The Scan  ·  July 2026

As covered previously in this newsletter, when you ask multiple leading large language models to pick a random number between one and ten and a disproportionate number will land on seven — the same number humans default to, for reasons buried in training data scraped from human behavior. A startup profiled by MIT Technology Review is tackling this directly, building techniques to deliberately diversify the outputs of LLMs so they don't all converge on the same "creative" answers. The problem runs deeper than number-picking: when models trained on similar corpora, fine-tuned with similar feedback mechanisms, and aligned toward similar notions of helpfulness all reason their way to the same conclusions, the epistemic diversity that makes multiple perspectives valuable simply evaporates. What's striking is how cleanly this mirrors the human trap — the more we optimize for approval and coherence, the more we sand away the productive friction of genuine originality. The fix for machines (deliberately injecting variation, rewarding divergence) turns out to be decent advice for knowledge workers, too.

Amazon has announced it will stop accepting new customers for Mechanical Turk, the crowdsourcing platform that launched in 2005 and at its peak connected hundreds of thousands of human workers — called "Turkers" — with businesses paying fractions of a cent to label images, transcribe audio, and perform the cognitive tasks that early AI couldn't handle alone. The platform's name was always its most honest feature: it was borrowed from an 18th-century chess-playing "automaton" that concealed a human grandmaster inside, and the modern version did something structurally similar — routing human intelligence through an interface that made it look like machine output. The closure doesn't mean the work disappears; it means AI has become capable enough that companies no longer need to paper over its gaps with invisible human labor at scale. What replaces Mechanical Turk is a world where AI does the labeling, the transcription, the micro-judgment — and the humans still needed are fewer, more specialized, and considerably better compensated. That's progress, with an asterisk for everyone who was one of those workers.

Programs like Alpha Forge Prep, charging upward of $40,000 a year, are marketing AI-driven personalized instruction to affluent families as a superior alternative to traditional schooling — with curricula that adapt in real time to each child's pace, gaps, and learning style. The Verge's reporting details how wealthy parents are betting that AI's ability to deliver endlessly patient, perfectly calibrated instruction outweighs whatever a classroom of peers, a distracted teacher, or an argument on the playground might offer. The pedagogical literature has a fairly clear position here: much of what children need to develop — frustration tolerance, collaborative negotiation, the experience of being wrong in front of others — cannot be optimized away without cost. The families who can afford to run this experiment on their children will generate useful data for everyone else. The families who can't afford it will find out what was lost after the fact, which is how educational inequality has always worked.

Archive of Our Own, the fan fiction repository hosting more than 13 million works, is in the middle of a community crisis: readers and moderators are using AI detection tools including one specifically calibrated for Claude outputs, to flag and shame works suspected of being machine-written, despite those detectors carrying well-documented false positive rates that have already implicated human authors. The anxiety is real, and it isn't irrational; fan fiction communities have always been built on the premise that the story is an act of devotion from one human to a shared cultural object, and that compact feels violated when a model can generate plausible fan fiction in seconds. But the witch-hunt mechanics — detection tools weaponized with confidence they don't deserve, social pile-ons, authors defending their humanity in public — reveal something the discourse keeps sidestepping: we do not actually have a reliable way to define authentic creative expression, we just know we want it. Until the community can articulate what it values more precisely than "not AI," the detectors will keep catching the wrong people.

In an internal address to Meta staff, Mark Zuckerberg acknowledged that AI agents have not advanced as quickly as he had anticipated, according to TechCrunch's reporting. This matters precisely because Meta has staked significant infrastructure investment and product roadmap on the premise that agents would soon be handling substantive cognitive work at scale. The gap between the boardroom narrative of autonomous AI and the engineering reality of systems that still struggle with anything requiring reliable judgment across novel contexts is not a secret, but CEOs rarely say it plainly. What Zuckerberg is describing is a hard problem with a specific shape: it's relatively tractable to get an AI to perform a known task reliably, and it remains genuinely difficult to get one to decide correctly which task to perform, when to stop, and when to ask for help. That problem has a name in the research community the alignment of agency with intent (quite similar to the prinicipal-agent problem in finance & insurance). Solving it will require more than another parameter scaling run, and continued advances in what we humans are saying AI doesn’t have (aka ‘taste’).

Google DeepMind and A24 — the studio behind Everything Everywhere All at Once, Midsommar, and Hereditary — have announced a formal research partnership, described as the first of its kind between an AI research lab and a major film production company. The collaboration, per DeepMind's own announcement, is framed around exploring how AI tools can support and expand the creative process for filmmakers, rather than replace it. The pairing is genuinely unusual: DeepMind's portfolio runs from protein structure prediction to reinforcement learning benchmarks, not screenplay notes. A24's brand is built on the proposition that strange, difficult, human-scaled stories find audiences — which is either a perfect complement to AI's tendency toward the conventional, or an irresistible target for optimization pressure dressed up as creative partnership. In my opinion, the most interesting output of this collaboration won't be whatever film they produce together. It will be: where did humans and machines diverge, and who won the discussion. We’ll report on it here first.

The human mind is the original generative engine.