2 stories, chosen and edited by hand.
01
Anthropic says it will apply machine-readable watermarks to content processed by its models, not just generated by them, to comply with the EU AI Act. All new models globally will mark output "from day one": invisible embedded watermarks in text, C2PA provenance metadata for other files. The Act covers models released after August 2, with a grace period to December 2026. No detection tool has shipped yet.
My readThe part that got me: a watermark applied at the model level can't tell wholesale generation from a comma fix, so Claude will stamp exactly the grammar-correction cases the EU carved out. If you run user prose through Claude for cleanup, that text now carries a mark you can't see and can't check, because the detection tool doesn't exist yet. I want to know how it survives a copy-paste and a rewrite. Until then it's a promise, not a signal.
Read the original at Ars Technica AI →
02
MIT Technology Review's Jen Swetzoff and Keeley McNamara interviewed kids aged 10 to 18 about AI for the print magazine Anyway. Most admitted using it at least a little; several refused, citing data-center water use or creativity. Pew Research Center, in a survey published February 2026, found 57% of US teens used chatbots to search, 54% for schoolwork, 47% for fun, and 12% for emotional support or advice.
My readThe product feedback buried in here is better than most reviews. Remy won't use AI for game dev because it writes sloppy code and can't blend mechanics, so correcting it costs more than writing it. Wesley noticed Character.AI's free tier forgets the chat and slows down. Krishiv built Aceflow because Khan Academy's tutor sat behind a waitlist. And 12% for emotional support, against the panic. I'd hire the 16-year-old who says it's the first airplane, interesting but crude.
Read the original at MIT Technology Review →