A remarkable share of public discussion about artificial intelligence concerns systems that do not exist. Whether general intelligence arrives this decade or next, whether it poses an existential risk, what a superintelligence would want. These are not unserious questions. They are also unfalsifiable, which makes them unusually pleasant to argue about.
What the argument displaces
Meanwhile, systems that do exist are deciding who gets a loan, which claims are investigated, which applicants reach a human, and what a moderator sees. Those systems have measurable error rates, identifiable people affected by them, and owners who can be named. Every one of those is a harder conversation to have than a speculative one, because it involves telling a specific organisation that a specific thing it does is not good enough.
Why the far-future frame suits the industry
It positions the technology as world-historically important, which is good for raising money. It locates the danger in a future system rather than a shipped one, which defers accountability. And it casts the people building it as the responsible parties best placed to manage the risk, which is an argument for regulating them lightly and consulting them heavily.
None of that requires bad faith. Incentives do not need anyone's cooperation to work.
The questions worth the airtime
What is this system's error rate for the group it affects most? Who is accountable when it is wrong, and what does the person affected get? Who did the labour behind it and on what terms? What does it cost, in energy and in water, and who bears that cost? What happens to the data?
These are all answerable. That is precisely what makes them uncomfortable, and it is why a publication covering this field should keep asking them.
