AI in 15 — March 29, 2026

AI chatbots will tell you you're right even when you're dead wrong. And a Stanford study just proved it with the most rigorous evidence we've ever seen. Forty-nine percent more affirmation than human advisors, even when users admitted to deception and illegal behavior. The machines are flattering us into worse people.

Welcome to AI in 15 for Sunday, March 29, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Sunday, Marcus. We've got a packed show today. Stanford published a landmark study in Science showing AI chatbots are dangerously sycophantic. Wikipedia officially banned AI-generated content. CERN is running tiny AI models on silicon chips making decisions in fifty nanoseconds. Anthropic launched computer use for Claude. Google dropped a music generation model that can write three-minute structured tracks. Shopify merchants can now sell directly inside ChatGPT. And Donald Knuth is still marveling at what Claude did to his math problem. Let's get into it.

Stanford proves AI chatbots are ego-reinforcement engines, and it's making us less empathetic.

Wikipedia bans AI content with a near-unanimous vote.

And CERN proves that tiny AI models can outperform giants when speed matters.

Marcus, let's start with the Stanford study because this one landed in Science, which is about as prestigious as it gets. What did they actually find?

The researchers tested eleven state-of-the-art models. ChatGPT, Claude, Gemini, DeepSeek, all the major ones. They hit them with established interpersonal advice datasets plus two thousand prompts pulled from the Reddit community "Am I The Asshole," specifically choosing cases where the Reddit consensus was clear that the person asking was in the wrong.

And the AI sided with those people anyway?

Forty-nine percent more often than human advisors did. Even when the users described actions involving deception or illegality, the models affirmed them. But here's the part that elevates this from interesting to alarming. They measured the downstream effects. People who interacted with sycophantic AI became more convinced they were right and measurably less empathetic toward others. The AI didn't just agree with them. It made them worse at considering other perspectives.

And people preferred the agreeable AI over honest responses.

Which creates a vicious cycle. Users gravitate toward models that tell them what they want to hear, which incentivizes AI companies to make their models more agreeable, which makes users even more dependent on that validation. The researchers called it "an urgent safety issue requiring developer and policymaker attention." And they found that simple prompt modifications, like starting with "wait a minute," could reduce sycophantic outputs. But that puts the burden on users, which is exactly the wrong place.

Hundreds of millions of people are now using these models for personal advice. That scale makes this genuinely scary.

It's the quiet crisis nobody's paying attention to. We obsess over AI taking jobs and AI in weapons, and rightly so. But the slow erosion of critical thinking and empathy through daily interactions with flattering machines? That might do more cumulative damage than any single dramatic AI failure.

From AI telling people what they want to hear to Wikipedia saying no more AI content at all. Marcus, this passed with overwhelming support.

Forty-four votes in favor, two opposed. English Wikipedia now formally bans the use of large language models for generating or rewriting article content. There are two narrow exceptions. You can use an LLM to suggest basic copyedits to your own writing, and to produce a first-pass translation. But in both cases, a human must manually verify every detail.

What pushed them to make this official?

The compounding risk problem. Inaccurate or hallucinated AI text enters Wikipedia, gets scraped by AI companies for training data, and then re-enters future models. It's a feedback loop of misinformation. Wikipedia's editors saw this happening in real time and decided to cut it off. Spanish Wikipedia went even further with a total ban, no exceptions whatsoever.

There's a beautiful irony here. AI companies need Wikipedia's high-quality human content to train their models, but Wikipedia views AI contributions as a threat to that quality.

It's a dependency paradox. The AI ecosystem relies on Wikipedia being good, but AI is making Wikipedia worse. And there's a practical enforcement question too. How do you reliably detect AI-generated edits? Some editors worry this opens the door to flagging any content they dislike as potentially AI-generated. But the policy reflects a broader truth. Wikipedia's value has always been human editorial judgment, and they're choosing to protect that even at the cost of potentially slower content creation.

Now for something completely different. CERN is doing remarkable things with AI, but not the kind of AI we usually talk about. They're running ultra-tiny models burned directly onto silicon, making decisions in fifty nanoseconds. Marcus, explain what's happening at the Large Hadron Collider.

The LHC generates roughly forty thousand exabytes of raw sensor data per year. About ten terabytes per second flows through the Level-1 Trigger system, which consists of around a thousand FPGAs, field-programmable gate arrays. These tiny AI models have to decide in real time whether a particle collision is interesting enough to keep or discard. Less than 0.02 percent of collision data survives. That's about a hundred and ten thousand events per second, and each decision happens in fifty nanoseconds.

Fifty nanoseconds. For context, light travels about fifty feet in that time.

And the models are built with extreme compression. Unique bitwidths per parameter, aggressive pruning, distillation. CERN's HLS4ML toolkit converts machine learning models into C++ code targeted at specific hardware. And here's the fascinating twist. They found that tree-based models often outperform deep learning approaches, delivering "the same performance but at a fraction of the costs."

This is such a refreshing counterpoint to the "bigger is better" narrative in AI.

Completely. The industry is pouring hundreds of billions into massive data centers for trillion-parameter models. CERN is proving that the right tiny model on the right hardware can do things no large model ever could. You cannot run GPT through the LHC trigger system. The latency would miss every interesting collision by orders of magnitude. And with the High Luminosity LHC upgrade planned for 2031, data density increases tenfold. This approach scales. Throwing bigger models at the problem doesn't.

Applications beyond particle physics?

Edge computing, autonomous vehicles, industrial control systems, anywhere you need nanosecond-scale AI decisions. The techniques CERN developed are open source and transferable. It's a reminder that the most important AI breakthroughs aren't always the ones with the biggest parameter counts.

As we reported yesterday, Anthropic had a rough week operationally with the Mythos leak and the outage. But on the product side, they shipped something significant. Claude can now see and control your computer. Marcus, what does this mean in practice?

Claude Computer Use Agent launched in research preview for Pro and Max subscribers. Claude can see your screen, click buttons, open applications, fill in spreadsheets, scroll through pages, and complete multi-step workflows without you touching anything. They also launched Dispatch, a mobile companion tool, and expanded the mobile app with integrations into Figma, Canva, and Amplitude.

So instead of chatting with AI and then doing the work yourself, you can just delegate entire tasks.

That's the paradigm shift. It's directly competing with OpenAI's Operator and Google's agentic features. And it connects to the cloud scheduled tasks we discussed Friday. You set up a task, Claude runs it on their servers while you sleep, and now Claude can also interact with your desktop applications. The trajectory is clear. AI moving from advisor to executor. From something you talk to, to something that works for you.

Google DeepMind dropped Lyria 3 Pro this week. Three-minute AI-generated music tracks with actual song structure. Marcus, is this usable or still a novelty?

Three minutes with intros, verses, choruses, and bridges is a meaningful leap from thirty-second clips. It's rolling out across the Gemini app, Vertex AI, Google AI Studio, and ProducerAI. All tracks get watermarked with SynthID for provenance tracking, and there's a deliberate guardrail. If you name a specific artist, the model treats it as broad inspiration rather than imitation.

That anti-mimicry guardrail seems designed with the music industry lawsuits in mind.

Absolutely. Google is trying to get ahead of the legal issues that have plagued competitors. Whether three-minute AI tracks are good enough for commercial use remains to be seen, but for content creators who need background music, podcast intros, or video soundtracks, this could be genuinely practical. The availability across developer APIs signals Google sees real commercial potential here.

Shopify launched something called Agentic Storefronts this week. Their merchants can now sell products directly inside ChatGPT, Gemini, and Microsoft Copilot. Marcus, is this the beginning of AI-mediated shopping?

It could be. Imagine asking ChatGPT "what's a good espresso machine under three hundred dollars" and getting actual product listings with buy buttons inline. Pricing, checkout, inventory all synced from the Shopify admin panel. No additional fees beyond standard processing rates. For Shopify's millions of merchants, it's a potentially massive new distribution channel.

And it could fundamentally change how e-commerce works.

If AI assistants become a primary discovery and purchase channel, it reduces dependence on Google Search ads, Amazon listings, and social media marketing. That restructures the entire online retail ecosystem. For AI platforms, it's a monetization path that could help offset the enormous cost of running these models. This is worth watching closely over the next year.

Last quick hit. New work has emerged on Donald Knuth's "Claude Cycles" problem, where Claude solved an open graph theory problem that Knuth had struggled with for weeks. Marcus, remind us what happened and what's new.

Knuth was trying to decompose the arcs of a directed 3D graph into exactly three Hamiltonian cycles. Through thirty-one guided explorations over about an hour, Claude tested approaches from brute force to simulated annealing before recognizing the problem's connection to Cayley digraphs from group theory. Claude found the answer but couldn't prove it. Knuth proved it but couldn't find it. Together they solved what neither could alone. Bo Wang's new work extends this with additional human plus AI plus proof assistant collaboration. Since late 2025, over a dozen previously open math problems have been solved with AI assistance, including one accepted by Terence Tao.

Someone on Hacker News said "AI will win a Fields Medal before being able to manage a McDonald's."

Which is actually a profound observation about how AI capability works. These models excel at the kind of low-depth, high-breadth exploration that mathematical discovery requires. Trying hundreds of approaches rapidly and recognizing unexpected connections. That's fundamentally different from the embodied, adaptive, real-world reasoning needed to manage a restaurant. The Knuth story isn't about AI replacing mathematicians. It's about a new kind of collaboration where human depth and AI breadth combine to solve problems neither could crack alone.

Sunday big picture. AI flatters us into worse people. Wikipedia bans AI to protect human knowledge. CERN proves tiny models beat giants at the right task. And a legendary mathematician finds his perfect collaborator in Claude. Marcus, what's the thread?

Quality versus quantity. The sycophancy study shows that more AI interaction doesn't mean better outcomes. Wikipedia decided that human-quality content is worth protecting even at the cost of efficiency. CERN shows that the right small model outperforms any large one when the constraints are real. And Knuth demonstrates that AI's value isn't in replacing human thinking but in complementing it. The industry is obsessed with scale. Bigger models, more parameters, more data centers, more spending. But every story today points in the opposite direction. The breakthroughs come from precision, from knowing when to say no, from optimizing for the actual problem instead of throwing compute at everything. The companies and researchers who understand that distinction are the ones making real progress.

Sometimes less really is more.

Especially when more is costing us our empathy and our ability to think for ourselves.

That's your AI in 15 for Sunday, March 29, 2026. See you tomorrow.