> daily_signal(2026_05_13)

Colorado just gutted its own AI accountability law before it could take effect. The federal page documenting safety reviews of three major AI labs disappeared the same week.

PickBits Daily Signal · Wednesday, May 13, 2026

By Mark Pickering · 7 min read · May 13, 2026

// tl;dr

Colorado lawmakers passed SB 26-189 on May 9 stripping the bias-audit and algorithmic-discrimination duty from the 2024 AI Act before it could take effect. The new bill replaces "high-risk AI system" rules with a documentation-and-notice ADMT framework that takes effect January 1, 2027. House vote was 57-6, Senate 34-1. Governor Jared Polis has said he plans to sign.
The Commerce Department page documenting pre-release AI safety reviews of Google, Microsoft, and xAI was pulled from the web by Monday afternoon. The May 5 announcement said the three companies would hand over their frontier AI systems for cyberattack-vulnerability, military-misuse, and national-security review before public release. The page now redirects to a generic homepage. Commerce and the White House would not explain why.
Hill County, Texas voted 3-2 Tuesday for a one-year moratorium on new data center construction in unincorporated areas - the first county-level data center moratorium in Texas. County Attorney David Holmes warned commissioners before the vote that Texas does not give counties moratorium authority and they could be sued for it. Cave City, Kentucky passed its own one-year moratorium 4-1 the same Monday.
White House economic adviser Kevin Hassett told CNBC on Monday that "there's no sign in the data that AI is costing anybody their job right now." The same day he confirmed the administration will not create a new federal agency to oversee AI and is drafting a voluntary executive order that does not require mandatory model testing. Tech layoffs YTD have hit roughly 92,000, with Coinbase, Cloudflare, Upwork, BILL, GitLab, and General Motors all naming AI in their public reasoning.
OpenAI shipped GPT-5.3-Codex-Spark on Cerebras chips this week - its first production AI tool that does not run on NVIDIA hardware. The Information reported OpenAI hired Gimlet Labs to optimize the models for Cerebras silicon. Gimlet says its optimizer accelerates inference up to 10x at the same cost and power.

The week started with formal AI oversight visible in four places: a Colorado state law about to take effect, a federal Commerce page listing three named pre-release safety reviews, a Texas Legislature stance against county-level moratoriums, and NVIDIA's grip on every production AI inference pipeline. By Tuesday afternoon, all four had been changed or removed. Colorado lawmakers stripped the bias-audit rule from the 2024 AI Act. The Commerce page got replaced with a redirect. A Hill County, Texas commissioners court passed a moratorium anyway. OpenAI shipped its first production model off NVIDIA. None of those moves was decisive on its own. Together they answer the same question: when AI is being put inside the loan decision, the hiring decision, the cyber-attack tool, the data center near your house, where does the override live?

Colorado removed the accountability rule. Commerce removed the federal review page. A Texas county tried to remove a data center anyway. OpenAI removed NVIDIA from a production pipeline. The White House removed the framing that AI is removing jobs.

1. Colorado just removed the rule that would have made AI explain why it rejected your loan.

On May 9, 2026, the Colorado General Assembly passed SB 26-189 on the second-to-last day of its 120-day session. The House voted 57-6. The Senate voted 34-1. Governor Jared Polis, per Bloomberg Law and the Colorado Sun, has confirmed he plans to sign. The bill repeals and reenacts the 2024 SB 24-205 framework that was scheduled to take effect in June - the only US state law that required AI developers and deployers to perform bias audits, document algorithmic-discrimination risk, and exercise a duty of "reasonable care" to avoid discriminatory outcomes when AI was used in a "consequential decision" (loan, hiring, housing, healthcare, insurance, government services). The new framework, called automated decision-making technology (ADMT), takes effect January 1, 2027. It keeps disclosure and notice obligations. It removes the duty of reasonable care, the algorithmic-discrimination-mitigation requirement, and the structured risk-management program.

This is the bookend on the story PickBits Daily led on May 6. A week ago the Department of Justice filed an amicus on Elon Musk's side arguing the Colorado law unconstitutionally burdened interstate commerce, and Colorado's own legislators introduced SB 26-189 to gut it. As of Saturday, the bill had cleared both chambers with near-unanimous Republican support and enough Democrats to pass. The sponsor told the Colorado Sun that "massive amounts of money" from industry made the original law impossible to defend. The 2024 framework that Colorado was the only US state to enact never took effect. Connecticut SB 5, awaiting Governor Lamont's signature, is now the only comparable comprehensive state-level AI accountability law moving forward in the country.

Colorado Sun coverage of SB 26-189 passing the Colorado legislature — coloradosun.com · May 12, 2026

Why this matters: If a credit algorithm denies your loan, an automated hiring system rejects your application, or an insurance carrier raises your premium based on a score you cannot see, the only US state-level rule that would have required the company to explain how it happened was Colorado's. That rule was supposed to take effect this June. It will not. The new bill keeps notice ("AI was used") and disclosure ("here are the categories of data") but removes the requirement that the company actually test for and mitigate discrimination. Action this week: if you live in Colorado, go to leg.colorado.gov/bills/sb26-189 and read what is now keep-vs.-cut in the ADMT framework; the consumer-rights section (right to opt out of profiling, right to a human review) is the part that survives, and it is what you will use if a decision goes against you. If you work in IT or compliance for a company that does business in Colorado, the bias-audit playbook the legal team built for the June 2026 deadline is now optional - which means the budget for it is the first line a CFO will cut. Pull the audit plan and have a position on whether to run it anyway before someone above you decides for you. If you're outside Colorado, check whether your state's pending AI bill (Connecticut SB 5, Texas HB 149, New York S-7522, California SB 53) still includes a duty of care - the National Conference of State Legislatures tracks state AI bills at ncsl.org/technology-and-communication/artificial-intelligence-2025-legislation.

https://coloradosun.com/2026/05/12/colorado-ai-law-rewrite-passes/

2. The federal page documenting safety reviews of three major AI labs disappeared from the web this week. No one will say why.

On May 5, 2026, the Commerce Department announced that its Center for AI Standards and Innovation (CAISI), the renamed AI Safety Institute housed at NIST, had signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI. Per the original announcement (mirrored by The Washington Post, CNBC, CNN, and Cybersecurity Dive), the three companies committed to share unreleased versions of their frontier AI systems with federal testers - sometimes with built-in safety guardrails pulled back - so government scientists could evaluate cyberattack vulnerability, weapons-of-mass-destruction misuse risk, and national-security flaws before public release. OpenAI and Anthropic were already in the program. The page also confirmed the testing happens inside classified environments.

By Monday afternoon, May 11, the page was gone. The original URL returned "Sorry, we cannot find that page" and was then redirected to the generic CAISI homepage, per Quartz, Bloomberg, and The Next Web. Neither Commerce nor the White House offered an explanation. The same Monday on CNBC, NEC Director Kevin Hassett said the administration is "studying" an executive order that would create an AI-vetting system "like an FDA drug" review - but explicitly said it will not require mandatory pre-release model testing and will not stand up a new oversight agency. Bloomberg on May 8 had previewed a draft of that EO and noted that mandatory model tests were omitted from the draft.

Quartz coverage of the US Commerce Department deleting its AI security testing page for Google, Microsoft, and xAI — qz.com · May 12, 2026

Why this matters: If you use ChatGPT, Gemini, Copilot, or Grok, the pre-release federal evaluation of those models was the only outside review of what those systems can do (and how dangerous what they can do is) before they hit your laptop. The page documenting it disappeared, the executive order being drafted to replace it explicitly leaves mandatory tests out, and the White House said Monday it will not stand up an agency to enforce anything. The system is moving from "tested before release" to "voluntary self-reporting" without anyone announcing it. Action this week: save the original CAISI announcement via web.archive.org (search "CAISI" and the commerce.gov URL) - the original press text is still on the Wayback Machine and is the only proof of what the agreements covered. If you work in security or cyber-risk at any company, the loss of the federal pre-release evaluation is something your incident-response plan should account for - the assumption that "the government will catch the model misuse before it ships" is the assumption that just got removed. Open caisi.nist.gov and find the residual public-facing material; that is the new ceiling on what federal AI safety oversight looks like.

https://qz.com/commerce-department-deletes-ai-security-testing-google-microsoft-xai-051226

3. A Texas county became the first in the state to try stopping a data center. Its own attorney warned them they would be sued for it.

On Tuesday, May 12, Hill County, Texas commissioners voted 3-2 to pass a one-year moratorium on new data center construction in the county's unincorporated areas, per the Texas Tribune, KXXV, and KWTX. Hill County is roughly 55 miles south of Fort Worth and was being looked at by Provident Data Centers for a 300-acre development in north Hillsboro. Before the vote, County Attorney David Holmes warned the commissioners that Texas state law gives moratorium authority to cities and municipalities, not counties, and they could be sued for passing this. They passed it anyway. Per the Tribune, this is the first Texas county to enact a data center moratorium.

The same Monday, Cave City, Kentucky city council voted 4-1 for a one-year moratorium of its own (per the Bowling Green Daily News). A Cave City resident told the council that "Large-scale data centers may promise economic development, but those promises often come at significant costs to communities like ours." Texas leads the nation in data-center construction. The Hill County vote does not stop a build in progress; the developer can sue, can ask the legislature for a preemption, can wait it out. What changed is the answer to whether a Texas county can take the position. One did.

Texas Tribune coverage of Hill County passing a data center construction moratorium — texastribune.org · May 12, 2026

Why this matters: If you live in a state that does not normally let counties block development, Hill County just tested whether the override exists anyway. The Saline (Michigan) moratorium that started this arc was passed by a township the developer sued and won; Hill County's moratorium starts the same fight from a different jurisdictional position. The point is that local governments are now willing to take the legal risk of passing the rule and forcing the developer to challenge it - which buys the community a year of construction pause while the lawsuit runs. Action this week: go to datacenters.ainowinstitute.org/local and look for any active data center permitting in your county; the AI Now Institute tracks proposed and active sites in a public database. If you're in local government, the Texas Association of Counties and the Michigan Township Association both have moratorium templates on file - the Hill County order itself is short (one year, unincorporated areas only) and is the cleanest template if your jurisdiction wants to try the same move. If you are a developer or own land near one of these sites, the property-value question is real and will not be resolved on a single news cycle - read the Hill County order text before the next zoning vote your county holds.

https://www.texastribune.org/2026/05/12/texas-hill-county-approves-data-center-construction-pause-ai/

Tomorrow's signal lands here. Subscribe to PickBits Daily Signal if you want the daily that leads with what changed for ordinary people, not the headline number.

4. The White House's top economic adviser said Monday there's no sign in the data that AI is costing people their jobs. Tech layoffs hit 92,000 the same year.

On Monday, May 11, NEC Director Kevin Hassett told CNBC that "there's no sign in the data that AI is costing anybody their job right now," and that companies adopting AI "tend to see rapid revenue growth and even employment growth." The same morning Bloomberg reported Hassett confirmed the administration will not create a new federal agency to oversee AI, calling the idea "a giant new bureaucracy" they do not need. Hassett did mention an active White House taskforce studying "the future of AI and what it means for the workforce."

The named-AI layoff record from the past 30 days, per CNBC, TechCrunch, Fortune, CBS News, and the Programs.com tracker: Coinbase cut about 700 employees (14 percent) on May 5, with CEO Brian Armstrong stating the restructuring was designed to make the company "AI-native" with no pure managers and AI agents replacing engineer-designer-PM workflows. Cloudflare cut 1,100 jobs (about 20 percent) in early May citing a 600 percent internal AI-usage increase. Upwork cut roughly 25 percent. BILL cut up to 30 percent. GitLab opened a voluntary-separation window on May 11 framed as "not AI" but accompanied by a promise to reinvest savings in AI. General Motors cut roughly 500-600 salaried IT workers on May 11 and said the same department is still hiring - for AI skills. Year-to-date tech layoffs, per the tech-insider.org tracker, are above 92,000.

CNBC coverage of Kevin Hassett saying there is no sign AI is costing people their jobs — cnbc.com · May 11, 2026

Why this matters: If you lost a tech job in 2026, or your spouse or child did, the federal position on whether AI did it is now public and disputed: the White House says it did not, the layoff-tracker data and the CEOs who fired you say it did. The practical effect is on unemployment-benefits eligibility (some states ask whether AI was a named cause), on COBRA timing decisions, and on whether the displaced-worker re-training programs your state runs actually open to you. Action this week: if you were laid off in 2026, open layoffs.fyi (free) and search your former employer; if the public filings or CEO emails name AI as a reason, save them as PDFs - state workforce programs increasingly require named-cause documentation for retraining grants. If you manage a team, take an honest count of which of your roles a coding agent or a customer-support agent could plausibly cover today; the question is not whether you fire those people, the question is whether you can explain to them what changed. If you work in HR or workforce policy, the gap between the White House's "no data" framing and the Programs.com / tech-insider tracker is the gap your CEO will quote in a board meeting; pull both numbers and have your own position ready.

https://www.cnbc.com/2026/05/11/trump-white-house-hassett-ai-jobs.html

5. OpenAI just shipped its first production AI tool that does not run on NVIDIA chips.

Continuing the loop from 5/10 #4 - NVIDIA's grip on every production AI pipeline. This week OpenAI launched GPT-5.3-Codex-Spark, a coding model, running in production on Cerebras wafer-scale chips, per Tom's Hardware, TechRepublic, and OpenAI's own product page. It is OpenAI's first production model that does not run on NVIDIA hardware. Last week The Information reported OpenAI also hired Gimlet Labs, a 2026-founded inference-optimization startup that raised $80 million in March, to optimize OpenAI's models for Cerebras silicon. Gimlet claims its software accelerates AI inference up to 10x at the same cost and power. The Codex-Spark launch sits on top of the $10-20 billion+ Cerebras supply commitment OpenAI signed earlier in 2026.

OpenAI is also reportedly negotiating an equity stake in Cerebras, which would tie the chip purchase to ownership upside ahead of Cerebras's $33 billion IPO target. The pattern is the inversion of what NVIDIA built. NVIDIA's stack works because every major AI lab buys from one supplier; OpenAI's Cerebras move works because the largest single buyer of NVIDIA inference chips is now buying somewhere else. NVIDIA's last 10-Q already itemized a portion of its revenue as coming from companies in which NVIDIA also holds equity (the loop on 5/10 #4). The Codex-Spark launch is the first production model where a major AI lab's inference traffic moves to a non-NVIDIA supplier and the developer interface does not change.

Tom's Hardware coverage of OpenAI launching GPT-5.3-Codex-Spark on Cerebras chips — tomshardware.com · May 12, 2026

Why this matters: If you pay for ChatGPT or your team uses the OpenAI API for coding tools, the chip that runs the response you got back today might not be an NVIDIA chip. That matters for two reasons that show up in your tooling. First, latency and price drop on Cerebras-backed endpoints - Gimlet's 10x inference claim, if it holds even halfway, means the OpenAI API is about to get faster and cheaper for the model classes Cerebras runs. Second, the "every AI tool you use is bankrolled by NVIDIA" framing PickBits Daily ran on May 10 has its first real counter: the largest buyer of NVIDIA chips is buying somewhere else for production traffic. Action this week: open openai.com/index/introducing-gpt-5-3-codex-spark/ and read the inference-cost benchmarks; if you maintain a build that hits the OpenAI API in any volume, run the same prompt on the default endpoint and the Codex-Spark endpoint and write down the latency delta - that is the empirical basis for your next vendor conversation. If you sit anywhere near a 2027 multi-year cloud or chip commit, the announced-vs-deliverable capacity question now has a second supplier in the comparison, which is the first time you have actually had leverage in that conversation. If you hold NVIDIA stock or your retirement portfolio is tech-heavy, pull NVDA's next 10-Q (due August) and check the customer-concentration disclosure - the OpenAI line is the one to watch.

https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-lauches-gpt-53-codes-spark-on-cerebras-chips

» What to watch this week

Polis signature on Colorado SB 26-189. The Governor has said he plans to sign. The signing date and any signing statement language are what trade associations will quote in their state-bill testimony in Texas, New York, and California. The push-forward trigger fires when Polis signs (or vetoes).
The Trump AI security executive order text. Hassett said Monday they are "studying" an FDA-style framework but will not require mandatory model testing or stand up a new agency. Bloomberg saw a draft May 8. The text - if and when it lands - is the formal answer to whether US frontier AI oversight is voluntary self-reporting or testing-with-teeth.
Whether Provident Data Centers sues Hill County. County Attorney David Holmes warned the commissioners they could be sued for passing the moratorium. If Provident files (or the Texas Attorney General files on Provident's behalf), that suit becomes the test case for whether Texas counties can pause data center construction at all.
Connecticut Governor Ned Lamont signing SB 5. With Colorado retreating, Connecticut SB 5 is the only comprehensive state-level AI accountability law still moving. Lamont's office said he plans to sign. The signing day is also the day the post-Colorado map of state AI law has exactly one state on it.
OpenAI's next non-NVIDIA production launch. Codex-Spark is the first. The next model class OpenAI moves to Cerebras (or to AMD, Intel Gaudi, Google TPU, or one of OpenAI's own announced custom silicon programs) is the data point that turns "diversification away from NVIDIA" from one product line into a strategy.

Tomorrow's signal lands here.