Everyone's drawing lines, and AI keeps crossing them anyway
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:55
Yesterday in AI — Weekend Recap | Monday, May 4, 2026
Everyone's drawing lines, and AI keeps crossing them anyway
This weekend, Hollywood, a Chinese courtroom, and a frontier AI benchmark all had something to say about what AI can and can't replace — and the answers might surprise you. We also dig into an economic insight hiding inside a coffee chain's staffing decision, a fraud wave that just got a lot harder to detect, and a brain implant that might finally crack a problem medicine has been stuck on for decades. Six stories. A lot to think about heading into Monday.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence in 10 minutes or less. I'm Mike Robinson. It's Monday, May 4th. May the fourth be with you. And this weekend everyone was drawing lines. Hollywood said AI can't take an Oscar, a Chinese court said AI can't take a job, and researchers quietly reminded us that AI still can't reason its way through a genuinely new problem. Let's get into it. The Academy of Motion Picture Arts and Sciences updated its eligibility rules last week, and the core decision is pretty direct. To be nominated for best actor or any acting category, a performance now has to be, quote, demonstrably performed by humans with their consent. End quote. Screenwriting nominees require human-authored scripts. The Academy says it can request documentation about AI usage anytime a question comes up. What's important here, this isn't a ban on AI tools and production. Films can still use AI for visual effects, sound design, or editing. But the credit, the nomination, the award, that has to flow to a human being who actually did the work. It matters because the Oscars aren't just a trophy, they're an industry contract. Actors negotiated sag after agreements around consent and compensation partly because studios were testing how far they could go with synthetic performances. A humans-only rule at the academy level forces studios to document what's human and what isn't before award season, which means the paper trail exists before legal disputes arise. That changes how contracts get written upstream. And it will ripple. Other awards bodies follow Hollywood's lead. Guild standards follow award standards. Taylor Swift filed trademarks on two audio clips of her voice this year for the same reason. Copyright law doesn't cover AI imitation well, so artists are racing to build legal structures that do. The Oscars just gave those efforts institutional backing. You'll see other creative industries watching this closely. From Hollywood to your neighborhood coffee shop. Well, actually to a really important economic story hiding inside a business decision that looked like a vibe shift. Starbucks CEO Brian Nichol told The Guardian last week that the company is rolling back its automation push. More baristas, handwritten names on cups, ceramic mugs, seats that people actually want to sit in. He called it a return to hospitality, and investors mostly nodded along. A University of Chicago economist named Alex Imus published an essay around the same time that may be the clearest framework I've seen for explaining what this actually means. His argument, when AI drives the cost of most things towards zero, what becomes scarce is what AI genuinely cannot replicate. He ran experiments on this. People paid roughly twice as much for identical items when they knew others were excluded from getting them. AI-generated art earned about half the exclusivity premium of human-made art, 21% versus 44%. His conclusion: a relational sector emerges where the human presence is the actual product. Teachers, therapists, baristas, craft brewers, live performers. Not because the AI output is worse, but because people will pay specifically for the knowledge that another human made it for them. The historical parallel is striking. Farming was 40% of U.S. jobs in 1900. It's under 2% now. Nobody starved. Spending shifted. Imus argues we're at the start of something similar. Starbucks isn't making a vibe choice. They're responding to a real market signal. Customers who can get good coffee anywhere are paying more to have a person make it for them. The labor market implications of that shift are going to take years to play out, but this weekend's coverage gave us a solid example of it already happening. If the Starbucks story is about where labor might win, the next one is about where the legal system is starting to push back on AI job displacement. A Chinese court handed down a ruling last week that deserves serious attention. A quality assurance supervisor named Zhu was hired in 2022. His employer replaced him with an AI model in 2025, offered him a 40% pay cut and a demotion, and then fired him when he wouldn't accept it. He sued. The Heng Zhu Intermediate People's Court sided with him on appeal. The ruling's logic, you cannot terminate an employee because AI has replaced their function. If you want to lay someone off legitimately, you have to show genuine operational difficulty or company-wide downsizing. Swapping in software doesn't count. That's a meaningful legal standard. China has hundreds of millions of service and manufacturing workers who are going to face exactly this scenario over the next decade. Setting a clear rule now that the AI is cheaper isn't a valid termination reason, shapes how employers deploy automation across the entire economy. Watch for this to spread. The EU's AI Act has some labor protections written in, but they haven't been tested in court yet. China just showed one way the jurisprudence can develop. And it matters for U.S. companies operating globally too, because their employment practices in these jurisdictions now have legal exposure they didn't have before this ruling. Let's talk about something the AI industry doesn't talk about enough. Where frontier models genuinely fall short? Because the honest answer is instructive. The ARC Prize Foundation published an analysis on Friday testing OpenAI's GPT 5.5 and Anthropics Opus 4.7 on something called ARC AGI 3. It's 135 game-like environments with rules the models have never seen before. Novel problems, no training data shortcuts. GPT 5.5 scored 0.43%. Opus 4.7 scored.18%. That's not a typo. The researchers went further than just reporting the scores. They watched the models work through the problems and identified three failure patterns that keep coming up. 1. The models understood local rules but couldn't build a consistent picture of how the whole environment worked. 2. They kept mapping new situations onto familiar games from their training data, essentially trying to play Tetris when the rules were something completely different. 3. They'd stumble through early levels without actually understanding the mechanic, then fail completely when things got harder. There's also a difference in how the two models fail. OPIS tends to lock in on one confident theory and stick with it even when the evidence is mounting against it. GPT-5.5 generates a lot of competing hypotheses, but can't narrow them down into a plan it can execute. Why does this matter for people deploying AI? Because real-world agents get dropped into systems and workflows they haven't been explicitly trained on. A model that's genuinely impressive in familiar territory and near useless in novel territory is fine for well-defined repetitive tasks. It's a liability in anything open-ended. This result should be somewhere in your mental model every time you're evaluating what to let an AI handle autonomously. Now one that should make you look twice the next time someone sends you a digital document. OpenAI's ChatGPT Images 2.0 has gotten good enough that people are using it to generate fake IDs, fraudulent bank alerts, forged prescriptions, and phony receipts. The Atlantic covered this weekend with specific examples, and the results are uncomfortable. OpenAI has guardrails in place. Researchers say those guardrails are being bypassed with regularity. The fraud risk here is real and practical. Banks, hospitals, insurance companies, and government agencies are going to face a surge in document fraud where the forgeries are photorealistic enough to fool human reviewers. The old advice, look for bad grammar, obvious Photoshop, watermarks, is effectively useless now. What hasn't caught up is detection. Most fraud verification systems were built around older attack patterns. An AI-generated fake driver's license doesn't trigger the same flags as a scanned altered document. The institutions most exposed are the ones that have been slowest to update how they verify identity and documentation. There's no clean answer right now, but anyone responsible for fraud prevention at a bank, an insurance company, or a health system needs to be treating this as an active problem today, not a future one. The volume is going to scale fast and the window to get ahead of it is closing. To close out the weekend's recap, a medical story that is AI adjacent and worth knowing about. A company called Motif Neurotech got FDA clearance last week to begin clinical trials of a brain implant for treatment-resistant depression. The device is called DOT. What's different about DOT compared to existing implants is where it sits. Most deep brain stimulation devices require electrodes that physically penetrate brain tissue. DOT sits above brain tissue on the surface and delivers programmable electrical stimulation without going in. Lower surgical risk, easier to recalibrate after implant. The target population is specific. About 3 million Americans live with depression that doesn't respond to medication, therapy, or any of the other standard treatments. They cycle through options and nothing holds. DOT is aimed squarely at this group. Clinical trials are just starting, so this is not a cure announcement. It's the FDA saying the design is safe enough to test in humans, but the combination of a less invasive design, a well-defined target population, and a real unmet need makes this one worth following. Brain computer interfaces for mental health have had a rocky path. A lot of early promise and early setbacks, then stalled trials. The DOT trial will tell us whether this approach is genuinely different. One more thing. If you like this podcast, please be sure to rate and review it so others can find it. It really does help. Thanks. That's all for this weekend catch up edition of Yesterday in AI. Stay curious, and I'll see you tomorrow.
