode reports. His own addition is to carry the mechanism over to public discourse, and he promises to show the cost of ownership on two anonymized projects. That part sits behind a cut, so it cannot be judged from the teaser.

An important caveat: forecasting has always been cheap, from science fiction to fears of automation, long before neural networks. Generative models did not create the genre. They removed the last remaining expense, the cost of packaging. A note with figures and an expert's tone can now be turned out in bulk for pennies' worth of tokens.

Where a forecast gets its anchor

Comments on the original article objected that verification is getting cheaper too. For code, that is true. Code has an independent anchor: a compiler, tests, a reproducible exploit. The answer does not depend on the model's opinion.

A forecast has no such anchor. There is no test that fails if the promised outcome does not arrive. There is also no party that pays for being wrong. So it is not only text that is cheap to produce, but unaccountability as well. Meanwhile, verification stays the same: find the primary source and find out who paid for the research.

The mechanisms side by side:

• Producing a report or forecast: minutes and a few cents of tokens.
• Verification: hours of a living person's work.
• Anchor for code: tests and a reproducible result.
• Anchor for a forecast: absent; there is no penalty for being wrong.
• What has to be established by hand: the primary source and the commissioner of the research.

What to do about it

Editorial teams, analysts, and investors who read about "disappearing professions" and "the next leap" would benefit from a simple discipline. Before quoting a figure, trace it to the primary source and work out who funded the work. Then ask whether the forecast contains a verifiable condition and a deadline. If not, it is an opinion, not a forecast. Texts whose confident tone is not backed by a single link to data deserve separate flagging.

It is better to build responsibility for verification into the process in advance, as projects already drowning in automated reports do. For maintainers, that means filters and reproducibility requirements. For media, it means a rule against publishing a figure without a source that has actually been found.

The piece has a limitation. It is an opinion column, not a study: the visible part contains little measurable data on the cost of verification, and the key examples from the two projects remain behind the cut. As a frame for discussing the quality of information around AI, it works. As proof, it needs figures that are not yet visible.