They also hoover up valuable data from users through the use of tools like reCAPTCHA, which ask visitors to solve problems that are easy for humans but hard for AIs, such as deciphering text from books that machines are unable to parse.
That does not just screen out malicious bots, but also helps digitise books.
People “pay” for useful free services by providing firms with the data they crave.
These data become part of the firms' capital, and, as such, a fearsome source of competitive advantage.
Would-be startups that might challenge internet giants cannot train their AIs without access to the data only those giants possess.
Their best hope is often to be acquired by those very same titans, adding to the problem of uncompetitive markets.
That, for now, AI's contributions to productivity growth are small, the authors say, is partly because of the free-data model, which limits the quality of data gathered.
Firms trying to develop useful applications for AI must hope that the data they have are sufficient, or come up with ways to coax users into providing them with better information at no cost.
For example, they must pester random people—like those blur-deciphering visitors to websites—into labelling data, and hope that in their annoyance and haste they do not make mistakes.
Even so, as AI improves, the amount of work made vulnerable to displacement by technology grows, and ever more of the value generated in the economy accrues to profitable firms rather than workers.
As the authors point out, the share of GDP paid out to workers in wages and salaries—once thought to be relatively stable—has already been declining over the past few decades.