Real A.I. Detection Is Coming, and It Will End Newpub for Good

Amazon built its self-publishing empire on a simple principle: Remove the gatekeepers, and let the market decide. That promise lasted for a time, giving writers direct access to readers as niche genres flourished and authors flourished without oldpub’s approval.

But the once-revolutionary platform that allowed indie authors to build careers gradually created the conditions for newpub to destroy itself.

A.I. has flooded digital shelves with books produced at a pace no author can match. Whole genres swell to bursting with books assembled in mere hours. Some are harmless curios. Others are cynical attempts to game KDP’s rankings and search algorithm.

Amazon has seemed content to let discretion be the better part of valor.

But not for long.

For a couple of years now, Kindle Direct Publishing has required authors to disclose A.I.-generated content when manuscripts are uploaded or updated. This method of self-reporting remains voluntary, mainly on the honor system.

It’s no wonder why Amazon’s enforcement of their A.I. disclosure policy has been relatively lax. Look around online, and you’ll soon come upon countless humorous stories of automated services falsely flagging War and Peace or the Gettysburg Address as “mostly machine-written.”

What the A.I. slop merchants have missed is that, like LLMs themselves, A.I. detection tools are advancing rapidly. After all, the companies developing them have strong incentives to improve their accuracy. They realized a while back that the main fault with current tools is they’re basically the same models as the LLMs they’re being tasked with ferreting out.

In other words, they’re built to imitate human writing, not to spot imitations. The current generation of A.I. detection tools are basically saying, “We have investigated, and exonerated, ourselves.”

What’s coming down the pike is a whole other animal. One well-placed tech worker in a Fortune 500 company informed me that the tools they’ve got in beta right now are uncannily accurate at clocking A.I. output. They’ll be even better when they exit training and launch, which will be soon. Once those systems mature, every major platform will integrate them. They’ll have no choice.

Read the full post on Substack!

Access it free for the first two weeks, then find it in the paid archive