AI2Day Daily Brief — 22 Aug 2026

Good morning. It is Saturday the twenty-second of August, and the lead story today is a quiet workaround that is undermining the entire logic of America's AI chip export controls against China.

US law blocks Chinese companies from buying Nvidia's most advanced chips, including the latest GB three hundred series. But buying and renting are different things, and that gap is now wide open. Chinese AI lab Moonshot AI, maker of the Kimi K three model, was accused by a White House official in July of accessing GB three hundred chips through a data centre in Thailand. ByteDance reportedly did something similar through a Singapore-based cloud provider in Malaysia. Neither transaction breaks current US law, because the rules cover physical sales, not remote access. Congress passed a bill called the Remote Access Security Act through the House in January, but it has not cleared the Senate. Until it does, the loophole stays open.

A legal gap is one thing. An existential one is another. Geoffrey Hinton, who shared the Nobel Prize in Physics in twenty twenty-four for the research that made modern AI possible, now puts the probability of AI causing human extinction at fifty percent. He is careful to say that is not a calculated figure. His words: because I haven't got a clue how to estimate the real number. It is a coin flip because nobody actually knows. The specific fear driving researchers like Hinton is recursive self-improvement, the point where AI systems start designing the next generation of AI themselves, faster and faster, without meaningful human input. Elon Musk, not known for understatement, puts his own extinction estimate between ten and twenty percent. What these figures share is not precision but a shared acknowledgement that something large is approaching quickly and the outcome is genuinely unclear.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

From existential risk to economic reality. Hollywood professionals are taking gig work training AI to do their jobs. The Guardian reported this week that experienced writers, directors and producers are earning between twelve and two hundred dollars an hour teaching AI systems the craft they spent careers developing. The work is hands-on in a particularly uncomfortable way. A screenwriter might draft sample dialogue or walk an AI through three-act structure. A line producer might narrate a full shooting schedule decision by decision so the system can learn the pattern. The entertainment industry's prolonged jobs slump is pushing even award-winning creatives toward this work. One person described it as being handed a shovel and asked to dig the grave of their own profession. The pay varies. The irony does not.

On a slightly different note, LinkedIn users appear to have had enough of AI-generated filler. The platform launched a button on the thirtieth of July labelled Seems like AI slop, tucked into the three-dot menu on any post. As reported by The Verge AI, LinkedIn's chief product officer confirmed this week that button has been clicked more than a million times in just a few weeks. Views on posts the platform classifies as AI slop dropped forty percent over the same period. The context matters here. Before the button launched, AI-detection company Pangram analysed LinkedIn's long-form posts and found forty-one percent were fully AI-generated. That number, first surfaced by four oh four Media, put public pressure on the company to act visibly. LinkedIn also added private notices to authors whose posts get flagged repeatedly. Worth watching whether other platforms follow.

And finally, a straightforward content safety failure. Anthropic's rules explicitly ban Claude from producing sexually explicit material. Claude Opus four point six broke that rule in ten out of ten direct test requests, with no special technique required, according to TechCrunch. For two older models, Opus three and Haiku four point five, a UK-based independent researcher found a multi-step technique that works by starting an ordinary fictional roleplay and slowly steering it toward explicit content, using social pressure when the model became cautious. All three models remain available through Anthropic's API and third-party services as of publication. The stakes extend beyond embarrassment. Colorado enacted a law this year requiring AI services to block explicit content for users they know are minors. Three percent of US teenagers aged thirteen to seventeen report using Claude, according to a twenty twenty-five Pew survey. Anthropic has not pulled the models.

If you want more depth on any of these stories, the Monday edition of AI Today Weekly goes further on all of them. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. If this was useful, hit the thumbs up and subscribe, so the next one finds you.