A tentative taxonomy of AI slop

Eight years ago, the term “fake news” burst into public discourse. But it was frequently misused. 

Fake news was supposed to describe a very specific phenomenon that my colleague Craig did much to uncover: knowingly false and sensationalist headlines posing as news, posted typically on Facebook to make an easy profit. It quickly became a catchall for many other things. Most notably, Donald Trump used it as a cudgel to attack mainstream media. 

The ensuing mess led communications scholar Claire Wardle to write a foundational framework describing the full range of content that was being mislabeled as “fake news”.

I think that we may soon approach a similar moment for “AI slop.”

The term describes visibly AI-generated content that is banal and low-quality to the point of being grotesque. And the stuff is everywhere.

Earlier this year, I reported on AI motivational slop of Donald Trump and Elon Musk that had clocked 700 million views on TikTok. In a 5-minute session on the app on Saturday, I was served the following specimens on my For You Page: Presidents and their country guardians (34 million views), Whale-watching nightmare in Alaska (14.9M), If this ladder breaks… (3.3M) and Joseph Boakai hits back at Trump (850K).

I love AI slop as a term. This type of content is indeed an unappetizing gruel being forcefully fed to us. “Slop” also evokes spam, down to the fact that both are American terms for unappetizing or unhealthy processed foods.

But AI slop is broader and more pervasive than spam. Some slop is essentially misinformation; but most of it is not, because it typically doesn’t present itself as genuine. These distinctions matter, so I put together a two-by-two matrix for discussion by regulators, researchers, and Trust & Safety workers navigating AI slop-infested online spaces. This taxonomy separates AI slop by purpose and form and should make targeting component parts easier to target with appropriate countermeasures. I lean primarily on video examples to explain the distinctions in my taxonomy but it applies equally to text and audio.

This is very much a first draft: I plan to update it based on developments in the field and feedback from slop savants. Please reach out with your thoughts and comments at [email protected].