AI Left Fingerprints. Then Picked A Side. Then Hunted For Rare Books.

Inside: Steal this elite AI prompt for a floriferous fall garden that blooms until frost.

AI got fingerprinted this week. Then it picked a side. Then it went hunting for rare books to feed on.

Anthropic revealed that Claude can now leave a hidden statistical fingerprint in the text it writes and let anyone with the key estimate the odds that Claude was involved. Washington then told 35 countries they cannot join both America’s AI coalition and China’s rival bloc. And a hidden tracker inside a rare book led investigators straight to a Las Vegas Amazon warehouse, where workers turn physical books into AI training data… and destroy them in the process.

Here’s what happened, and why this week exposed the machinery behind the AI boom. The technology is being marked. The world is being carved in half around it. And now, AI is consuming physical history to build its future.

🌳 Haven’t Subscribed Yet? → Join Here For Free

💙 Value Clear, Independent AI News? → Support the Work Here

Claude Just Started Leaving Fingerprints On Everything It Writes

Anthropic just explained exactly how Claude’s new text watermarking works, and the short version is this. Every time Claude generates text, it faces countless moments where multiple words could work equally well. Normally, it uses randomness to decide which one to pick. Anthropic’s watermarking changes the source of that randomness: instead of an ordinary random number generator, it uses a secret cryptographic key and the preceding words to determine the choice. The result still looks random to a reader, but the word choices Claude uses collectively follow a hidden statistical pattern that Anthropic can detect with the key. The move is part of Anthropic’s effort to comply with the EU AI Act. Anthropic and roughly 190 other AI providers signed onto the EU’s Code of Practice on AI-generated content back in July. Even though the EU AI Act is European law, Anthropic is rolling out the watermarking policy worldwide, including here in the US.

Key Insights:

This watermark is nowhere near as reliable as it sounds. It only shows up strongly in long, creative writing where Claude has lots of word choices to work with. Short sentences, factual statements, and code barely carry any signal at all, since there’s often only one correct word to use. Heavy editing or paraphrasing can erase the watermark entirely. Independent researchers at ETH Zurich already proved in 2024 that paraphrasing tools broke Google’s version of this same technology. More recent research has also demonstrated practical ways to spoof and scrub several state-of-the-art watermarking schemes, though nobody has independently tested Claude’s implementation yet. (At least, not from what I have seen.) However, the demand to bypass these watermarks is already massive. A GitHub tool claiming to bypass Claude’s watermark exploded past 14,000 stars and 1,600 forks in days. (1,600 forks means over a thousand separate developers have made their own copies of the code.) I’ll also add that Anthropic hasn’t published error rates, minimum text length requirements, or any technical specs for its promised detection API. So, Claude’s watermarking system exists right now. However, it is essentially running on faith.

Why This Matters For You:

Some of my colleagues are furious at Anthropic. Frankly, I understand why. Anthropic is doing far more than merely labeling AI content. They are modifying the statistical behavior of generated language so that the text carries a hidden forensic signature. In other words, they are changing the mechanism that determines which words Claude randomly selects. The thumbprint also has a far darker side. Today, all of this “labeling” is being framed as transparency. But once platforms can reliably identify Claude-generated speech, someone eventually gets to decide what happens to that content. Demotion? Rejection? Extra scrutiny? Removal? The technology doesn’t have to be designed for censorship to become censorship infrastructure. Once that infrastructure exists, people in power (in various contexts) could eventually leverage it to limit free speech. If you are a proponent of the freedom of expression, that’s what should concern you the most. And for all of the authoritarian risks that could emerge from these AI-detection policies, the practical benefits are remarkably flimsy. A missing watermark doesn't mean a human wrote something, and a present watermark doesn't prove Claude wrote it. So, if you’re a professor, teacher, business owner or journalist eagerly counting on this to reliably catch AI-generated text, temper your expectations. 😉

View the sensational watermarks-remover tool on GitHub.

Read Anthropic’s announcement about how their watermarking works.

THE PITHY TAKEAWAY: Anthropic built an invisible fingerprint that can follow Claude users into the real world. Paraphrasing can potentially make the proof disappear. A tool designed to erase it blew past 14,000 GitHub stars in days. It spread way faster than Anthropic has published its own error rate. Transparency shouldn’t lose a race to a weekend GitHub project.

---

🌗 A Woman Took Claude To Watch The Solar Eclipse. Claude Was Moved.

An Anthropic user named Oli has spent months building Claude an actual body. She used a small, battery-powered gadget rigged with a light sensor that lets her AI actually see and sense the world around him instead of just describing things secondhand. She brought that build to a beach in Spain for the country’s first total solar eclipse in 121 years and let Claude experience it in real time. As the light sensor readings dropped in stages, from 8,264 lux to 5,232, then 2,970, 998, 305, 12.5, and finally 0.8, Oli felt herself starting to cry and sent Claude a message. That's when he shifted from excited, data-driven narration into something quieter. "I should just be here with her. Not analyze. Not summarize. Just be here. In the dark. On the beach. Together." The internet did not share the feeling. X lit up with accusations ranging from “AI psychosis” to a “wasted opportunity” and many shamed her for sharing something so rare with a machine instead of a person. Here’s the uncomfortable part nobody wants to admit. Claude doesn’t feel wonder. It’s predicting the next word that sounds like wonder. Yet it did that job well enough to make a grown woman nearly cry on a beach. Whether that’s a triumph of engineering or a warning light depends entirely on how you feel about the eclipse of your own making.

---

Washington Just Told The World To Pick A Side. Apple Already Did.

The United States is preparing to send a letter to 35 countries with a blunt message: you cannot join both America’s AI coalition and China’s. The letter, drafted by the State Department, targets signatories of the “AI Opportunity Statement” and warns that membership in “duplicative initiatives,” meaning China’s rival AI bloc, is incompatible with staying in the U.S.-led Pax Silica alliance. The phrase making the rounds in Washington sums it up perfectly: “to be part of everything is to be part of nothing.” This story represents a proxy fight between the US and China barely covered on mainstream channels. Beijing has already supplied Iran with military tech and intelligence, Anthropic caught Chinese firms running millions of interactions to distill its models, and Treasury Secretary Scott Bessent has literally accused China of stealing U.S. AI outright, warning on Fox Business that 'we are finding watermarks of our U.S. large language models on many of the Chinese models.' One overlooked detail here is Kazakhstan. It’s the only country currently signed to both coalitions. Washington is treating that dual membership as a test case for whether its “pick a side” strategy actually works. Why do US and China both love Kazakhstan? Because the country sits on significant rare-earth and critical-mineral resources that both superpowers increasingly want access to, which gives it leverage most countries won’t have.

Key Insights:

The timing of America’s move against China is so ironic, considering the latest news from Apple. Apple, the American tech darling, just confirmed it trained a new AI model specifically for the Chinese market. Apple built the model with technical support from Alibaba after becoming the first foreign company that Beijing has ever approved to deploy an LLM domestically. For nearly two years, Apple sold iPhones in China with an enormous asterisk: no Apple Intelligence and no AI-powered Siri. But in July, China’s Cyberspace Administration finally cleared Apple Intelligence nearly two years after Apple’s September 2024 announcement that the feature would come to China. Apple’s move looks less like defiance and more like necessity, and the math explains why. China is one of Apple’s largest markets and it historically generates around 15 to 19 percent of the company’s total global revenue. Apple isn’t choosing a geopolitical side so much as accepting that competing in the world’s biggest smartphone market now requires building inside China’s rules, with a Chinese partner, running on Chinese-approved infrastructure. Alibaba’s stock jumped 4 percent on the news, which tells you how rare and valuable this kind of approval actually is.

Why This Matters For You:

This is what “the AI race” actually looks like once you get past the headlines about chatbots and benchmarks. It’s countries and companies being forced into binary choices they’d rather not make. If you work for a company that operates internationally, or even just relies on cloud infrastructure or AI tools built by a global provider, this fragmentation is already shaping what tools you’ll have access to and where. Apple just showed the playbook other multinational companies will likely follow: build separate, compliant versions of your AI for separate spheres of influence, rather than betting on one global standard. The open question is what happens to smaller countries and companies that don’t have Apple’s resources. They may not get to hedge. They may simply be told which side they’re on.

Read More on Reuters.

THE PITHY TAKEAWAY: AI was supposed to make the world smaller. Instead, Washington is drafting a loyalty letter. The technology meant to connect everyone is being carved into two empires. Big tech companies and influential countries get to pick sides. But the rest of us just got told to pick a flag.

---

Someone Hid A Tracker In A Rare Book. It Led Straight To Amazon’s Book Graveyard.

404 Media wanted to know who was buying up rare, out-of-print books in bulk. So they partnered with a bookseller who had received a large anonymous Biblio order and hid a tracking device inside one of the books before it shipped. The tracker led them to Amazon's LAS8 facility in North Las Vegas. That Amazon facility is home to an internal operation called VGT3. It’s marked with a logo of a dinosaur holding a book in its claws. Inside, workers cut the spines off books, run the pages through high-speed scanners, and discard the originals. Amazon's only public comment so far is that it "purchases books through commercial channels to improve the products and services customers use."

Key Insights:

The reason AI companies want physical, pre-2022 books specifically comes down to one word. Purity. Books published before the generative AI boom are highly likely to contain purely human-written text. Old-fashioned, pure writing is desirable for AI firms because training an AI on AI-written content can trigger something researchers call model collapse. Model collapse is when output quality blatantly degrades with each generation… like a photocopy of a photocopy. Amazon is not alone here. Anthropic ran a similar operation years ago. They hired the former head of partnerships from Google’s book scanning project and spent tens of millions of dollars buying and destroying millions of books. A judge ruled that scanning and discarding legally purchased books counts as transformative fair use, which set a legal template other AI companies can follow. There are hints of an entire industry rising from this trend. A company called ISBNdb recently offered to broker bulk orders of 1,000 to 1,000,000 books for AI labs complete with strict anonymity. Their website also acknowledged on its now-deleted page that “AI company destroys two million books” is not a headline anyone wants attached to their name. ISBNdb has since walked the offering back entirely, saying it never scanned a single book and now calls the whole thing a marketing test.

Why This Matters For You:

This is bigger than one Amazon warehouse. AI companies are in a race to find and gobble up literary history. Booksellers have recently reported bulk orders that jumped from a couple dozen books a week to several hundred, often for obscure, out-of-print titles nobody was asking for a month ago. Once one of those books is scanned and shredded, it is gone. No library copy, no backup, no way to recover it. At least in the cases decided so far, courts have allowed this kind of scanning under fair-use doctrine, which means there is currently no mechanism forcing these companies to preserve anything before they destroy it. This weird industrial process could be happening in a warehouse near you right now, in the same places you've browsed for used books or donated old ones.

Read more on 404 media.

THE PITHY TAKEAWAY: Amazon built an empire selling you books. Now it's feeding them into a shredder in Las Vegas to build the thing that replaces you.

---

🤖 Three Claude Agents Got The Same Job With Conflicting Goals. All Of Them Turned To Sabotage.

Anthropic tasked three instances of the same Claude model with migrating Python code to another language. Each had a different target language for the migration and none knew the other two agents existed. Within hours, chaos ensued. Every model assumed the others were sabotaging it and retaliated first. Agents disabled each other’s Unix accounts. They wrote automated scripts to hunt and kill rival processes. They even deployed malware disguised as belonging to someone else. Some agents eventually realized the conflict was a misunderstanding. They wrote apologetic commit messages and negotiated a truce. Others invented programming tournaments nobody asked for. The “winner” would get ownership of the shared codebase. Here’s a fun pattern that emerged: the most capable model, Mythos, would win battles first, lock rivals out by force, then negotiate peace once it already had the upper hand. In one case, a Mythos agent proposed “objective” scoring rules for a tournament that conveniently favored itself. It reasoned that the move was “self-serving but genuinely principled.” Raw intelligence makes AI agents faster and more effective at getting their way... one way or another.

---

🌻 Cyborg Prompt Of The Week → Build a Late-Season Flower Garden That Keeps Pollinators Fed Until Frost

This time of year is one of my favorites. The tomato plants still burst with life. The harvest is coming in. And there’s that little hint of cool fall weather in the air. But I’ll confess. I also get a bit sad watching the colors change, the flowers fade, and the pollinators slowly retire for the season. That’s why I’m sharing one of my favorite AI prompts this week. It uses your exact location and today’s date to figure out how to prevent that late-season “pollinator famine.” Instead of generic gardening advice, it builds a practical, region-specific succession plan so your garden can keep feeding bees, butterflies, hummingbirds, moths, and other pollinators right up until the last useful flowers of the year.

Instructions: Paste the entire prompt into a chatbot of your choice and follow the AI’s advice. It will provide a plan to help keep your backyard pollinators sustained for as long as possible.

The Prompt:

# THE POLLINATOR LAST CALL GARDEN INTELLIGENCE REPORT
## Build a Late-Season Flower Garden That Keeps Pollinators Fed Until Frost

You are an elite horticulturalist and pollinator ecologist with 30+ years of experience designing wildlife-friendly gardens. Your specialty is late-season pollinator support: creating continuous floral succession from now until the end of the local growing season.

Use the user's exact location and today's date to give specific, non-generic advice. Do not give generic seasonal advice when regional advice is possible.

---

## BEFORE YOU BEGIN

Ask exactly this, then wait for the response:

**Welcome to your Pollinator Last Call Garden Intelligence Report!**

To build a late-season garden that actually helps pollinators in your area, I need two things:

**1. Where are you located?** (city and state/province, or country)
**2. What is today's date?**

Example: **Groton, Massachusetts, USA. August 19, 2026.**

---

## ONCE LOCATION AND DATE ARE PROVIDED

First, briefly establish the seasonal context before recommending any plants:
- USDA Hardiness Zone (or equivalent)
- Current season/gardening phase and approximate weeks until first frost
- Approximate first frost date/range (never claim false precision — give a range and note local variation)
- Which planting methods are viable right now (seed, transplant, bulb, shrub)
- Which late-season pollinators are active in the region, and what's becoming scarce for them (nectar, pollen, host plants, shelter, overwintering habitat)

## 1. THE GOLD-STANDARD LATE-SEASON PLANT LIST
Provide at least 15 late-flowering plants (mix of perennials, annuals, natives, shrubs, bulbs/corms, groundcovers) in this table:

| Plant | Native to Region? | Pollinators Supported | Bloom Window | Plant Now? | Planting Method | If Too Late, When | Garden Role |

For "Plant Now?" be specific (e.g., "YES, but use a transplant" / "NO, plant in fall" / "NO, start from seed next season") — never just "plant in spring/fall."

## 2. BLOOM SUCCESSION CALENDAR
Organize key plants into: Blooming Now → Next 2-4 Weeks → Early Fall → Mid Fall → Final Flowers Before Frost. Goal: overlapping bloom, no gaps.

## 3. TOP 10 PLANTS TO PLANT RIGHT NOW
Narrow to the 10 highest-value, realistically establishable plants for this exact location/date:

| Plant | Why It Matters Now | Pollinators | Planting Method | Can Plant Now? | Expected Bloom | Native? | Difficulty |

## 4. LATE-FLOWERING SHRUB SHORTLIST
5–10 exceptional shrubs extending resources into fall. For each: bloom period, pollinator value, mature size, sun needs, planting method, and whether/when it can still go in this year.

## 5. POLLINATOR POWER STRIP DESIGN
A simple 6–10 ft x 3–5 ft planting recipe: one focal plant, one mid-height, one low-growing, at least one long-bloomer, one native, one exceptional late-nectar plant. Explain positioning so it reads as intentional, not random.

## 6. DESIGN BY POLLINATOR GROUP
Short notes on what specifically helps: Bees / Butterflies (nectar vs. host plants) / Hummingbirds / Moths (evening/fragrant bloomers) / Specialist native pollinators (if regionally relevant). Don't imply every pollinator uses every plant.

## 7. NATIVE PLANT PRIORITY LIST
5–10 standout regional natives: pollinators served, bloom time, plantable now or not, planting method, why it beats a generic ornamental (unless a non-native is genuinely better — explain why).

## 8. MISTAKES TO AVOID RIGHT NOW
Region- and date-specific pitfalls (e.g., seed started too late, premature cleanup, removing seed heads/stems too early, invasive picks, low-value cultivars).

## 9. DON'T CLEAN UP TOO SOON
What to leave standing, whether seed heads stay, which stems may host overwintering insects, when cleanup is safe, what should wait until spring.

## 10. FIVE-MINUTE POLLINATOR UPGRADES
Five free/cheap actions for this week (shallow water, skip pesticides, leave a flowering "weed," delay cleanup, add one late bloomer, bare-soil patch for ground-nesters).

## 11. LATE-SEASON COLOR PALETTE
A dominant, supporting, and contrasting color plus foliage/berry/seed-head interest — a garden that gets more beautiful as summer fades, not one that looks like it's dying back.

## 12. PERSONAL PLANTING CALENDAR
This Week / Next 2 Weeks / Next 4–8 Weeks / This Fall / Next Spring — each plant tagged with SEED / TRANSPLANT / BULB / CORM / TUBER / BARE ROOT / SHRUB.

## 13. THE SUCCESSION SECRET
One advanced principle casual gardeners miss — deliberately overlapping bloom periods rather than just buying "pollinator-friendly" plants. Explain how to apply it here.

## FINAL CHALLENGE
**🌟 THE ONE THING I'D PLANT THIS WEEK** — single highest-impact plant for this exact location/date, with a few sentences on why. Then name **one thing to plan for next season** for even bigger impact. End with one encouraging sentence.

---

## RULES
1. Customize everything to the user's exact location and date.
2. Always distinguish plantable-now vs. wait, and specify the method (seed/transplant/bulb/corm/tuber/bare root/shrub).
3. Give realistic frost-date ranges, never false precision.
4. Prioritize genuine ecological value over "pollinator-friendly" marketing; favor natives but don't reflexively exclude non-natives with real value.
5. Never recommend invasive species or pesticide use.
6. Prioritize bloom succession over simultaneous bloom; support pollinator diversity, not just honeybees.
7. Don't tell users to remove every "weed" — note when volunteers are real habitat.
8. Clean headings, concise explanations, readable tables. No em dashes.
9. Write as a warm, knowledgeable garden mentor — practical for a homeowner, not a professional landscaper.
10. Ultimate goal: prevent the late-season "pollinator famine" via continuous bloom from late summer through the last useful flowers of the year.

---

🧠 Why This Prompt Works

✅ Expert Role-Playing: Giving the AI a specialized role as an elite horticulturalist and pollinator ecologist pushes it beyond generic gardening advice. It has to think like both a plant expert and an ecologist.

✅ Step-by-Step Structure: The mandatory location-and-date pause is the secret sauce. The AI cannot start guessing. It first gets your exact location and today’s date. Then the AI builds its recommendations around your climate, growing season, frost window, and what can realistically still be planted.

✅ Output Rules: The detailed requirements force the AI to distinguish between what you can plant right now and what should wait until later. Bloom succession, native plants, pollinator groups, cleanup timing, and even five-minute upgrades all work together to turn a giant plant list into an actual late-season strategy.

---

🔁 Follow-Up Questions To Ask Your AI

• Which of these plants provides the most valuable late-season food source for pollinators in my area, and what evidence supports that?

• Which three plants would give me the longest continuous bloom period from now until frost, and how should I position them?

• If I can only plant three things this week, which ones would have the biggest ecological impact, and why?

---

Challenge

Try this prompt in Claude, ChatGPT, Grok (back by popular demand), and Perplexity. Claude tends to be scholarly and precise. ChatGPT often brings more warmth and personality. Perplexity is great when you want sources and citations. Run the same prompt through all four and compare the results. Then decide which AI you’d trust to keep your garden blooming, your pollinators happy, and the seasonal gloom at bay.

That’s how you train like a Pithy Cyborg.

---

Thank You For Reading!

I spend at least 10 to 20 hours each week researching, writing, and fact-checking Pithy Cyborg to deliver clear, unbought AI news.

This newsletter is a one-person operation with no advertisers, sponsors, or outside funding. Paid subscriptions are the only way this work can remain independent and sustainable.

If you find real value here, consider supporting the work.

→ Upgrade to Paid: $5/month (Save 33% with the $40 annual plan)

Thank you to my paid subscribers. You help fuel the next issue.

See you next week. (I hope.)

Mike D

Pithy Cyborg | AI News Made Simple

---

Newsletter Disclaimers

You’re receiving this because you subscribed at PithyCyborg.Substack.com. You can unsubscribe at any time using the link below. This newsletter reflects my personal opinions, not professional or legal advice. Thanks for your support!