“Before a person asks a machine to write for them, they should know what it feels like to write from the depths of their own uncertainty and identity. Before a person lets a machine imitate their voice, they should have endured the long, chaotic, and often humiliating process of forging one.”
— Debbie Milman, via Umanz (August 2026)“Everything is a song to me.”
— Dolly Parton (1946–2026), via Austin Kleon
People: Socializing less. Young workers in AI-exposed occupations (ages 22–25) are now 19% below their expected employment trend. Dolly Parton died; also Yayoi Kusama died the same week Tim Curry (RIP RHPS!). Hard vs soft fascination.
AI strat: Zuckerberg’s AI manifesto. AI and writing (WITI, Noah Brier): a high school English teacher worried that AI would make the “write longs” assignment pointless. Design AI that strengthens reasoning.
Security: Russian backdoor in speed cameras. ICE ankle monitors.
Geopolitics: AI sovereignty paradox. Recursive self-improvement has ceilings.
Society: AI rejection letters. Magnifica Humanitas. AI Resist List. Philippine digital workers coalition (2025). I’ve been Substacked (workfutures): Stowe Boyd discovered mid-read that a newsletter he was enjoying was AI-generated. Personal nostalgia vs political nostalgia.
Tech: Token asymmetry. Tree canopy mapping.
Foresight: The Clarke inversion — any sufficiently advanced technology becomes indistinguishable from Nature. Boussole Anti-chaos: a one-page vision-and-strategy visual deployed across 450 people in a single session; the one-pager as an artifact of alignment.
AI: Qwen 3 doing frontier work. Kimi K3: the team’s own account of building a frontier model under export controls. RSI limits.
![Chicago, Ill., 12th St. Bascule Bridge] — PICRYL - Public Domain Media Search Engine](https://substackcdn.com/image/fetch/$s_!xbI8!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fef0ff8ac-11c2-43b1-92a6-d0d7f70735ab_1024x811.jpeg "Chicago, Ill., 12th St. Bascule Bridge] — PICRYL - Public Domain Media Search Engine")
The Voice
Priya had been subscribing to Nour’s newsletter for three years. It was one of about forty she’d kept from the cull.
The Tuesday edition arrived at 7:14 AM. Priya read it on the train, standing, one hand on the overhead bar. The first paragraph was about a drawbridge on the Chicago River — the kind you wait for while a barge passes, the feeling of impatience becoming something else. She thought: this is good. Then she thought: this is almost too good.
At the office she ran it through the content analysis tool. 71% confidence: machine-generated.1
She emailed Nour through the newsletter footer. Nour replied in eleven minutes — faster than she had ever replied before. Yes, she’d been using assistance for about six months. The ideas were still hers. The structure was still hers. Was it really so different from using Grammarly?
“It’s not so different,” Garrett said, reading over her shoulder. Garrett had opinions about this. He’d run his last fourteen client proposals through the same pipeline. “The ideas are what matter.”
Priya thought about the Milman line she’d screenshotted months ago and never deleted: Before a person asks a machine to write for them, they should know what it feels like to write from the depths of their own uncertainty and identity.2 She hadn’t thought about it in a while.
“Did you read the Pope?” she said.
Garrett had not read the Pope.
“He published an encyclical. In May.” The Vatican had been unusually fast — the text argued that the authenticity of the human voice was not reducible to any simulation, however accurate. Magnifica Humanitas. Six hundred footnotes. Already cited in three legal briefs.3
“The Pope has a ghostwriter,” Garrett said.
This was technically true.
Priya looked at the Nour newsletter again. The drawbridge paragraph was still good. The barge of crushed cars, the mood shift, the way it turned on a single waiting figure — these were not things a machine assembled by accident. Maybe the ideas were still Nour’s.
She opened her own archive. Three years of performance reviews. The oldest one, from 2024, when she’d been working eighty-hour weeks and writing everything herself — a long paragraph about her team’s approach to uncertainty. She ran it through the tool.
68% confidence: machine-generated.
She had not used AI in 2024.
Garrett looked at the result and then looked away in the manner of someone who had already had this conversation with himself. “The tool’s not perfect,” he said.
“No,” said Priya.
She archived Nour’s newsletter in a subfolder she’d been maintaining for two years called Keep This. Then she went to get coffee. There was a new machine in the kitchen that made it automatically when you walked in. It was quite good.
1 Anthropic’s watermarking system, as of August 2026, reports a confidence score rather than a binary flag. A score above 0.65 is classified as “likely AI-assisted.” The company notes this threshold produces roughly 12% false positives on human writing from non-native English speakers. Nour was not a native English speaker. Priya did not know that the threshold existed.
2 Debbie Milman, quoted in the French newsletter Umanz (August 2026). The full sentence continues: “Before a person lets a machine imitate their voice, they should have endured the long, chaotic, and often humiliating process of forging one.” The word forging is doing a lot of work here — it implies heat and resistance, not just formation.
3 Pope Leo XIV, Magnifica Humanitas (May 15, 2026). The Vatican’s encyclical is available in nine languages. The 600-footnote figure was reported by La Croix. One of the three legal briefs that cited it was filed in a wrongful termination case in which the plaintiff argued that an AI-generated performance review had misrepresented their work. The document is 47,000 words. Priya had read the executive summary and three footnotes.