Howdy, folks.
A link landed in my orbit last week: a blog post from an herbalism school called Rowan + Sage, written by Sarah Corbett, arguing that herbalists need to categorically refuse AI. Not "use it carefully." Not "watch your sourcing." Refuse it. All of it, forever, on principle.
My first instinct — and I'd bet yours too, if you spend any time around tech-critical writing — was to brace for the thing this genre usually is. Vibes wearing footnotes. A real feeling reaching for citations that can't carry the fucking weight put on them.
So I did what I do. I followed every link. All of them. If you have never done this to a piece you agree with, I recommend it and I do not recommend it.
Credit where it's due
Most of it holds up.
Start with the headline stat: 82% of herbal remedy books on Amazon likely AI-generated. Originality.ai scanned 558 titles published in Amazon's herbal-remedies subcategory between January and September of 2025, and flagged 82% of them.
Now ask who ran the scan. Originality.ai, which sells AI detection.
Their tool flagged the samples with what got reported as "100% confidence." But their own explainer says the output isn't confidence at all. It's a likelihood score, printed on screen as "100% Likely AI." Same page, their words: "Even our own AI detector is not perfect, and can produce false positives."
Watch what happened in transit. A likelihood became a confidence. That happened one hop upstream of Corbett, before anybody in herbalism laid a finger on it.
So 82% is a vendor's estimate with real slack in it, and the load-bearing word in that headline is likely. Don't let me skate past it any more than I'm letting her.
What doesn't wobble is the reporting underneath. The Guardian went looking for the people behind the bestsellers and came back empty-handed.
Natural Healing Handbook is a number one bestseller in three Amazon subcategories. Its author is Luna Filby, a thirty-five-year-old herbalist from Byron Bay, founder of My Harmony Herb. Sarah Wynn, of Wildcraft Journal, calls the book "a resource and an inspiration."
None of those four exist. Not the herbalist. Not her brand. Not the reviewer. Not her journal. No footprint anywhere beyond the page you'd buy it from.
At least twenty-nine of the flagged books send you to Barbara O'Neill and Alfredo Bowman, who both promoted unproven cures for cancer. That's what's on the other end of the listing you're scrolling.
That isn't a detector's guess. That's shoe leather, and it's the real one. If you've ever bought a herbal off that shelf, you played those odds whether you knew it or not.
The environmental numbers hold up too, and I mean that specifically, because I've spent five months building these receipts myself. Data-center electricity demand really could reach 945 terawatt-hours by 2030 — the same IEA figure sitting in my own notes. AI-related water consumption really could equal the basic annual domestic needs of 1.3 billion people by the end of the decade.
You want to know how rare that is? I checked every viral AI-environment claim I could find back in April. Most were wrong by one to three orders of magnitude. These two weren't. They're the real ones.
And there's a passage in her piece better than anything I'm about to say about it: "No one really wants your AI-generated essays and monographs. [...] We want your stillness and your slowness. We want your messy words, your imperfect art…we want to witness you as yourself."
That's not a citation. It's also not wrong. It's the ache under the whole piece, real regardless of what the footnotes do. If you make things, you know the feeling she's naming.
The tell
Here's where it stops holding.
Corbett cites a study claiming AI use is "directly linked to cognitive decline and the erosion of our critical thinking skills." So I read it. The paper is a 666-person self-report survey. Not clinical. Not longitudinal. It found a correlation between self-reported AI reliance and self-reported critical-thinking scores. It has since been through a published correction.
It's also one of the most argued-over papers in this discourse, disputed for the precise reason it just got misused. Everyone who cites it swaps "correlation, self-reported, in this specific sample" for "directly linked." The second phrase sounds like science. The first one sounds like what it actually is.
She cites a Brookings piece on AI labor in the Global South as proof that models "are built and maintained through the exploitation and abuse of people in the Global South." The piece documents real conditions. Twenty-hour days. Subcontracting that hides accountability. Real psychological harm.
Then it turns, for its last third, to the people organizing against exactly that — the African Content Moderators Union, the Data Labelers Association, Turkopticon — and closes by calling for binding directives on the rights of data laborers.
It's a reform piece. Its argument is that this is fixable and the people inside it are already fighting.
She quoted the wound and skipped the part where the patient gets up.
Neither of these is fabrication. Nobody's lying.
But look at the job these two are doing that the honest ones can't. The solid stuff — the missing authors, the water, the energy — gets you to "be discerning, demand sourcing, watch the platforms." It does not get you to "refuse the tool, categorically, forever."
For that, you need a citation that sounds more damning than its source is. Which is a bullshit way to build an argument. Also, apparently, an easy one, if you need it badly enough.
Meanwhile, in a Mississippi courtroom
Twenty-five minutes from where I live, xAI built a data center called Colossus in South Memphis, then crossed into Southaven, Mississippi, and installed twenty-seven gas turbines without an air permit. There are sixty-nine out there now. Across the state line sits Boxtown, carrying a cumulative cancer risk four times the national average.
The NAACP sued. I wrote about it in August: "data center" was doing camouflage work, and arguing about the category let the specific violator hide inside a class of building that mostly just files its paperwork.
The story kept moving, the way these always do.
The Department of Justice tried to intervene in the NAACP's suit. Not to help it — to help dismiss it. The district court effectively said no, by doing nothing. A "constructive denial," which is a genuinely useful phrase for a company getting told no without anyone having to put it in writing.
So DOJ appealed to the Fifth Circuit. Then, days later, asked the district judge to freeze the entire case while that appeal plays out. Denied entry through the front door, they're trying to stop the clock from the parking lot.
The hearing that actually matters — the one that decides whether the turbines get shut off now, not eventually — was postponed in August and, as of this week, still has no new date.
And the Memphis moratorium I told you to watch in August finally moved. Second reading passed. The final vote, originally set for mid-September, got pushed to October 6th. And the ordinance carries an exemption — for "existing lawfully operating data centers that do not expand their building footprint, electrical demand, utility capacity, or operational intensity."
Read that list again. Electrical demand. Operational intensity.
Now hold it against the Colossus Gas Plant, which is trading sixty-nine temporary turbines for a permanent 1.2-gigawatt power plant. Eleven are already dark. The rest go by mid-2027. Nobody is making them do it; it's their own timeline. Does that sound like not expanding to you?
Nobody has tested it. The vote hasn't happened, and I don't know which way it goes.
None of that is a tidy update. This is what a real fight looks like from the inside. Procedural. Slow. Contested at every step by people with more lawyers than you have. Still open five months later, and nobody in it gets to feel finished.
Now put that beside refusing a chatbot, and tell me which one you'd rather be asked to do.
The actual crisis in herbalism
I'm not an herbalist. I don't have a dog in whether Corbett's students learn wildcrafting from a real teacher or a search bar.
What I do have is a front-row seat to the identical fight, one lane over. Tarot, astrology, the whole apprenticeship-and-lineage economy that trained me. Same platform, same slop, same instinct to answer it with a pledge instead of a receipt.
Publishing a deck used to cost real money. Benebell Wen put the numbers in public back in 2018: about twenty-six thousand dollars for a thousand-unit run, and five hundred hours, before a single card gets drawn. Nobody designed that as a gate, but it sorts for conviction anyway. It's why the tarot shelf used to be mostly people who meant it.
That filter is gone. Print-on-demand means no run, no inventory, no upfront. Generated art means no illustrator. The whole pipeline costs an afternoon.
So which way did the tarot world jump? Mostly pledges. Badges. Human-made banners on shop pages.
Not everyone. A friend of mine makes decks, and sells other people's through her shop. She won't stock AI decks and won't work with AI creators, and when she learned that one of the decks she was carrying was AI-produced, she took it off the shelf.
That one costs something. She had to find out, which nobody made her do, and then pull a product she was already selling. And it does what a banner on a homepage cannot: one less unattributed deck reaches a buyer who had no way to tell.
She runs a shelf, and she used it.
That's the whole difference, and it isn't about how strongly anybody feels. It's whether the refusal is attached to anything. Hers is. Corbett doesn't run Amazon's shelf, and no amount of not opening ChatGPT puts her hand on it.
So the actual crisis in herbalism isn't AI. It's Amazon, and it's the shelf you buy from.
A platform with no mechanism to tell a book written by someone who apprenticed for a decade from one written by a model prompted forty times in an afternoon. It rewards both identically. Reviews come in, the algorithm doesn't notice, everybody gets paid.
That platform predates ChatGPT. It rewarded volume over relationship before AI made volume free. AI didn't create the incentive — it removed the last bottleneck, the part where faking an herbal guide used to at least cost you a weekend.
"Refuse AI categorically" is a purity test. Ask what passing it requires of you. Nothing. Never open the tool. Never touch the platform's incentives. Never say a word about the specific sellers with no footprint beyond a listing page, or the category algorithm that put an unverifiable author at the top of "Herbal Remedies."
You get to feel clean. Nothing about the mechanism moves.
"Name the fake books, name the sellers, demand Amazon label AI-generated health content, point your students at teachers you can actually verify a lineage with" is a fight.
People are already in it. Dan Conway, who runs the UK's Publishers Association, says any fully AI-written book should be labelled as such, and the slop "must be removed as a matter of urgency."
Amazon's answer? It has content guidelines. It has "proactive and reactive methods" to enforce them. That's a company with a number one bestseller by an author who does not exist, telling you it has methods.
You can lose this fight. Amazon has no incentive to fix it and every incentive to keep the listings coming, which is the whole goddamn point. That's why it's the version that matters. Something is at stake, instead of a feeling dressed up like one.
The word doing the camouflage work
Three pieces in now, and it's the same mechanism every time, wearing a different coat. In April it was the numbers. In August it was a noun. This time it's a pledge.
"Data centers" is a big enough category to hide Colossus inside every boring, permitted building that never hurt anybody. "AI" is a big enough category to hide "Amazon has no mechanism for craft or accountability" inside a fight about whether you, personally, ever open ChatGPT.
Vagueness doesn't fail by accident. The people reaching for it are plenty smart enough to be specific. Specific is where you can lose. Vague is where you always, technically, win.
Which one are your pledges?
A category can't beat you. A permitting hearing can. A vibe can't beat you. A verified teacher, named in public, standing behind their own work, can.
Corbett is right that herbalism is losing something real. If you have ever learned anything from a person instead of a search bar, you know what she's grieving.
I just watched her reach past the winnable fight — name the books, name the sellers, name the platform — for the one that costs her a sentence and changes nothing.
I know what you're going to say: easy for the guy who uses the tools every day to tell someone else their refusal doesn't count.
Fair. It's the same move anyone makes when they want permission badly enough. But run it the other way for a second. What does my refusal buy you? If I close the laptop tonight and never open a model again, how many fake herbals come off Amazon tomorrow?
None. Not one. My hands get clean and the shelf doesn't move an inch. That's not an argument for using the tools. It's an argument that my personal consumption was never the lever, and neither is hers.
The hearing in Mississippi still doesn't have a date. The moratorium vote is October 6th. Go find out what happens. That's the fight. Refusing a chatbot was never it.
That's it for this week.
Stay feral, folks.