When a local newspaper closes, the town pays more to borrow money. Three finance researchers, Pengjie Gao of Notre Dame and Chang Lee and Dermot Murphy of the University of Illinois Chicago, went looking for this in municipal bond data and published what they found in the Journal of Financial Economics in 2020. In the three years after a closure, offering yields on that municipality’s bonds rose 5.5 basis points, and secondary market yields rose 6.4. For revenue bonds, where the risk of a badly run project actually lands on investors, the effect was closer to 10 basis points. On a $65 million issue with a ten-year duration, ten basis points runs about $650,000. The authors also checked the obvious objection, which is that online outlets absorb the work, and found no sufficient substitution.
The part worth sitting with is not the number. It is that nobody at the newspaper was ever measured on municipal borrowing costs. No editor ran a report on the spread between what the county would have paid and what it did pay. The monitoring function was real, it was expensive to lose, and it was invisible to every metric the business ran on. When the paper folded, the cost did not show up on the paper’s ledger, because the paper no longer had one. It showed up on the county’s.
It is fair to ask why a piece about AI in media opens with municipal bond spreads. The answer is that the newspaper closure is the rare substitution where somebody went back and priced what disappeared. Media companies are making the same category of decision right now on a much faster clock, choosing which human functions AI can absorb in ad ops, QC, localization, scheduling, and programming, and every one of those choices rests on a measurement of what the function visibly produces. The bond study is the receipt for what that kind of measurement misses, documented in the one industry that happens to be ours.
That is the shape of the problem. Organizations are bad at identifying which human functions are load-bearing when the value of those functions sits outside the metrics used to evaluate them. This is not merely bad management, because even competent measurement tends toward what is attributable and legible. Which means the instrument you would naturally reach for to answer the question is the instrument least equipped to answer it.
Where the effect lands
The bank teller curve is the version everyone has heard, usually as reassurance. The numbers, from the Boston University economist James Bessen, are real. The number of tellers required to run a branch in the average urban market fell from 20 to 13 between 1988 and 2004, and over the same period urban bank branches grew 43 percent. Cheaper branches meant more branches, more branches meant more tellers, and teller employment held up while the job drifted toward relationship work and sales. This gets cited constantly as proof that automation creates jobs.
It shows something narrower than that, and the precision matters. Automating part of the task changed the economics of the unit the job lived in, and demand for the occupation rose as a second-order effect. That effect was a property of the economics, not of the job. Teller employment has since gone into decline, and the Bureau of Labor Statistics attributes that to branch closures and online banking rather than to the ATM, projecting a further 13 percent drop between 2024 and 2034. Whatever saved the job the first time was a feature of a moment, and a later technology reset the moment.
So the teller story works as a warning rather than a consolation. The first-order substitution is rarely where the consequential effect lives, and the decision you make today tells you very little about where things land two technologies later.
Why the dashboard looks fine
There is a name for the specific failure here, and it comes out of accounting rather than technology. Surrogation is what happens when people mistake the measurable representation of a function for the function itself. The accounting researchers Willie Choi, Gary Hecht, and Bill Tayler demonstrated it experimentally in 2012 and found something that should worry anyone running a metrics-driven organization: the effect intensifies when compensation is tied to the proxy. Pay people on the measure and they stop treating it as a stand-in. It becomes the objective, and the thing it was standing in for stops being anyone’s responsibility.
Run that forward into a substitution decision and you get the trap. You evaluate the function by its proxy, because the proxy is what you have. And the proxy is the countable, attributable part of the work, which is exactly what made it the proxy in the first place. Whatever else the function does that is diffuse, delayed, or landing on someone else’s ledger never made it into the measurement, and not through carelessness. Those effects are hard to attribute and expensive to trace, which is why they sit outside the operating measurement system rather than inside it. Gao and his coauthors measured one of them, and it took a research design and a bond database to do it. So after the substitution the dashboard shows the proxy holding steady, and the dashboard is correct. What it cannot do is establish that the proxy captured the whole function. The substitution can succeed on every measure you chose and still have removed something consequential.
That should land harder now than it used to. Handing a function to an AI system means writing down what you want it to do, and what gets written down is whatever you could already measure. I would treat that as analogy rather than finding, since nobody has documented the failure mode at scale yet. It is the direction I would bet on.
Lepore’s mechanism, and her consolation
The Harvard historian Jill Lepore has a new book out, The Rise and Fall of the Artificial State, and it argues that governments and private corporations are automating politics and public discourse into something she calls the artificial state. Most of the coverage has picked up her secondary argument, that Silicon Valley’s leaders are bad readers of science fiction who mistook cautionary tales for blueprints. It is a good line and it travels well, and Farrah Jarral’s review in the Guardian last week already pushed back on the imprecision underneath it, faulting her for treating very different technologies as a single category.
The pushback is fair, though I would put it differently. Her mechanism is stronger than her frame. The substitution she is describing, where a measurable artificial process takes over a function that used to carry consequences nobody was counting, is the durable part of the book. The bad-readers argument is the part that will get quoted.
Where I would push is on the ending. Her title promises a fall, and the argument leans on the idea that these arrangements have historically opened back up. That idea has a well-known version in Tim Wu’s The Master Switch, which describes a Cycle in which open information industries consolidate into closed ones and then reliably crack back open. The Princeton sociologist Paul Starr took it apart in 2011, arguing that Wu never supplies a mechanism that produces the reopening and that the pattern does not travel well outside the twentieth-century United States. I am not arguing the opposite. I am saying that a historical cycle without a mechanism is not something to plan around.
The operator question
The useful version of this is not that you cannot know. It is that you cannot assume the measurable output exhausts the function, which is a smaller claim and a more demanding one, because it does not resolve into a decision rule.
What follows from it is procedural. Before substituting a mature human process, record what else it touches, notices, escalates, suppresses, or makes possible. The operator who has been catching the same upstream fault for six years is not measured on catching it, and will not appear anywhere in the business case for replacing them. You are not going to capture every hidden dependency that way, and you should not pretend otherwise. You are preserving enough evidence that you have not erased the only people who knew where some of them were.
None of this makes the next substitution decision easier, and it is not supposed to. The reason to keep a record of what a process was doing besides its output is that the record is the only way anyone finds out later what got deleted.