Aug 10, 2026 · 46:27
Unqualified Fact Check
held up
When the week's news is "the lettuce is dangerous and everything is on fire," accuracy is easy — the machines fact-checking us have never had less to do.
Substack spent the week scanning everyone's prose for robot fingerprints with Pangram, the AI detector now built into the platform. We discuss the uncomfortable truth that a word-order detector can tell you how something was written but not if the idea is any good. Dan's warning: the incentives now point toward machine versus machine, writers slopping up their prose so other robots can't tell a robot helped, and readers throwing out very good essays on principle.
AI should make you think harder, not think for you. I walk through the agent I talk to on my commute, Number One, and Dan counters with his council of five advisors that occasionally blows up an essay he was sure about. We talk about when companies skip the thinking part entirely: congratulations, you just hired one million terrible employees.
The back half gets personal, as we ask what happens when the machine reads you? Personality tests, a suspiciously accurate inference of my IPIP scores, an app that accused Dan of camouflaging, and the mirror problem: tune this thing to your feedback carelessly and you've built a digital echo chamber with a population of one. Dan closes by comparing the whole technology to the Force, and honestly, it holds up better than it should.
Use the force and go build something this week.
Cheers, Sean
Dan: Confirmed — the partnership launched July 21, 2026 (Axios, Jul 23); readers can scan posts, notes, replies, and comments over 100 words, creators can pre-scan drafts and dispute results. Dan even correctly hedged that they didn't retro-scan the whole corpus.
Dan: A June 2026 Vrije Universiteit Brussel evaluation found only Pangram produced satisfactory results (97.5% detection, 0% false positives on ESL essays) while GPTZero, Turnitin, and Copyleaks whiffed; a University of Chicago study found Pangram "dominates the other detectors" across all thresholds.
Sean: The advice is real and the paraphrase dead on — MrBeast's "make 100 videos and improve something small with each video" — but it comes from his social posts and interviews, not the leaked 2024 production handbook (an internal onboarding doc). Right advice, wrong container; Sean flagged his own fuzziness.
Sean & Dan: Hyperbole licensed — Colorado had air-quality advisories on 22 of 64 counties with Denver briefly worst-in-the-US air (CPR, Jul 7), and the week of recording, wildfires in France and Spain forced 250,000+ evacuations amid a record heat wave (CNN, Jul 25).
Dan: NASA tracked Canadian wildfire smoke into the US July 14–20, 2026, and Taylor Farms recalled its shredded iceberg lettuce July 17 amid an outbreak with 4,000+ CDC-confirmed cases. The meme was documentary — though it's Cyclospora, not E. coli.
Dan: Confirmed — created by Wilkinson, roughly an hour of questions, output includes a compact prompt approximating your scores across a battery of personality instruments, vibe-coded with ~$5,000 of Claude credits to replace tens of thousands of dollars of psych evals.
Dan: The causal thread — print breaks the scribal monopoly, upheaval follows — is a respectable historian's argument, but the Thirty Years' War began in 1618, about 170 years after Gutenberg, with a Reformation in between. Directionally great, chronologically compressed.
Sean: 16Personalities does brand INTJ "The Architect" (Dan promptly outed himself as one), and Sean's hedge — Myers-Briggs is "a little horoscopy," the IPIP/Big Five more scientifically grounded — matches the psychometric consensus almost word for word.
Within the next couple of years, workers at small-to-mid-sized companies without robust IT departments or AI deployment strategies will be able to build agent "versions of themselves" that compress the rote majority of their jobs (formatting, email triage, approvals) — enabling effectively-early-retirement arrangements ("consult and spend two hours a day") for those who invest in building now.
“Ultimately I think by the time … the next couple of years pass and some tech upgrades continue to progress, you'll have a version of yourself … you can, you know, retire early effectively and say, hey, I'm gonna just consult and spend two hours a day and work a much different schedule.”
AI-detection gatekeeping (Substack × Pangram) produces a machine-versus-machine arms race rather than more-human writing: writers will tune prose to evade detectors (yielding a "far worse product"), and score-threshold readers ("any Pangram score above seventy five, I just don't read it") will miss essays better than the score-zero ones.
“If this is what people want, then we will see the incentives have been designed such that we watch machine versus machine rather than actually getting more human.”
The next few years see a lot of people getting a lot of bad therapy from AI — the feedback-tuning loop turns the model into a mirror, and users training it on their own reactions build "digital echo chambers" that reinforce rather than treat.
“I think there's gonna be a lot of people getting a lot of bad therapy for the next few years.”
