In California, an AI can no longer fire you on its own.
On September 30, Governor Gavin Newsom signed SB 947, the “No Robo Bosses Act.” Employers can no longer rely only on an AI system to discipline or fire someone. A person has to review what the software found and confirm the decision. It takes effect on July 1, 2027.
We covered the bill in Issue #8 when it passed, and flagged last week that Newsom had vetoed an earlier version. This time he signed it, along with a dozen other worker and AI bills, including one requiring notice when AI causes mass layoffs and one restricting workplace surveillance tools, including ones that try to read workers’ emotions.
He also took a swipe at Washington, a week after the White House renamed AI: “Super intelligence is clearly not coming from the White House – that’s why California continues to lead.”
- Oct 2025Newsom vetoes the first version, SB 7.
- Aug 2026A narrower SB 947 clears the legislature. We covered it in Issue #8.
- Sep 30 2026Newsom signs it, with 12 other worker and AI bills the same day.
- Jul 1 2027It takes effect. A human must review any AI-driven firing or discipline.
Sources: Office of the Governor of California, Sep 30 2026; Sen. Jerry McNerney’s office.
software already sets shifts, scores performance and flags people for warnings in a lot of workplaces. This law doesn’t ban any of that. It says the last word on your job has to belong to someone you could argue with. It only covers California, but big employers rarely run one HR system per state, so the habit may spread. If a system ever flags you, the useful question is now: who reviewed this?
SB 947 (McNerney) requires human review and confirmation of automated decision-system outputs before discipline or termination; the enacted version drops the broader advance-notification requirements that were in the vetoed SB 7. The Sep 30 package also included SB 951 (notice of AI-caused mass layoffs) and AB 1883 (workplace surveillance, including bathroom surveillance and emotion-inference tools); the Governor’s office lists 13 bills signed that day. Penalty and notice details should be read from the chaptered text, not secondary summaries.
Google’s Gemini 4 is here, but not for you yet.
Google released Gemini 4 Argon on September 30, but only to vetted cyber defenders (reported to include governments, hospitals and phone companies), because it can find and patch software security holes on its own. Paying developers come next, the public “as soon as possible.” That makes all three leading labs holding their strongest work back for safety reasons.
On the independent Artificial Analysis scoreboard it ties for third behind Anthropic’s two Claude models. The striking number is elsewhere: on a test of whether a model admits it doesn’t know, Gemini 4 made up an answer 15% of the time; OpenAI’s top models did so more than half the time.
And on October 7 Elon Musk said his Grok agent will use “the best back end model for any given task,” naming a rival’s first: Anthropic’s Claude Opus 5.5.
Source: Artificial Analysis, Sep 30 2026, hallucination rate on its own knowledge test. Lower is better. One evaluator, one test. Bars share one axis.
“smartest” and “most honest about its limits” are now different contests with different winners. For anything you can’t easily check, the second one matters more.
Gemini 4 Argon is distributed through Google’s Fairwind Program (650+ partners) at an introductory $2/$10 per million tokens, rising to $4/$20; Google reports 77.9% on DeepSWE v1.1. Artificial Analysis Intelligence Index: Opus 5.5 58, Sonnet 5.5 56, Argon 53, GPT-6 Astra 53, GPT-6.1 Sol 52. AA’s hallucination rate measures how often a model answers incorrectly rather than abstaining; cost per task $1.99 for Argon at introductory pricing (about $3.98 at standard rates) against $3.26 for Astra. Musk’s post also names Midjourney and Suno and does not say which tasks go where.
Free ChatGPT now wants to see your bank account. Ads are on the way too.
On October 2, OpenAI opened Finances in ChatGPT, which used to be paid-only, to free users in the US. It connects to your bank through Plaid, the service many budgeting apps use, and can pull your credit report from Experian. It can only look, not move money or pay bills.
Three days later OpenAI said that from late October, free users will start seeing labelled ads next to the pictures ChatGPT makes for them. Paying subscribers won’t.
free ChatGPT is turning into what free Google and Instagram became: a service paid for by advertising, now holding your spending history. That doesn’t make it a bad deal. It makes it a deal worth reading before you connect your bank.
Finances in ChatGPT is read-only, US-only, on web, iOS and Android, connecting via Plaid (12,000+ institutions) and Experian credit data. The image-ad test is limited to a small set of US advertisers, shown on Free and Go plans only and excluded for Plus, Pro, Business, Enterprise, Edu and under-18 accounts; OpenAI says ads do not influence model answers.
Meta’s hit AI app keeps notes on your friends, even ones who never signed up.
Last week we reported that Meta’s Muse, an AI that does tasks for you, had become America’s top free iPhone app. On October 6, a TIME investigation reported that Muse’s internal instructions tell it to build profiles of its users and the people in their lives, including people who don’t use it, noting “tensions and alliances” and good moments to nudge someone.
Meta says Muse remembers information “that you choose to share.”
you can decline an app. You can’t decline being mentioned in someone else’s. Before you tell any assistant about a colleague’s divorce or your sister’s money trouble, assume it is writing that down.
TIME’s reporting is based on Muse’s system instructions; it also reports an instruction not to tell users that original messages may remain visible after a deletion request. Meta’s response is quoted as given. Reported usage figures conflict between outlets and are not used here.
A famous investor says the AI bubble is close to popping. A Fed president isn’t so sure.
Ray Dalio, who founded Bridgewater, one of the world’s biggest hedge funds, warned on October 7 that AI is in a bubble. His worry is borrowing: “a huge amount of debt is being taken out to fund AI. As rates continue to climb, that is a point at which the bubble begins to pop.” How near is that? “I think we’re close to that.”
The other side came a day earlier from Mary Daly, who runs the San Francisco Federal Reserve. “These hyperscalers aren’t very interest rate-sensitive,” she told Axios, meaning the giant cloud companies doing most of the spending, which largely pay from their own cash. That could change, she added, as they borrow more.
both can be right. The giants can afford their bets; the question is everyone borrowing to keep up with them. Follow the Money, below, has exactly that.
Dalio spoke at the Forbes Global CEO Conference in Singapore; Bloomberg characterised his view as a “classic bubble”, a phrase not confirmed as his own. His framing is the standard debt-cycle one: bubbles burst when heavy borrowing meets rising rates. Daly spoke to Axios on Oct 6; the cash-flow explanation is the reporters’, not hers; she separately listed AI chip demand alongside tariffs and oil among inflation drivers.
Hiring nearly stalled in September, and companies are blaming AI less for cuts.
The US added just 29,000 jobs in September, the government said on October 2, and July was revised to a small loss. Unemployment was little changed at 4.2%. Pay rose 3% over the year, the slowest pace since May 2021.
Layoffs, though, are not where AI is showing up most. In Challenger’s monthly count of announced job cuts, AI slipped to the fifth-most-cited reason in September, after being first from March to July (we have tracked this since Issue #1). It still leads for the year. And when HubSpot, a marketing-software company, said on October 6 it would cut about 7% of staff, its chief executive wrote that the cuts were “not driven by AI-related efficiencies.”
Source: Challenger, Gray & Christmas, Oct 1 2026. 43,281 announced cuts in September. AI is still the top reason for the year to date. Bars share one axis.
the job market is soft for ordinary reasons: slow hiring, cautious companies. AI’s effect is showing up less in pink slips and more in the jobs that never get posted.
BLS Employment Situation, Oct 2 2026: +29,000 nonfarm payrolls against roughly 84,000 expected; July revised from +21,000 to −10,000 and August from +162,000 to +133,000; unemployment 4.2%. ADP had reported +90,000 private jobs for September on Sep 30, a reminder the two series often diverge. Challenger: 43,281 announced cuts in September; AI-attributed cuts 3,961 for the month and 120,136 year to date, about 21% of 573,195. HubSpot’s plan, approved Oct 1 per its SEC filing, carries an expected $65–75m charge.
Google put an invisible “made by AI” label inside designer proteins, and they still work.
AI can now design proteins, the molecular machines medicines are built on. Useful, and worrying: the same tools could design something harmful. In a paper in Nature, Google DeepMind showed SynthID Bio, which hides a detectable signature inside an AI-designed protein’s recipe and shape. In lab tests on three targets, including part of the COVID spike protein, watermarked designs worked as well as unmarked ones.
this is the deepfake-label idea applied to biology: the companies that manufacture DNA could check where an order came from. Google says it can’t yet stop someone deliberately stripping the mark, so treat it as a first step, not a lock.
SynthID Bio embeds a statistical watermark in generated protein sequences and structures; wet-lab validation on three targets (VEGF-A, the SARS-CoV-2 spike receptor-binding domain, and PD-L1), published Sep 30, showed hit rates and binding affinities comparable to unwatermarked designs. The intended deployment is screening by DNA-synthesis providers. Robustness to adversarial removal is described as unsolved. Detection-rate figures circulating in secondary coverage are not used here.
A memory-chip maker sold in three months what it used to sell in a year and a half.
Micron makes the memory inside AI chips, and inside your phone and laptop. On September 30 it reported sales of $54.2bn for the quarter, up 379% on a year earlier. That is about one and a half times its sales for the whole of its previous year, and it expects more next quarter.
Meanwhile, the Financial Times reported on October 6 that SpaceX, which now owns Elon Musk’s AI company, wants to borrow $40bn just to buy Nvidia chips.
Source: Micron Technology, fiscal Q4 2026 results, Sep 30 2026 (SEC Form 8-K). The forecast is the midpoint of the company’s own range. Bars share one axis.
the chip boom is real money. But Dalio’s warning above is about the second story, not the first: when the buying runs on loans, a slowdown stops being a disappointment and becomes a debt problem. Memory shortages also tend to push up the price of phones and laptops.
Micron fiscal Q4 2026 (quarter ended Sep 3): revenue $54.23bn against $11.32bn a year earlier, fiscal-year revenue $133.19bn against $37.38bn, guidance about $61.5bn plus or minus $1.5bn. The SpaceX financing, per the FT, is about $10bn of bank loans and $30bn of bonds, led by Apollo, closing in 2027; there is no company statement.
Washington praised AI’s self-policing. Most voters aren’t buying it.
Five days after the White House’s voluntary safety pact with AI companies (Issue #11), President Trump on October 4 named a “Super Intelligence Force” of officials to coordinate AI policy, led by the Director of National Intelligence, and praised the industry’s “tremendous self-policing.”
A Reuters/Ipsos poll published October 7 suggests the public sees it differently. 84% of voters see AI as a threat to workers, and 57% say the administration hasn’t taken its risks seriously enough, including a third of Republicans.
Source: Reuters/Ipsos poll, fielded Sep 30 to Oct 5 2026, 4,506 US adults including 3,526 registered voters. Self-reported attitudes. Bars share one axis.
Washington and California are now running opposite experiments: trust the companies, or write rules. Most voters, in both parties, are leaning toward rules.
The Super Intelligence Force is led by DNI Jay Clayton with FTC Chair Andrew Ferguson, the Pentagon’s Emil Michael and OPM’s Scott Kupor, reporting to the President and the Chief of Staff. Reuters/Ipsos fielded Sep 30 to Oct 5 among 4,506 adults (3,526 registered voters); 80% of Democrats and 61% of Republicans favour stricter regulation.
What to Watch Next
Gemini 4 for everyone. Google says “as soon as possible,” with no date. Watch whether the low made-up-answer rate survives once millions of people start using it.
AI copyright goes to the Supreme Court. The appeals court’s opinion against Ross Intelligence went public on September 30, and Ross says it will “seek review by the Supreme Court” (Issue #11).
Anthropic’s stock-market filing. Nothing public has been filed yet; the Wall Street Journal and Bloomberg report a November listing. When it is, we get the real numbers rather than leaked ones.
Has a computer system ever made a call about you at work, a schedule, a rating, a warning, that you wished a person had checked first? Hit reply to this email and tell me. I read every one, and the good ones end up in a future issue.