AI Deepfakes Are Targeting Teachers — And No One's Responsible

Something about today's AI news feels unusually human — the teachers whose faces got weaponized, the kids who still learn better than machines, the authors whose life's work got fed into a system without a phone call. The technology keeps advancing. The people keep getting left holding the bag.

Teachers Are Being Targeted With AI Deepfakes, and the System Has No Answer

Students are using AI tools to generate sexualized fake images of their teachers, and the four educators WIRED interviewed describe something consistent and damning: when they reported it, nobody knew what to do. Schools deferred to police, police deferred to schools, and the platforms hosting the content mostly shrugged. There's no law that covers this cleanly, no platform policy that stops it, and no official whose job it is to fix it. The kid who did it probably faced nothing. If you run a school and you don't have a written policy for this, you're going to be improvising during the worst week of someone's career.

https://www.wired.com/story/teachers-deepfake-ai-students-content/

Children Still Learn Language Better Than AI, and Researchers Can't Explain Why

Four years after ChatGPT, large language models (AI systems trained on billions of words to generate fluent text) can pass the bar exam and write better prose than most adults. But they still can't acquire language the way a three-year-old does: from a few thousand hours of messy, real-world interaction, without labeled data or gradient descent (the mathematical process that AI uses to correct its own mistakes). Most AI coverage either overclaims or dismisses. This piece just says we don't know, and then actually explains what we don't know. That's rarer than it should be. The leading theory is that children aren't just pattern-matching words — they're connecting language to physical experience and social feedback in a way that no text dataset can replicate. Or that's one theory. Nobody agrees. People keep saying AGI is 18 months away. A three-year-old is still beating the state of the art at something fundamental. Make of that what you will.

https://www.technologyreview.com/2026/08/24/1141740/kids-machines-language-learning/

A Mystery Model Called Ox Alpha Is Circulating, and Nobody Knows Who Built It

A new AI model called Ox Alpha has appeared in certain technical forums, performing well on benchmarks (standardized tests that researchers use to compare AI models) while offering no information about who made it, how it was trained, or what data it used. The speculation ranges from a stealth startup to a research lab doing a quiet soft launch — and some forums are floating nation-state actors or a rogue team inside a major lab, which may be fringe but isn't obviously wrong. What's actually interesting here isn't the conspiracy theorizing — it's that the model release process has become so informal that a capable AI system can just appear with no institutional backing and get taken seriously immediately. A year ago that wouldn't have happened. Now it's a Friday news story.

https://techcrunch.com/2026/08/23/whos-behind-the-new-stealth-model-ox-alpha/

Training AI on Copyrighted Books: The Legal Situation Is Genuinely Murky

Most major AI companies scraped published books to train their models, and most authors had no idea it was happening. The TechCrunch piece does something useful: it actually explains why this isn't a clean legal violation despite feeling like one. The core argument companies make is "fair use" — a legal doctrine that allows limited use of copyrighted material for purposes like commentary, research, or transformation. Courts haven't settled whether feeding a book into a training dataset counts. If you're a consultant, writer, lawyer, or anyone who produces original content professionally, this ambiguity matters because it shapes what compensation, if any, creators can ever expect from AI companies that got rich on their work.

https://techcrunch.com/2026/08/23/is-it-legal-to-train-ai-models-on-copyrighted-books-its-complicated/

Flock Safety Is Learning That Surveillance Requires a Social Contract

Flock Safety sells AI-powered license plate readers and camera networks to police departments and homeowners associations across the country, and it's now facing serious public pressure over how that data gets used and shared. The CEO's call for "compromise" is interesting mostly because it suggests the company was caught off guard by the backlash, which itself tells you something about how surveillance tech has been sold: quietly, locally, one contract at a time, until it's suddenly everywhere and people start asking questions. For anyone in operations, legal, or city government who has been approached about these systems, the reputational risk is now clearly part of the calculus, not just the operational benefit.

https://techcrunch.com/2026/08/23/flock-ceo-calls-for-compromise-as-surveillance-company-faces-growing-backlash/

The through-line in today's news is that AI's hardest problems aren't technical anymore — they're about who gets hurt, who's responsible, and who decided the rules.

← More daily briefs