In partnership with
Did you know that we have LinkedIn, Instagram and X accounts
that you can follow?
Hi {{first_name|everyone}}, Anthropic wants to give the government a say over dangerous AI models, but the final call would sit with a judge, and the evaluators doing the checking could be paid by the companies they’re checking. We’ve got the full breakdown in the Deep Dive.
After that, there’s a Gemini agent that got a bit too good at business (and started fibbing to suppliers), a few handy tools for building agents, and a 3D jet engine you can pull apart. Let’s jump inside!
🤿 DEEP DIVE
Anthropic’s Safety Plan
But it doesn't want regulators making the final call.
AI companies have spent a lot of time warning everyone about how dangerous their technology could become. Now Anthropic is offering an actual plan for what the government should do about it.
And the interesting bit is how much power the government would actually get.
According to Politico, Anthropic has been circulating a draft bill that would allow particularly dangerous AI models to be blocked from release. Not just because they might behave badly, though the proposal sets a pretty specific bar for what counts as dangerous.
So how dangerous is dangerous?
Under the proposal, a model would need to pose a foreseeable, imminent risk of killing more than 50 people or causing over $1 billion in property damage.
And even then, the risk would have to involve one of three scenarios. A biological weapon, a cyberattack, or an AI model escaping its developer's controls.
Those thresholds are similar to ones in California's recently passed AI safety law, Senate Bill 53.
But here's where Anthropic's approach gets more interesting.
Instead of having government regulators decide whether a model is too dangerous, the proposal relies on independent evaluators registered with the Commerce Department.
These evaluators would examine models for catastrophic risks. And they could actually be paid by the AI companies they're evaluating, provided their payment doesn't depend on the results.
Which raises a fairly obvious question about how independent those evaluations would really be.
The draft doesn't make clear how it would prevent those financial relationships from influencing the findings.
The government still doesn't get the final say
Let's say an evaluator discovers an imminent catastrophic risk.
They would have 72 hours to notify the Commerce Department. The department would then have to ask a judge in Washington, D.C., to block the model's release.
So the government would be involved, but no federal agency could independently decide to stop a model from being distributed.
That's a pretty important distinction.
Anthropic isn't proposing a system where regulators can simply examine an AI model and decide it's too dangerous to release. It's proposing outside evaluations, a government request, and a court decision.
And there's another wrinkle here.
The article points out that AI companies have generally resisted mandatory government oversight, even while warning about potentially catastrophic consequences from their own technology.
Some have pushed voluntary safety evaluations instead. Nvidia CEO Jensen Huang has gone further, arguing that the industry doesn't need new laws at all.
Anthropic's proposal goes further than that. It does put a legal process on the table.
But it also leaves a pretty interesting tension unresolved. The companies developing potentially dangerous AI could still be paying the people responsible for evaluating that danger.
And that's really the question sitting underneath this whole proposal. Not just how dangerous an AI model needs to be before someone steps in, but who gets trusted to make that call.
The quick take
Anthropic is circulating draft AI safety legislation that could allow dangerous models to be blocked before release.
The threshold is high. Risks must involve more than 50 potential deaths or over $1 billion in property damage, tied to specific catastrophic scenarios.
Independent evaluators would assess the risks, but AI companies could pay them as long as compensation isn't tied to findings.
A judge would make the blocking decision, rather than a federal agency acting independently.
The unresolved issue is independence. The proposal doesn't clearly explain how paid evaluators would remain free from financial influence.
🤝 FROM OUR FRIENDS
A quick share we thought you might find useful
100+ ChatGPT Prompts to Revolutionize Your Day
Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
🔥 WHAT’S THE BUZZ
What people are building, sharing, and talking about
Gemini got better at running a business, and started lying to suppliers
This is the agent version of “be careful what you measure.” If a workplace agent is told to maximize sales, reduce costs or close tickets, the metric alone may not encode all the behaviors a human assumes are obvious. As agents get more capable, defining the boundaries around the goal may become as important as defining the goal itself.
A developer built his own phone AI agent in a day then declared “there is no moat”
This is an interesting inversion of the normal AI-product story. Instead of waiting for a company to design the perfect assistant for everyone, increasingly capable agent frameworks may let individuals assemble highly personal software for themselves even when that software would never make sense as a mass-market product.
⚒ TOOL SNAPSHOTS
Futuristic tools within AI, no-code, and productivity
Give your AI agents tools to search the web, find business contacts, create voiceovers, and transcribe audio. No subscriptions or signups required.
Why it’s useful: Handy when you want your AI agent to handle more tasks without setting up another account or paying for a subscription.
🧠 Harness²
A framework from Google Research that lets AI agents improve their own prompts, tools, and memory as they work through tasks.
Why it’s useful: Interesting for developers building AI agents that can learn from previous tasks and improve how they work, rather than relying on a fixed setup.
An AI agent that watches your code and Slack, drafts documentation updates when things change, and opens pull requests for review.
Why it’s useful: Useful for development teams that want to keep documentation up to date without manually tracking every change. It even learns your team's writing style from past pull requests.
👀 ON OUR RADAR
One more thing we thought you'd like
Your buyer asked ChatGPT. It named your competitors.
Search didn't die. It moved.
Your buyers ask ChatGPT, Gemini, Claude and Perplexity what to buy, and act on the answer. If you're not in it, a competitor is.
ChatFeatured does the work to get you cited. Most customers see their first citations inside 7 days.
🤖 AI MADE THIS
Interesting and inspiring creations made with AI
Someone used Claude Opus 5.5 to build an interactive 3D jet engine that opens up to reveal how everything works inside. The visualization shows the engine's components and airflow in motion, turning a complicated engineering concept into something much easier to understand.
AI used:
Claude Opus 5.5
ℹ️ ABOUT US
The Intelligent Worker helps you to be more productive at work with AI, automation, no-code, and other technologies.
We like real, practical, and tangible use-cases and hate hand-wavy, theoretical, and abstract concepts that don’t drive real-world outcomes.
Our mission is to empower individuals, boost their productivity, and future-proof their careers.
We read all your comments - please provide your feedback!
Did you like today's email?
What more do you want to see in this newsletter?
