The Endgame for AI-generated Writing

“What will the internet look like when it is populated to a greater extent by soulless material devoid of any real purpose or appeal?” wrote Atlantic senior technology editor Damon Beres this week.

Beres’s concern is that media corporations less scrupulous than his own are beginning to pollute the internet with cheap, machine-synthesized content. He argues we’re headed towards a dystopian future that looks an awful lot like the dead-internet conspiracy theory, which posits that most of the people and content we see in our feeds are fakes—created by an evil AI to manipulate us into buying stuff we don’t need and consuming information that isn’t true so that advertisers can reach us.

On what fact pattern is Beres basing his projections? His article names two recent developments.

The first is BuzzFeed’s announcement that it will use AI to create personalized quizzes (e.g., “Answer 7 Simple Questions and AI Will Write a Song About Your Ideal Soulmate”).

The second is CNET’s experiment using AI to generate SEO-optimized explainer articles, such as “What is an Annual Percentage Yield?,” some of which apparently contained significant inaccuracies and plagiarism.

One writer employed by CNET’s parent company, Red Ventures, wrote an anonymous op-ed that reveals just how deep anxiety over AI-generated text runs:

I wonder about what the future will be like for my children. I wonder if they’ll have the same dreams of being a writer like I did when I was young. I wonder if that job will even be there when they grow up. Twenty years from now, will they cut their teeth on freelancing, learning and developing their style and getting their beat? Or will it all be dried up? Will the door be closed forever, the ladder pulled up behind us, the last writers, our words used to feed the ever-starving algorithm?

Later in the op-ed, the anonymous writer predicts the same dystopian future that Beres did:

Google’s going to be clogged with AI-generated content of dubious accuracy. Will it turn into an endless prism of echoes, as the algorithm scrapes articles from other algorithm-generated articles, over and over again? Will the cultural vernacular be changed when the majority of content we read is filled with the syntax and semantics of a robot?

Is this the endgame for AI-generated content? Has the last writer already been born? Will future generations be forced to wander aimlessly in a junkyard of regurgitated content, desperately searching for morsels of meaning?

Or will the future play out differently? Is there reason for hope?

I believe so. Here’s why.

AI-generated text will be easy to detect

People want to know the source of the information they read. Therefore, watermarking tech* will become a standard feature of LLMs provided by large AI companies that want to keep their reputations intact. Only those on the fringes will attempt to evade this—either as an LLM creator or user. Even the CNETs of the world have realized transparency is in their best interest.

* Side note: the way the watermarking tech works is incredibly cool. LLMs generate one word (technically one “token,” which can be part or all of a word) at a time. Before the algorithm decides what the next token should be, it runs some math on the previous token to deterministically generate a “greenlist” of allowed possible next tokens and a “redlist” of tokens it should not say next. For instance, after the word “for” it might be allowed to say “instance” but not “example.” The more red-listed tokens a text contains, the less likely it is to be generated by AI. You can remove the watermark by intentionally replacing green-listed words with red-listed alternatives, but because the lists are dependent on the words that came before, this is complex and requires substantial revision to the original text.

So what about the marginal players that try to evade detection? I doubt their prowess will be able to match the much bigger players that will invest in AI content detection classifiers, like the one OpenAI released just yesterday. These models aren’t good enough yet, but over time it seems likely that they will be able to tell with reasonable accuracy whether a text was AI-generated or not—even when there is no watermark.

Today we live in the Wild West of AI, but once you can discern AI-generated text 80–90% of the time, anyone who wants to avoid it will be able to do so with a browser extension. I wouldn’t be surprised if Apple builds something like this into their OS. It’ll be like digital privacy—if you care, you can opt out by adopting a few simple tools like ad blockers and VPNs.

The question is: will people care to intentionally evade AI-generated content? Will they have to?