AI Slop Is Breaking the Internet. Can We Save It?

Key takeaways

• Bot traffic recently passed half of all internet activity, Spiro says, and could reach 99%. A study he ran with Stanford and the Internet Archive found 40% of web text was AI-generated by May 2025. Pangram's own analysis puts 41% of long-form LinkedIn posts and 29% of Twitter posts over 250 words in the machine-written column. The endpoint he fears is "dead internet theory," where bots mostly talk to bots and real signal is buried.
• Spiro's original 2022 worry was fake reviews and supercharged disinformation. The real surprise was mundane economics. Text became nearly free to produce, and platforms like LinkedIn reward posting volume with career visibility, so people let ChatGPT do the writing. Derek adds that this erodes trust between people, since polished prose no longer signals real effort or a real mind behind it.
• Pangram is a classifier, not a chatbot. Its team took human writing from before ChatGPT, such as a five-star Denny's review and a 500-word Moby Dick essay, then had models produce matching pieces and trained on the stylistic gap. Spiro compares it to Waymo's pedestrian detector. He claims a false positive rate near one in 10,000, and cites a University of Chicago study that ranked Pangram best by a wide margin with zero false positives across about 2,000 human texts.
• Spiro argues AI's fingerprints have climbed one level at a time—word, then sentence, then paragraph. Early models overused words like "delve" and "intricate." Newer ones lean on negative parallelism, the "not just X but Y" pattern, and sentences built around dashes. Today's models re-summarize themselves compulsively, because reinforcement learning rewards answers that feel impressive, so each sentence competes to restate the point better than the last.
• Both men agree AI can write a convincing Cormac McCarthy sentence but not Blood Meridian. It can fool even MFA-trained readers over a few hundred words. What it cannot do is sustain meaning, character change, and payoff across a novel or a 7,000-word feature, partly because so little of its training data runs that long. Derek recalls an Atlantic editor telling him to stop signposting and let readers get lost before the reveal.
• Spiro frames Pangram as a counterweight to the money pushing outlets toward high-volume AI output and fewer staff writers. He names three risks to his company: the writing gap closing, readers ceasing to care, and big labs building their own detectors under AI's "bitter lesson" that scale beats niche expertise. His bet is that an independent referee will be trusted more than a lab grading its own homework. He also expects Anthropic's new watermarking to over-credit Claude on documents with mixed human and machine authorship.

Notable moments

• Derek's cold open: people dislike AI because of its first letter, since it is so often used to make fake things, from fake LinkedIn monologues to Facebook's "shrimp Jesus" and Google's glue-topped pizza recipes.
• Live fact-check: Spiro corrects the New York Times' rendering of Pangram's Twitter data, giving 29% for posts over 250 words and about 9% for shorter ones.
• "We're trying to save the internet. We're trying to save the written word and communication."
• The publishing reckoning: Hachette pulled a horror novel flagged as substantially AI-generated, the Times fired a book critic, and The Atlantic reported a Modern Love column that seemed half machine-written.
• The Guardian dispute: after the paper defended a sports columnist, Spiro ran about 900 of that writer's articles through Pangram and published a time series showing rising AI reliance from a visible start date.
• Baseball analogy: a machine could throw far faster than 105 mph, but nobody wants to watch a robot pitch for the Brewers or Pirates; the human record is the only version with value.
• Photography and painting: painting survived the camera by moving away from realism, and Derek notes his favorite painters all worked between roughly 1870 and 1925.

Time saved: 46 minutes.