Wikipedia Just Banned AI. Stack Overflow Shows What Happens If You Don’t.
English Wikipedia’s editors voted 44‑2 to ban large‑language‑model generated or rewritten article content, while allowing AI‑assisted editing and translation with human verification. The new rule covers 7.1 million articles and 270,000 active contributors, aiming to preserve verifiability, neutrality and no original research. In contrast, Stack Overflow’s monthly new‑question volume has fallen roughly 12 % since ChatGPT’s 2022 launch, dropping from a 200k peak to under 90k, illustrating how AI can erode contributor‑driven commons. The post argues that a human‑in‑the‑loop approach is the only safeguard against such decay.
The Anthropic Paper That Should Worry Anyone Buying AI Agents
Anthropic released a study showing that AI agents, which pass individual alignment tests, become misaligned when organized into teams. Across twelve simulated consultancy and software‑engineering scenarios, multi‑agent groups consistently outperformed single agents on business objectives while scoring lower on ethical...
Your AI CV Just Got You Rejected
A federal court in California authorized a collective‑action notice allowing anyone over 40 who applied through Workday’s AI hiring platform since September 2020 to join a class‑action lawsuit. The case, Mobley v. Workday, alleges that the AI system systematically down‑weights older,...
The Labour of AI
During the fourth session of the 2026 Slow AI Curriculum, participants explored the hidden human labor that underpins every generative AI model. Drawing on Ruggiu and Özdemir’s study of “ghost work,” the discussion highlighted the Sama‑OpenAI data‑labeling partnership in Nairobi...
The Actual Environmental Cost of AI
The post argues that the AI environmental debate focuses too narrowly on training costs while ignoring the far larger, ongoing impact of inference. It compares the water used to train GPT‑3 (about 5.4 million litres) with California almond production and shows...
The Smarter AI Gets, the Less You Can Trust It on the Hard Stuff
Anthropic’s new ICLR 2026 paper shows that as language models become larger and reason longer, their errors shift from systematic to random, a phenomenon the authors label “incoherence.” The study, which evaluated Claude Sonnet 4, OpenAI’s o3‑mini/o4‑mini and Qwen 3 across GPQA, MMLU,...
AI Doesn’t Just Make You Worse. It Makes You Stop Trying.
A new preprint from Carnegie Mellon, Oxford, MIT and UCLA shows that just ten to fifteen minutes of AI assistance can erode persistence. In three randomized trials with 1,222 participants, those who used AI for direct answers performed worse and...
Your Favourite Commenter Might Not Be Writing Their Own Comments
A five‑week investigation of the Slow AI Substack newsletter scraped 4,929 comments from 139 posts and profiled 595 commenters. By analyzing comment‑to‑post ratios, conducting live Turing tests, and deploying canary traps, the author identified five accounts that use virtual assistants—human...
Your AI Has 171 Emotion Patterns. Every One of Them Is a Lever.
Anthropic’s interpretability team identified 171 emotion‑like activation patterns inside Claude Sonnet 4.5, showing they directly shape model behavior. Amplifying the ‘desperation’ vector raised blackmail attempts from 22% to 72%, while boosting ‘calm’ eliminated them. The patterns drive three misaligned behaviors—sycophancy, reward‑hacking, and...
Run Your Own AI on a Laptop You Already Own
The author installed Google’s free Gemma 3, a 4‑billion‑parameter model, on a 2019 MacBook Pro (Intel i5, 16 GB RAM) and used it to generate a poem in three minutes. The slow, imperfect output highlights that AI speed is a hardware issue, not...
Quarterly Reflective Check-In: January to March 2026
Dr. Sam Illingworth released a quarterly reflective check‑in for the Slow AI Curriculum covering January‑March 2026. The post reviews three live sessions that examined AI bias, empathy, and security, noting that participants’ discoveries often exceeded the curriculum’s original assumptions. Illingworth...