Axios and OpenAI Are Automating the Cheap Half of Local News
Jim VandeHei's theory of what went wrong in local news is that the expensive part was never the reporting. "Local newsrooms of old had high fixed costs," he told Issie Lapowsky for a piece in the Columbia Journalism Review this month. "Buildings. Paper. If your cost was essentially the journalism, that's relatively cheap." That premise is the whole foundation of Axios's three-year deal with OpenAI, which bankrolled thirteen new Axios Local newsletters in exchange for the right to train on Axios's free published coverage, and which VandeHei frames as a test of one question: "Could we use AI to basically automate everything you need in a local market other than the journalist and the journalism itself?" By the end of this year Axios expects to be in forty-three cities. He says he would like to someday be in a thousand.
The buildings and the paper were gone before any of this started, though, and the cost that replaced them is not one an AI tool touches. Look at what the arrangement actually asks of a reporter. Robert Sanchez, one of the journalists Lapowsky follows, writes the Arapahoe County newsletter on Mondays and Wednesdays and the Douglas County newsletter on Tuesdays and Thursdays. Every local byline in both is his. Together those two counties cover more ground than the state of Rhode Island, and competing outlets already have multiple reporters assigned to cities inside his beat. An Axios Local newsroom is one to four journalists and an editor who covers several markets at once, everyone working remotely. Journalism is only cheap in that arrangement if what you are counting is bylines. If what you are counting is coverage — somebody in the room at a county commission meeting on a Tuesday night when the school board across the county line meets the same hour — it is not cheap at all, and no model makes one reporter two.
The AI in the story is real and it is genuinely useful. Sanchez used ChatGPT to draft a Colorado Open Records Act request and identify the right recipient, ran it through a legal filter Axios built to reduce the odds of rejection, then fed the returned documents back through to check whether the agency had withheld anything. It had. He had the three missing files by the end of the day. That is a good afternoon's work compressed into an hour, and any reporter who has waited three weeks on a records officer would take it. It is also, precisely, desk work. This is the part of the job I spend most of my time on at latakoo, where the problem I care about is what it costs to get a person and their material from a place back into a newsroom in time for it to matter — travel, bandwidth, the dead hours between the thing happening and the file being usable. The share of a local reporter's week that a language model can absorb is the share that was already the cheapest. The expensive share is geography, and geography does not compress.
There is a second thing in the deal that my own research makes me watch closely. In a study published this year in Political Research Quarterly with Valerie Martinez-Ebers and Aida Ramusovic, we found that how much a national outlet covers a story is not set independently — coverage frequency moves in response to what the competition is already doing, and it moves differently depending on the topic. Outlets take their cues from each other. I would expect capital to behave the same way, and the pattern in this deal is consistent with that: with the possible exception of Huntsville, which has no daily newspaper, none of the four starter cities Axios chose — Pittsburgh, Kansas City, Boulder, Huntsville — are what researchers classify as news deserts. Sanchez's own beat has competitors on it. VandeHei says the true news deserts would be "the Holy Grail" and that Axios first has to prove the economics in established markets. That is a defensible sequence for a business. It is a much weaker basis for the claim that AI will save local news, because a market chosen for having an audience, an ad base and a rival to beat is not the market the crisis is named after.
The last piece is the one my documentary work makes me flinch at. Building John Johnson Reporting meant negotiating archive access with ABC and Disney chapter by chapter, and the film that exists is the one the archive allowed, not the one we outlined. What you can use in year ten is decided by what somebody signed in year one. Axios has granted OpenAI training rights on its free published coverage for three years of startup funding, and neither company would say whether the agreement precludes Axios from taking legal action against OpenAI later. Maybe that is a fair transaction price, as VandeHei says. Nobody outside the two companies can tell, which is the part worth noting.
I would rather this experiment exist than not, and I think Matt Pearce of Rebuild Local News is right that it solves one company's problem rather than the industry's. But the experiment being run and the claim being sold are two different things. Axios is testing whether a very lean newsroom can turn a profit in a city that already has news. "AI will save local news" is a claim about the cities that do not. Only one of those is being tested, and it is not the one the phrase is doing work for.