Two years ago we worried AI would take our jobs. This week the internet reminded us of a more immediate problem: AI is taking our attention, our trust, and our patience. The biggest story of the week wasn’t a model release. It was a full-blown rebellion against unedited machine output, complete with a suspected fake think tank, a book-pulping scandal, and one genuinely excellent vision model. OpenAI also quietly disbanded the team responsible for assessing catastrophic model risks. That’s the kind of headline that should worry you more than it did.
Everyone Is Done Reading Raw ChatGPT Output
Two posts dominated the week, and they said the same thing in different tones. “Don’t paste the AI, please” is a blunt little site telling people to stop dumping chatbot walls of text into conversations: “A short answer from a human will always be worth more than one from a robot, even if the robot is right.” It even ships an angrier version for people who want to burn bridges. Days earlier, Rick Manelius published “AI;DR (AI; Didn’t Read)”, a policy that spread like wildfire: if you won’t bother editing the AI output, he won’t bother reading it. His tweet pulled in 346,000 views and 16,600 likes, and the two pieces combined for more than 2,000 Hacker News points and over 1,200 comments. A wall of Claude output in a Slack thread is now the fastest way to get ignored.
Why it matters: when the people who actually use these tools start flinching at unedited output, that’s not a niche complaint. It’s a market signal that AI slop has become a brand problem for the entire industry, and the tools that make output feel human will win the next phase.
Don’t paste the AI, please | HN discussion
A Fake Think Tank Was Built to Manipulate AI Chatbots
An investigation by Responsible Statecraft alleges that Israeli operatives created a fake think tank designed to shape what AI chatbots say. The scheme, as described: seed the web with apparently neutral, expert-sounding content, let AI crawlers absorb it as training data, and let the models launder the narrative back to millions of users as objective fact. The report went nuclear on Hacker News, pulling over 1,000 points and 855 comments, many of them arguing about how strong the evidence really is. That ambiguity is itself the point: chatbots are now treated as a propaganda channel worth attacking, which is a milestone in itself.
Why it matters: if a coordinated group can bend what chatbots “believe” by flooding the internet with fake scholarship, then model answers are only as trustworthy as the least-scrutinized corner of the web. Every AI company now has a disinformation problem, whether it admits it or not.
Anthropic’s Watermark Scheme Is “a Perversion of Writing”
John Gruber’s Daring Fireball took a blowtorch to Anthropic’s watermarking approach, calling the technique “text adulteration” and “a perversion of writing.” The argument: watermarking is supposed to help detect machine text, but this version quietly corrupts the text itself and breaks the trust between writer and reader. The post clearly struck a nerve, racking up 822 points and 728 comments within days.
Why it matters: how AI companies mark machine text is becoming a three-way fight between detection, transparency, and the integrity of human writing. Anthropic just volunteered to be the battleground.
Amazon Is Destroying Rare Books to Train AI
An Ars Technica investigation planted an AirTag inside a shipment of rare, out-of-print books and traced them to a pulping facility, alleging Amazon is destroying them to feed AI training pipelines. Yes, the company that started as an online bookstore is allegedly shredding the very inventory it once sold. Anna’s Archive responded with an urgent, widely shared plea: scan rare books before they’re gone for good, because the window is closing fast and there’s no undo button for a burned library.
Why it matters: the training-data scramble just turned physically destructive. Whether the books end up shredded or scanned, the industry is literally consuming the world’s archives to build models, and the ethics of that deserve far more than a footnote.
AirTag investigation (Ars Technica) | Anna’s Archive plea
GPT-5.6 Sol Might Be OpenAI’s Best Vision Model Yet
Roboflow put GPT-5.6 “Sol” through its vision benchmarks and came away impressed, calling it the best vision model OpenAI has ever shipped, with standout gains in document understanding and OCR-heavy tasks. For people building automation around receipts, forms, and screenshots, that’s not trivia, that’s a workflow upgrade. It was also the one piece of unambiguously good news in a week dominated by trust stories.
Why it matters: capability keeps climbing even as confidence wobbles. The gap between what models can do and what people will trust them to do is now the defining tension of the industry, and Sol is a reminder that the underlying tech is still moving fast.
Roboflow’s review | HN discussion
What This Means for You
First, the AI;DR rule is now in force everywhere. If you use AI at work, edit, trim, and personalize the output before you send it. The people on the other end have the same tools you do, and they can smell raw model text from across the room. Second, treat every chatbot answer as one input, never as the answer. The fake think tank story is a warning that training data is now an active attack surface, and confident-sounding text is not evidence of anything. Third, audit your own content. Unedited AI blogs and comment walls are becoming a reputation liability, and both readers and search engines are getting better at spotting them by the week. That doesn’t mean quitting AI; it means doing the human half of the job: judgment, context, and taste. Finally, don’t sleep on the actual progress. Vision models like GPT-5.6 Sol are doing real work in document processing and automation today, and the teams that pair them with solid human review are pulling ahead. Use the tools, keep the judgment. That combination is the only durable skill left.
If you’re tired of your AI output sounding like a robot wrote it, we built a practical guide to fixing exactly that.
The through-line this week is trust. Models get more capable by the week while the people who use them get more skeptical by the day, and that gap is where the real action will happen. Add in a disbanded safety team, a shrinking financing guarantee, and an industry literally pulping books for data, and the picture is clear: the AI era is entering its accountability phase. The winners of the next year won’t be the teams shipping the most tokens. They’ll be the ones who figure out how to use AI without making everyone around them flinch.
About The Author
