The Ghost in the Manuscript: Why AI Detection Is Publishing's Least Interesting Problem

Here is what the publishing industry has apparently decided: the best way to fight machine-written prose is with machine-reading tools. The logic is airtight. Or it would be, if those tools worked. As critic Maris Kreizman noted this week in a widely shared essay on Literary Hub, AI detection software is "notoriously flawed at best, ruinous to new authors at worst" — a remarkable description of a solution to a problem that publishers helped create.

The problem Kreizman actually identifies is not technology. It is time. Editors are overworked. Houses have shed staff for years in the name of efficiency. The people most qualified to catch suspicious prose — editors who have read thousands of manuscripts, who know what human uncertainty looks like on the page — are the same people being stretched thinner each season. When a manuscript that reads too smoothly lands on an overloaded desk, the red flags are easy to miss. Or easy to ignore.

What Kreizman is describing, in more careful language, is a labor crisis dressed as a technology crisis. The publishing industry has spent two decades automating what it could, outsourcing what it could, cutting what it could, and is now surprised to find that its quality filters are porous. The solution — as she argues convincingly — is not a better algorithm. It is a better-staffed editing floor. Give people time to actually read. Give editors the conditions under which suspicious writing becomes obvious.

There is an irony here worth noting: the same corporate logic that accelerated the AI threat — more output, lower costs, faster timelines — is the same logic that makes the solution feel impractical. Investing in human expertise is expensive. Slowing down is expensive. Treating a manuscript as something that deserves unhurried attention is, in contemporary publishing arithmetic, a luxury.

Meanwhile, readers are adjusting. A skepticism has crept in — the slight hesitation before picking up an unfamiliar author, the wondering whether the rhythm of the prose is real or statistical. Trust, once eroded, takes considerably longer to restore than a quarterly earnings call. What publishers are risking is not just the odd suspicious title. They are risking the basic contract between writer and reader that the industry exists to facilitate. The good news, if there is any, is that a labor argument is winnable in a way that an arms race with language models is not. Whether anyone with the authority to change things is listening is another matter entirely.