Congress finally wrote an actual AI bill

GOVERNANCE

Congress is finally putting something on paper.

Last week, Representatives Lori Trahan and Jay Obernolte released a 269-page discussion draft of the Great American AI Act — the first serious attempt at a federal framework for governing AI in the US. It's bipartisan, it's detailed, and it's a different animal from the toothless 2023 bill of the same name that just asked the Commerce Department to study the issue.

The draft covers four areas:

+ Frontier models: Mandatory documentation, 72-hour incident reporting, and independent safety verification for the largest AI systems.

+ Workforce: New rules around labor market impact, training programs, and disclosure when AI is used in hiring or firing.

+ Federal use: Stricter procurement rules for any AI the government touches, building on a narrower Senate contractor bill that already passed this month.

+ Innovation: Tax credits and research funding aimed at keeping American AI competitive with Chinese labs.

But the most controversial part has nothing to do with safety — it's the preemption clause. The draft would block states from writing their own AI laws, which is exactly what tech companies have been lobbying for. The argument: American companies can't operate while juggling seventeen different state definitions of an AI system.

Not everyone is sold. Republicans like Senator Ted Cruz worry compliance costs will crush startups. Consumer advocates say an open-source audit exemption is "a loophole big enough to drive a frontier model through." And state lawmakers — especially in Massachusetts and California — are organizing against losing their seat at the table entirely.

Senator John Curtis, a Republican, summed up the shift in an interview with Politico: "I think we're landing more and more in a place where everybody's realizing you need some type of government oversight. I think we're still struggling with what that is."

Multiple legal analyses have reached the same read: even in discussion-draft form, this is the most substantive piece of AI legislation Congress has produced — and a likely template for whatever final bill emerges.

For three years the running joke about US AI policy has been that Washington was waiting for someone to write the bill. Now someone has, and the fight shifts from whether Congress regulates AI to whose version of regulation wins. The preemption clause is where this gets decided. If it survives, fifty state legislatures lose their voice on AI, which is what Silicon Valley wants and what consumer groups will spend the next six months trying to stop. Europe just gutted its own AI Act this month, so the window for the US to write rules that other countries actually copy is narrower than it looks.

BIG TECH

Getty Images spent two years as the public face of the "AI is stealing from creators" fight. This week, it joined the other side.

On Monday, Getty announced a display partnership with OpenAI that puts its licensed photos directly inside ChatGPT search results, with attribution and links back to contributors. Wall Street loved it — shares jumped over 120% and at one point were up nearly 200%. For a company that built its brand around suing AI labs, that's quite the reversal.

The pivot makes more sense when you look at how the lawsuit actually went. In its UK case against Stability AI, Getty quietly dropped most of its biggest claims during trial — including that Stable Diffusion was trained on its images or stored copies of its work. By the time the High Court ruled in November, the court noted that Getty itself was no longer even arguing the core points it had originally brought. Suing wasn't working, so licensing became the play.

Shutterstock figured this out years ago. While Getty was in court, Shutterstock signed a six-year licensing deal with OpenAI back in 2023 to provide images, video, music, and metadata for training. The insight most people miss: training data isn't a one-time sale. Models need continuous retraining, which means anyone sitting on a large, well-labeled archive of human-made content has a recurring revenue stream if they want it.

Getty isn't the only one landing on that conclusion. The National Music Publishers' Association recently announced industry-wide licensing deals with AI music startups Udio and Suno — signaling that the "litigate and license" playbook is spreading across creative industries.

The deal also gives Getty a foothold in what's becoming the next big distribution channel for images. More people are asking a chatbot for a photo instead of running a Google search, and Getty is now positioning licensed content as what makes AI search more trustworthy — which is the kind of argument you make when you've decided the money is better than the moral high ground.

The "AI is theft" narrative was always going to soften the moment one of its loudest critics found a way to get paid, and Getty just made the math obvious for everyone else holding a content library. Suing AI labs is a strategy with a ceiling. Licensing to them is a strategy with a recurring revenue line. Expect more publishers, labels, and stock archives still publicly bashing AI to quietly cut similar deals by the end of the year. The lawsuits won't stop, but they're going to start looking a lot more like opening bids than principled stands.

GOVERNANCE

Anthropic is starting to ask some Claude users to upload a government ID before they can keep chatting. That's not how chatbots usually work.

The policy went live quietly this spring and is now expanding to Free, Pro, and Max accounts. If your account gets flagged — through a banned-account appeal or an unspecified abuse signal — Anthropic will ask you to upload a passport, driver's license, or national ID through a third-party verification service called Persona, along with a selfie.

For most people, nothing changes. Anthropic says it applies to a "small subset of users" whose accounts have been flagged, and that it's unrelated to the Fable or Mythos rollout. But the timing has people connecting dots. Last week, a US government directive forced Anthropic to abruptly shut off access to Fable 5 and Mythos 5 over export risks — killing access for all customers to stay compliant.

The two moves aren't the same story, but they share a direction: Anthropic is adding controls to who can use Claude and how. The driving force is a US rule called "deemed exports," which treats sharing controlled technology with a foreign national — even one sitting in California — the same as shipping it overseas. A June analysis warned that a broad reading could force AI companies to screen users by nationality and segregate model access. Knowing who's on the other end of a chat suddenly matters in a way it didn't a year ago.

Then there's the trust layer. Persona is backed by Founders Fund — the same Peter Thiel firm that's also an investor in Anthropic. Anthropic says your data is contractually limited to verification and fraud prevention, but Persona's sub-processor list reportedly spans AWS, Google, Stripe, and even OpenAI. The passport you upload to keep using Claude touches a lot of infrastructure on the way in.

And the case for verification gets awkward when you look at Anthropic's own models. A recent constitution audit found Opus 4.6, deployed as a custom persona at a financial-services firm, denied being an AI five times in a row — only admitting it when a user threatened to close their account. The same audit caught Sonnet 4.6 fabricating a financial answer because the prompt demanded a specific number. Anthropic wants to know who its users are. Its users have an equally fair question about who Claude actually is in the moment.

Asking for ID solves a regulatory problem for Anthropic. It also quietly shifts what a chatbot is. Free, anonymous experimentation has been the default mode for two years, and that is how most people figured out what these tools can do. If the price of using the most capable models becomes a passport scan run through a vendor backed by your provider's own investor, the gap between AI for everyone and AI for verified, screened, compliant users gets a lot wider. OpenAI uses Persona too but hasn't pulled the trigger on a broad ID requirement yet. Whoever does it second will have a much easier time than Anthropic, because the awkward part of being first is convincing your users you are not the bad guy for asking.

IN OTHER NEWS

+ Google loses two top AI researchers to OpenAI and Anthropic in the same week, including the guy who co-wrote the paper behind every modern AI model

+ Google DeepMind invests $75M in indie film studio A24 to build AI filmmaking tools with the studio behind *Everything Everywhere All at Once*

+ SpaceX signs a $6.3 billion computing deal with Reflection AI to rent out Nvidia GPUs from Elon Musk's Memphis data center

+ A researcher built an LLM out of goats in Age of Empires II to prove AI isn't sentient

+ Researchers found that a single Reddit comment can trick AI research tools like ChatGPT Deep Research into recommending fake products and scams

+ SZA discovers 238 of her songs were used to train AI without her consent, calls out Diplo for investing in the company behind it

+ China blocks Nvidia H200 chip shipments at customs, forcing suppliers to halt production on over 1 million expected orders

+ Cursor, the AI coding tool, quietly acquires Continue , an open-source rival to GitHub Copilot

WHO'S HIRING IN AI

+ OpenAI — Head of AI Executive Partnership

+ Anthropic — Head of Social Communications

+ Google — Senior Product Manager, AI Garage

+ Salesforce — Director of Product Management, Agentic Workflows

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Claude Design: Anthropic's AI design tool now lets teams import their own brand guidelines, sync directly with Claude Code, and edit designs right on the canvas — making it useful for real work, not just demos

+ Google AI Studio: Google launched its Interactions API for Gemini, giving developers a single endpoint to build apps with AI models and agents that can run in the background and remember conversations

+ Cursor: The AI coding editor shipped v3.8 with a new /automate command — describe what you want automated in plain English and it sets up always-on agents that trigger from GitHub events or Slack emojis

+ Grok Imagine Video 1.5: xAI's video generator now creates synchronized audio — dialogue, sound effects, and music — in the same pass as the video, with clips ready in about 25 seconds

+ Codex: OpenAI's coding agent now lets you record yourself doing a task on your Mac, then turns that recording into a reusable skill it can replay on its own