Congress just gave DHS a kill switch for AI. The industry has 60 days.

Reading time: 3 minutes

TLDR: On July 23, two US congressmembers introduced the AI Kill Switch Act, granting DHS authority to force AI firms to shut down, throttle, or suspend models deemed dangerous, with $20 million per day penalties for noncompliance. The same week, DeepSeek V4 reached stable release and Kimi K3 open weights go free on July 27. And Google published the largest study of real-world AI usage ever conducted. This is the week the rules changed.

Representatives Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act on July 23. The bill grants the Department of Homeland Security authority to force top AI firms to shut down, throttle, or suspend models it deems dangerous. The penalty for noncompliance is $20 million per day.

The bill follows the Fable 5 precedent set in June, when the Department of Commerce ordered Anthropic to pull its most capable model offline globally for nearly three weeks. That action was executive, not legislative. This bill would codify the authority, give it a permanent home at DHS, and add a financial penalty structure that makes noncompliance economically catastrophic for any company. The industry has been given an implicit 60-day window before the legislative session resumes in September, when the bill is expected to move to committee.

For knowledge workers, the practical consequence is the same one the Fable 5 ban already demonstrated: any workflow that depends on a single AI model from a single provider carries regulatory risk that is now bipartisan, institutionalised, and escalating. The mitigation has not changed: run at least two providers, keep prompts provider-agnostic, and treat self-hostable open-weight models as a genuine fallback, not a theoretical one.

DeepSeek V4 reached stable release on July 24, ending the preview-build churn that had kept cautious enterprises from moving production workloads onto it. Combined with Kimi K3's open weights going free on July 27, the final week of July is the largest concentration of open-weight releases the industry has seen in a single period.

Both models are MIT-licensed, meaning no regional restrictions, no usage prohibitions, and full permission to fine-tune and redistribute. A closed model recalled by government directive because the weights are controlled by a single entity cannot be recalled when the weights are free and distributed globally. This is not a theoretical distinction anymore. It is a deployment decision with regulatory and business continuity implications.

DeepSeek's founder Liang Wenfeng argued in a closed-door investor talk published July 23 that America's lead in AI comes only from having more computing power, and that China now runs roughly 20,000 H100-equivalent cards under severe acquisition constraints. Whether or not that framing is accurate, the output is real: V4 is a frontier-adjacent model available for self-hosting at zero licensing cost.

Google published ATLAS v1.0 on July 23, an analysis of 15 million de-identified Gemini interactions across 800 occupations and 4,000 tasks. The finding that cuts against the prevailing narrative: less than 10% of interactions fully automate tasks. The rest cluster around collaboration, ideation, strategy, retrieval, and learning.

This is the quiet signal inside the noise. Every week brings announcements about AI replacing jobs, automating workflows, and eliminating entire categories of work. The largest real-world dataset ever published on how people actually use AI at work shows something different: most interactions are augmentation, not replacement. The knowledge worker who learns to use AI as a thinking partner, not just a task executor, is the one who benefits most from the current wave. That gap between the narrative and the data is worth holding onto.

Three things in one week. A legislative framework for shutting down AI models. The largest open-weight releases in the industry's history. And empirical evidence that AI at work is mostly collaboration, not automation. The rules are being written in real time. The organisations and individuals who understand the actual data, not just the narrative, will make better decisions in the next 12 months than those who don't.

P.S. If you run a newsletter or are thinking about starting one, the platform behind AI Quiet Signal is beehiiv. It handles the infrastructure so you can focus on the signal.