Joe Weisenthal Says AI Writing’s Comma Tell Is Already Slipping

Joe Weisenthal spent an April podcast episode explaining how to catch AI writing by its rhythm and its punctuation. By July he was posting that the trick had stopped working, even for him.

The Bloomberg executive editor for news and co-host of the Odd Lots podcast has spent 2026 tracking a strange corner of the AI story: what happens when machines write so much that spotting them becomes its own skill, and then that skill stops being reliable. His own timeline shows the shift in real time, and this month he pointed followers to something else worth noticing too, a deleted post from Airbnb co-founder and chief executive Brian Chesky about blockchain that vanished from Chesky’s own feed but not from anyone else’s screenshots.

A Tweet That Undercut His Own Advice

On April 2, 2026, Weisenthal and Odd Lots co-host Tracy Alloway published an episode built around a single question, how do you actually tell if a piece of writing was generated by AI. Their guest was Max Spero, chief executive of Pangram Labs, a company whose software scores whether text was written by a machine. The episode description leaned on a specific, almost backhanded observation from the hosts: AI is already better than most people at placing a comma, which is precisely why so much of it reads as suspiciously clean.

Three months later, Weisenthal changed his mind in public. On July 10 he wrote on X, the platform formerly called Twitter, that his old confidence was gone.

I used to cynically think ‘AI writing sucks, but it’s better than 90% of people at it.’ But I’m less sure. I think it’s gotten worse.

He went further in the same post, writing that recent Anthropic model outputs were so larded with Claude-isms that he was getting responses he called flatly unintelligible. A separate reply in the thread, from an AI researcher who posts under the handle roon, floated a different theory: that model writing has not actually declined, readers have simply been exposed to its mannerisms so often that the tics now grate.

That same week Weisenthal used his feed to flag something unrelated to language models at all. He drew attention to a post from Chesky touching on blockchain that Chesky had since deleted, the kind of moment that used to disappear from the public record and now mostly doesn’t, because someone always has a screenshot.

How Do You Actually Spot AI Writing?

Detection still leans on a handful of measurable signals: unnaturally even sentence length, a smoothed-out vocabulary with few surprising word choices, and punctuation that lands with mechanical regularity instead of the bursts and gaps typical of a human draft. None of these signals is proof on its own, and researchers who study the field describe them as probabilities, not verdicts.

The earliest widely cited tool for this, called GLTR, was built by researchers at Harvard and IBM who found that giving readers a visual guide to word-by-word predictability raised human detection of generated text from 54% to 72%, without any additional training. That gap between chance and near-certainty is the whole game: detection has always been a probability exercise dressed up as a checklist.

Writers, editors and researchers watching the space point to a similar cluster of habits when they suspect a machine wrote something.

• Signposting openers such as announcing a breakdown before delivering one, a structure that mimics an overeager instructor more than a person telling a story.
• Flattened burstiness, meaning sentence lengths that barely vary, where human drafts swing between a clipped fragment and a much longer, comma-heavy build.
• Stacked triads, three parallel phrases or adjectives used for rhythm rather than because three distinct facts existed, a pattern Wikipedia’s own editing guidance now explicitly flags in suspected machine-written entries.
• Suspiciously uniform punctuation, where commas and periods appear at almost predictable intervals instead of clustering the way a person’s actual thinking does.

None of that is new to linguists. What is new is how fast the target keeps moving underneath it.

Pangram Labs, Havelock.AI and the Detection Business

Spero built Pangram Labs specifically to make a business out of the problem Weisenthal was describing. The Odd Lots episode framed the company’s core challenge honestly: the advanced techniques it uses still carry real risk of false positives and false negatives, an admission that would be unusual marketing copy from most vendors.

Weisenthal has his own, much smaller version of the same instinct. He built a free web tool called Havelock.AI, now at version 0.4.1, that scores how oral versus literate a passage of text reads, drawing on the corpus-linguistics work of scholars Walter Ong and Eric Havelock. He built it through what he calls vibe coding, iterating conversationally with a language model rather than writing the code by hand, and he is upfront that it is a science project rather than a finished product.

The irony sits close to the surface: a tool meant to help sort human writing from machine writing was itself assembled largely by prompting a machine.

The Tools Keep Flagging the Wrong People

The harder problem is not what detectors catch. It is who they catch by mistake. Research from Pindrop and the Authors Guild, an advocacy group for professional writers, found that skilled human writers get systematically misidentified as AI, and that the pattern is not something better engineering will fix.

The Authors Guild called it a troubling paradox: the more refined and controlled a writer’s style becomes, the more it can resemble the polished, low-surprise output detectors are built to catch. Separate research cited in that same coverage used formal probability theory to argue that any one-shot, text-only detector with real detection power will inevitably misidentify some writers whose style overlaps statistically with machine output, a structural limit rather than a bug to patch.

Regulators are leaning on the first tool in that table more than the last three. The EU AI Act’s Article 50 transparency rules take effect on August 2, 2026, and they lean on AI companies labeling their own output rather than on any outside detector catching what slips through. That is a tacit admission that third-party detection, the very category Pangram and Havelock.AI both sit in, is not considered reliable enough to build law on.

The Same Editor Is Also Watching Wealth and Soccer Clubs

The writing debate is one thread in a broader habit. Weisenthal has separately reported that the world’s 500 richest people added $366 billion to their combined fortunes in a single day, the largest one-day gain on record for that group. He has also argued that European soccer clubs could become more valuable as artificial general intelligence advances, on the theory that assets with a genuine, irreproducible history turn scarcer, not more common, once a machine can generate almost everything else.

That same week he was tracing leverage-driven gains in chip stocks and the rising cost of equity financing, a rally that has kept Nvidia holding the world’s top valuation even as its stock trails the broader chip boom. Wealth concentration, soccer club scarcity and chip-stock leverage look like three unrelated stories. Read together with his comments on AI writing, they trace one instinct: figuring out what stays genuinely scarce, and what quietly loses its value, once machines can copy almost anything.

A comma used to be a small, reliable signal of which side of that line a piece of writing sat on. Weisenthal’s own week suggests that signal is fading faster than the replacement for it is arriving.

Frequently Asked Questions

Does Joe Weisenthal think AI is bad for writers?

Not straightforwardly. In a separate interview on Semafor’s Compound Interest show, Weisenthal said a lot of the criticisms people level at the emerging AI world really just feel like descriptions of the past, and he has said he tries not to be too rigid in his expectations about where the technology goes next.

What is Pangram Labs?

Pangram Labs is a company led by chief executive Max Spero that builds software to score whether a piece of text was written by AI. Its own public description of the work, laid out on Odd Lots, acknowledges real risk of both false positives and false negatives in that scoring.

How accurate are AI writing detectors in 2026?

Independent evaluations put accuracy anywhere from roughly 60% to 90%, depending on the specific tool and the type of text being scored, with shorter or heavily edited passages proving hardest to judge reliably.

What does the EU AI Act require for AI-generated text?

Starting August 2, 2026, Article 50 requires providers of generative AI systems to mark their outputs in machine-readable form, and certain AI-generated text published on matters of public interest must be visibly labeled to readers.

What is Havelock.AI?

Havelock.AI is a free tool Weisenthal built himself that scores how oral versus literate a passage of text reads, based on corpus-linguistics research rather than AI-detection science. He describes it explicitly as a science project and cautions against using it for anything mission-critical.