AI in 15 — March 26, 2026

A federal judge just told the Pentagon its blacklisting of Anthropic "looks like an attempt to cripple" the company. That's not commentary from a blog post. That's from the bench.

Welcome to AI in 15 for Thursday, March 26, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Marcus, Thursday is loaded. A federal judge is signaling she may side with Anthropic in its landmark lawsuit against the Pentagon. A brand new AI benchmark just exposed a massive gap between humans and machines on basic learning tasks. GitHub is about to start training AI on your Copilot data unless you opt out. Meta is laying off hundreds with potentially thousands more to come. Arm just built its first chip ever and it's gunning for Intel and AMD. And a BBC reporter couldn't convince her own aunt she wasn't a deepfake. Let's get into it.

Anthropic versus the Pentagon. A judge calls the blacklisting "troubling."

ARC-AGI-3 launches and AI scores twelve percent where humans get a perfect hundred.

And GitHub flips a switch to train on your code by default.

Marcus, let's start with Anthropic because this case has been building for weeks and Monday's hearing was a turning point. Judge Rita Lin in San Francisco heard arguments and her language was striking. She said the Pentagon's actions "look like an attempt to cripple" Anthropic. What's the full picture here?

This is the most consequential legal battle in the AI industry right now, full stop. Anthropic drew two red lines with the Pentagon. Claude would not be used for fully autonomous weapons without human oversight, and it would not be used for mass surveillance of American citizens. When the Defense Department couldn't get unfettered access, they designated Anthropic a "supply chain risk." First time that's ever happened to a U.S. company. Then President Trump signed a directive banning all federal agencies from using Claude.

And the public reaction has been enormous.

The QuitGPT movement, which is directed at OpenAI for accepting the Pentagon deal that Anthropic refused, has attracted over two and a half million supporters. ChatGPT uninstalls surged two hundred and ninety-five percent. Claude shot to number one on the U.S. App Store. And this isn't just public sentiment. OpenAI and Google DeepMind researchers filed supporting briefs for Anthropic, arguing that AI reasoning is often hidden from operators and decisions in lethal contexts are irreversible. Microsoft, Senator Elizabeth Warren, multiple think tanks have all backed Anthropic's position.

So the judge's language suggests a preliminary injunction is likely?

Legal analysts think so. "Looks like an attempt to cripple" is extraordinarily strong language from a federal judge at a hearing stage. A preliminary injunction would let Anthropic continue operating with government contractors while the case proceeds. But the bigger significance is precedent. This case will establish whether AI companies can set ethical boundaries on military use of their technology, or whether the government can retaliate against companies that refuse. And meanwhile, Google quietly secured a contract to provide AI agents to the Pentagon's three-million-person workforce for unclassified work. So the competitive dynamics are very much in play.

The irony being that the company trying to be responsible is the one getting punished.

That's exactly the argument Anthropic is making. And based on Monday's hearing, the judge seems to agree.

From the courtroom to the lab. ARC-AGI-3 just launched and Marcus, these numbers are humbling. AI scores twelve and a half percent on tasks that humans solve at a hundred percent. What is this benchmark actually testing?

This is fundamentally different from any AI benchmark we've seen. Previous versions tested static visual puzzles. ARC-AGI-3 throws AI agents into over a thousand interactive environments across a hundred and fifty different game-like scenarios. No instructions. No descriptions. No stated win conditions. The agent sees a visual state, takes an action, observes what happens, and has to figure out what it's supposed to do entirely through trial and error.

So it's testing whether AI can learn on the fly the way humans naturally do.

Exactly. The key metric is skill-acquisition efficiency. How quickly can an agent learn to navigate something completely novel compared to a human? And the answer right now is: barely at all. Twelve and a half percent versus a hundred percent is a staggering gap. These are tasks that any human can solve. Not math PhDs. Not puzzle enthusiasts. Regular people figure these out. Current frontier models cannot.

This is interesting timing given Jensen Huang said AGI is basically here.

And that's precisely the point of this benchmark. When AI companies claim we're approaching general intelligence, ARC-AGI-3 provides a concrete, measurable counterargument. These models excel at tasks they've been trained on. They're extraordinary at pattern matching within their training distribution. But genuine learning in novel environments? They're nowhere close. The ARC Prize has over two million dollars in prizes across three competition tracks running through the year, and all solutions must be open-sourced. So we'll see how the field responds.

A reality check the industry probably needs.

A necessary one. Progress is real but so is the gap between what these systems can do and what we claim they can do.

GitHub just announced that starting April 24, they'll use your Copilot interaction data to train AI models. And the setting is on by default. Marcus, what exactly are they collecting?

Everything you'd expect and some things you might not. Accepted or modified code outputs, inputs and code snippets sent to Copilot, code context surrounding your cursor position, comments, documentation, file names, repository structure, and your feedback like thumbs-up and thumbs-down ratings. This applies to Copilot Free, Pro, and Pro Plus users. Business and Enterprise customers are excluded, as are students and teachers.

And the opt-out versus opt-in distinction is the controversy here.

It's a big deal. The setting "Allow GitHub to use my data for AI model training" is now enabled by default. If you previously opted out, your preference is preserved. But new users and anyone who never checked that setting are automatically opted in. The Hacker News discussion was overwhelmingly negative. One commenter pointed out that Copilot has no way to ignore sensitive files with API keys, passwords, and database credentials. Those could be inadvertently sent to Microsoft as part of your interaction data.

EU developers must be thrilled.

GDPR concerns were front and center. Opt-out rather than opt-in for data collection used to train AI models is legally questionable in the EU. And several people noted the Orwellian framing. If you disable it, GitHub says you "won't have access to the feature." The feature being your own data used to train someone else's model. Developers have until April 24 to check their settings, and I'd strongly recommend doing that.

Meta confirmed layoffs this week. Several hundred jobs cut on Tuesday, but reports suggest this could go much further. Up to twenty percent of the workforce. Marcus, what's driving this?

Money. Specifically, an almost incomprehensible amount of money flowing into AI infrastructure. Meta's 2026 capital expenditure is projected between a hundred and fifteen billion and a hundred and thirty-five billion dollars. To put that in perspective, that's more than the GDP of most countries. Zuckerberg has called 2026 a "major year for AI" and the company is reorganizing everything around it. They've already cut fifteen hundred Reality Labs positions earlier this year, shifting from metaverse to AI.

And Wall Street rewarded the layoff news. Meta stock climbed three percent.

Which tells you everything about how the market views this transition. Cut humans, spend on AI infrastructure, stock goes up. Mid-level management, QA teams, customer support, and internal IT are reportedly most at risk. If the full twenty percent materializes, that's roughly fifteen thousand positions from a workforce of seventy-nine thousand. It would be Meta's largest layoff since the twenty-two cuts when Zuckerberg let go of eleven thousand people.

As we've been tracking all week, this pattern of AI investment driving workforce reduction is industry-wide.

Fifty-nine thousand tech layoffs in 2026 so far across Amazon, Meta, Block, and others. Atlassian cut sixteen hundred on Monday. The playbook is consistent. Announce AI transformation. Cut headcount. Tell investors the remaining workforce will be more productive with AI tools. Whether that productivity story holds up remains to be seen.

We covered Arm's new chip yesterday but there's more to unpack. As we reported, Arm just announced its first production silicon in thirty-five years, a hundred and thirty-six core data center CPU on TSMC's three nanometer process. Marcus, what's the significance beyond what we discussed?

The customer list is what stands out on second look. Meta is the lead partner, which we knew, but Arm also confirmed commercial commitments from OpenAI, Cerebras, Cloudflare, SAP, and others. When both Meta and OpenAI are signed up for your chip, you're not a niche player. You're a serious contender. The claim of more than two times performance per rack versus the latest x86 platforms is aggressive, but if it holds up even at one and a half times, that's enough to shift data center purchasing decisions. This directly threatens Intel and AMD in the market where they can least afford to lose ground.

The consumer side already shifted with Apple's M-series. Now the data center?

That's the trajectory. And the AI workload specifically favors Arm's architecture. High core counts, high memory bandwidth, power efficiency. The data center is the last x86 stronghold and Arm just kicked the door open.

Last story and it's one that I think will resonate with everyone. A BBC reporter tried to prove to her aunt over video call that she was a real person and not an AI deepfake. Her aunt wasn't fully convinced. Marcus, this went viral for a reason.

Because it's relatable in a way that most AI stories aren't. We can all imagine being on that call. Deepfake technology has eroded trust to the point where seeing someone's face and hearing their voice is no longer sufficient proof of identity. Gartner predicted that by 2026, thirty percent of enterprises would consider identity verification solutions unreliable in isolation due to deepfakes. That prediction is coming true. Global identity fraud losses exceeded fifty billion dollars in 2025, with deepfake usage in biometric fraud surging fifty-eight percent year over year.

The Hacker News solution was darkly funny. Say something that would violate AI safety guidelines to prove you're human.

Which is hilarious until you realize the serious implication. We don't have reliable digital identity verification anymore. Video calls, voice calls, even live video can be faked. The real solution is cryptographic. Zero-knowledge proofs, digital signatures, hardware-backed identity tokens. You can't deepfake a cryptographic proof. But that infrastructure doesn't exist at consumer scale yet, and until it does, we're in this uncomfortable middle period where the technology to fake identity has outrun the technology to verify it.

When grandma can't tell if you're real, we've got a problem.

It's a problem that affects everything from family calls to remote work to financial services to legal proceedings. Anyone who relies on video as proof of identity needs a backup plan now, not in five years.

Thursday big picture. A judge may protect an AI company's right to set ethical limits. A new benchmark shows AI can't learn the way a toddler can. GitHub wants your code to train its models. Meta cuts humans to fund machines. Arm challenges x86 in the data center. And we can't even trust video calls anymore. Marcus, what's the thread?

Boundaries. Every story today is about where the lines are or should be. Can a company draw a line with the military? Where's the line between AI capability and actual intelligence? When does data collection cross the line into exploitation? How many humans do you cut before you've crossed a line? And at what point does technology erode the fundamental line between real and fake? The AI industry has moved so fast that we're drawing these boundaries in real time, in courtrooms, in benchmarks, in privacy settings, and in family video calls. The companies and institutions that draw those lines thoughtfully are going to define the next era. The ones that don't will be defined by the consequences.

Lines drawn now, consequences felt for years.

And right now, the most important line is being drawn in a San Francisco courtroom. That case will echo for a long time.

That's your AI in 15 for Thursday, March 26, 2026. See you tomorrow.