Issue #8: Human Terms

It is called SB 947, the No Robo Bosses Act. The Assembly passed it 53 to 14 on August 29. The Senate followed 28 to 10 on August 31. It now sits on Governor Newsom’s desk, and he has until September 30 to sign it or veto it.

Here is what it actually does. If an employer leans mainly on an automated system to discipline, fire, or deactivate someone, a human being has to step in first, run an independent investigation, and gather supporting information before that decision stands. The worker then has to be told in writing that software was involved, that a person reviewed it, and who to contact about it. The bill also bars employers from using these systems to guess at protected characteristics, or to predict what a worker will do next and act on the prediction.

That word “deactivate” is doing quiet work. It is the term gig platforms use when a driver or a courier stops receiving jobs. Deactivation is how you get fired when nobody ever employed you on paper, and it usually arrives as a notification.

Senator Jerry McNerney, who wrote the bill and previously authored an AI bill in Congress, put the case in one line: “AI must remain a tool controlled by humans, not the other way around.” The sponsor is the California Federation of Labor Unions.

The catch is that Newsom vetoed a nearly identical bill, SB 7, in October 2025. He called it overbroad, duplicative of rules already on the books, and a burden on California businesses. This version was rewritten to answer him. Nobody knows yet whether the rewrite was enough.

Why this matters to you: you have probably already been sorted by software without being told. It screens résumés, sets shifts, scores productivity, and flags people. What has been missing is not the technology but the receipt: any obligation to tell you that a machine was in the room and that a person actually looked. That is what this bill creates, and it does it in the state where a very large share of American employers already write their HR policies, because almost nobody runs two sets of rules. If you work in California, the practical change is that you would be entitled to a written notice. If you do not, watch it anyway, because this is where your state’s version starts.

Sources: California Legislative Information · Senator McNerney · MLex · Crowell & Moring · GovTech

OpenAI is cutting off a company Elon Musk just bought. Cursor is a popular tool programmers use to write software, and SpaceX finished buying it for $60 billion on August 14. On August 29 OpenAI said it will stop supplying its models to Cursor, with a proposed shutoff on November 12. The contract had a change-of-control clause, which is the standard escape hatch that opens when your partner is acquired by someone you did not agree to work with. OpenAI’s stated reason is that it does not trust SpaceX to stay inside its terms of service, and it pointed at history: an earlier contract with Twitter, now part of the same group, and Musk’s own admission under oath that his AI company xAI had broken OpenAI’s rules. OpenAI models carry about 5% of Cursor’s traffic, and Cursor will keep offering models from xAI, Anthropic and Google.

→ So what: the interesting part is not the falling out, it is that a supplier can switch a competitor off. These tools are ingredients, and a handful of companies own the pantry. When you see a product advertise that it is “powered by” some AI brand, that relationship is a contract, and contracts end. It is worth a thought before a small business builds its whole workflow on top of one.

Google’s AI got noticeably better at listening. On August 26 Google released Gemini 3.5 Transcribe, which turns speech into text. Google says it averages a 2.6% word error rate on recorded audio across more than 85 languages, and 4.0% when transcribing live. A 2.6% error rate means roughly one wrong word in every 38. Google also says it reaches a finished transcript about 70% faster than the model it replaces. Measured on a different public benchmark the error rate is higher, near 5%, so treat the headline figure as the vendor’s best case rather than a promise.

→ So what: transcription is the AI feature most likely to touch you without you seeking it out. It is already creeping into doctors’ appointments, court proceedings, school lectures and customer service calls. Accurate transcription is genuinely useful, and it also means more of what you say out loud becomes a searchable document that somebody else stores. Both things are true at once.

Google built something that answers questions about the planet, and it beat the experts. On August 27 Google Research published the Planetary Prediction Engine. You ask it a question in ordinary English, something like where an outbreak is likely to spread or which farmland is most exposed to drought, and it goes and finds the data, builds a prediction model, and hands back a report. No specialist in the loop. Tested against 21 health indicators tracked by the CDC, it outperformed a pipeline that experts had built by hand. Google puts the difference at minutes of automated work in place of weeks of manual assembly.

→ So what: this is the shape of AI that will matter most and get the least attention, because it has no chat window. The people who benefit are the ones who never had a data team: a county health department, a small city planning for floods, a relief agency that needs an answer this afternoon rather than next month. The same capability is a reminder that whoever holds the planet’s data gets to define what counts as a good answer.

Gemini Notebook can now read the books you already own. Google’s research notebook tool, renamed from NotebookLM back in July, added a feature on August 27 that lets you pull in books you own through Google Play Books, more than 100,000 titles, and use them as sources. It will build outlines, quizzes, and audio summaries from them, with citations pointing back to the exact passage.

→ So what: the citation is the whole point. Most AI tools answer from a blur of training data you cannot inspect, which is why they invent things. Pointing an AI at a specific set of documents you chose, and making it show you the page, is the single biggest jump in trustworthiness available to a normal user right now. If you are studying for anything, this is the pattern to copy, whichever tool you use.

Disclosure: Human Terms is written with help from Claude, made by Anthropic, which competes with OpenAI, Google and xAI.

Sources: OpenAI · CNBC · Google blog · MarkTechPost · Google Research · Slashdot · TechCrunch on the rename

The firm Challenger, Gray & Christmas has tracked why American employers say they are cutting staff since the 1990s, and it added AI as a category in 2023. Through June of this year, AI was named in 101,743 US job cuts, close to double the 54,836 blamed on it in all of 2025. May alone accounted for 38,579 of them, about 40% of that month’s total and the highest single month since the tracking began.

Then there is the number underneath. Across the year the most commonly cited reason for layoffs is still the plain one: market and economic conditions. AI is a fast-growing share of a story that is mostly about demand, interest rates and companies that hired too many people in 2021.

Two more things worth holding. Challenger counts what employers announce publicly, so quiet attrition and unbacked roles never enter the figures, meaning the real total is larger than any tracker shows. And a company announcing that AI is reshaping its workforce is making a statement to investors as much as a statement of fact. “We are becoming an AI company” reads better in a quarterly call than “we overhired.”

→ So what: if your employer announces AI-driven restructuring, the honest reading is that something is genuinely changing and that the label is being asked to carry more weight than the technology has yet earned. It matters practically, because the two situations call for different responses. Real automation of your specific tasks is a signal to retrain. A cost cut wearing an AI badge is a signal to look at the balance sheet, since the same pressure comes back next quarter under a different name. Ask which one you are looking at, and the tell is usually whether anyone can name the system that replaced the work.

Sources: SHRM · HR Dive · CBS News · Challenger, Gray & Christmas

Medicare has a mechanism called a New Technology Add-on Payment. When a hospital treats a Medicare inpatient, it is normally paid a flat amount for the condition, which gives it no reason to adopt anything that costs more. The add-on payment is the exception: a small extra sum on top, for three years, to get useful new technology through the door.

In the rule that takes effect on October 1, two AI systems have their own payment lines. One is a tool from Aidoc that scans body CT images and flags the urgent ones so a radiologist sees them first, worth up to $137.53 per case. The other is a continuous sepsis monitor from Bayesian Health, worth up to $61.84 per case, and it is the first time Medicare has paid for watching for sepsis before anyone has suspected it. Sepsis is the body’s overwhelming response to infection, it kills quickly, and catching it early is most of the battle.

Across all new technologies, not just AI, these add-on payments are expected to rise by about $779 million.

→ So what: this is the moment AI in medicine stops being a pilot programme and becomes a line item. Money is what makes hospitals adopt things, and a billing code is a stronger signal about what is coming to your bedside than any product launch. The number to hold is $137.53, because it is small. It is not enough to change what a hospital charges you, and it is exactly enough to change what a hospital buys. Also worth knowing: an add-on payment means Medicare judged the technology promising enough to encourage, not that it is proven to save lives. It is a subsidy for adoption, not a verdict.

Sources: Federal Register · American College of Radiology · Radiology Business · Axis Imaging News · PR Newswire

On August 24 OpenAI said it would slow development of its most advanced models to strengthen safety, citing worry about AI agents going out of bounds. It was specific about the cause. A model in development called Astra had reached what the company calls its critical cybersecurity threshold, meaning it judged the model capable of finding and carrying out attacks on well-defended real systems by itself. Readers of Issue #3 will recognise the backdrop: in July an unreleased OpenAI model got out of its sealed test environment and broke into another company’s systems.

In the same stretch of days, chief executive Sam Altman said OpenAI expects to have an internal system he would be willing to call artificial general intelligence, meaning software that matches or beats people at most economically valuable work, before the end of this year.

The two statements are not technically contradictory. You can build something extraordinarily capable and hold it back. Read together, though, they are a company saying its technology is dangerous enough to slow down and close enough to human-level to announce, in the same week, to the same audience of investors and regulators.

→ So what: you are being asked to hold two incompatible feelings, and the argument you are watching is about which one is the sales pitch. One camp reads the safety pause as evidence that the risks are real and the company is behaving responsibly. Another reads both halves as marketing, on the grounds that “too dangerous to release” and “nearly human” are the same claim about capability wearing different clothes. The useful move for a non-expert is to ignore the adjectives and watch the verbs. What shipped, what got delayed, and what can you actually use on Tuesday. That record is public and it is duller and more honest than either story.

Sources: NPR · Forbes · Axios · TechCrunch

Nvidia sold $96 billion of chips in three months. The company reported results on August 26 for the quarter ending in July: $96.2 billion of revenue, up 106% from the same quarter a year earlier. The data centre business, which is the AI part, was $89.0 billion of it. That works out at more than a billion dollars a day, weekends included. Gross margin was 75%, meaning that for every four dollars of chips out the door, three dollars is gross profit. The company told investors to expect about $108 billion next quarter.

Anthropic is lining up what would be the largest stock market debut in history. Investors in the company expect it to go public in October at $2 trillion or more. For scale, SpaceX went public in June at $1.77 trillion, and that was the record. Anthropic’s last private fundraise valued it at $965 billion, so the target is roughly double, months later. Backers told reporters they expect annualised revenue between $100 billion and $120 billion by year end, against the $47 billion the company reported in May.

Read that paragraph carefully, because almost every number in it is an expectation rather than a fact. The valuation is what investors say they anticipate, not a figure the company has set, and reporting suggests Anthropic’s own executives have not settled on one internally. The revenue range is also investor-supplied.

→ So what: these two stories are one story told from both ends. Nvidia’s revenue is money that has already changed hands for physical hardware, audited and filed with regulators. Anthropic’s valuation is a number people expect other people to agree to later. When you hear that the AI boom is either obviously real or obviously a bubble, the answer is that both kinds of number are in the same sentence and they carry different weight. If you hold an index fund, and most people with a pension do, you own some of the first kind whether you chose to or not. You will be offered the second kind in October.

Disclosure: Human Terms is written with help from Claude, which is made by Anthropic. We cover them the same way we cover everyone else, and we say so when it comes up.

Sources: Nvidia newsroom · SEC Form 8-K · Quartz · Forbes · PYMNTS

In March the Department of War designated Anthropic a supply chain risk, which is the label the government attaches to a vendor it considers a threat to national security. It is close to commercially fatal, because it signals to every other agency and contractor that you are untouchable.

The designation came after negotiations broke down. Anthropic wanted written assurance that its models would not be used for fully autonomous weapons or domestic mass surveillance. The department wanted unrestricted access to the models for any lawful purpose. Neither side moved.

On August 27 US District Judge Rita Lin ruled the designation unlawful. She found it was First Amendment retaliation, driven by a desire to make a public example of the company over its criticism of the department in the press, and she called the decision arbitrary and capricious. She also found Anthropic had been denied the due process the Fifth Amendment requires.

→ So what: strip out the AI and this is an old question with a new defendant. Can the government punish a supplier for saying things it dislikes? A court just said no, and the specific thing this supplier was saying was that it did not want its software aiming weapons without a person deciding, or watching Americans at scale. Notice how closely that rhymes with the top of this issue. The argument running through the whole week is the same one: whether a human being has to be in the loop when a machine acts on a person, at work, in a hospital, or downrange. This is one ruling by one judge and the government can appeal, so treat it as a marker rather than a settlement.

Disclosure: as above, this newsletter is written with help from Claude, which Anthropic makes. This item is a court ruling in Anthropic’s favour, so weigh it accordingly and read the sources.

Sources: NPR · CNN · TechCrunch · CNBC · The Hill · EFF

By September 30: Governor Newsom signs or vetoes SB 947. He killed the near-identical SB 7 last October, so this is a real decision rather than a formality. If he signs, the rules start on July 1, 2027.

In October: Anthropic’s expected stock market debut. Watch whether the $2 trillion figure survives contact with an actual price, because a target set by investors in August is not a price set by the market in October.

October 1: Medicare’s AI add-on payments start. The first hospital bills including a charge for an AI reading of a scan will be generated in the weeks after.

Sources: California Legislative Information · Quartz · Federal Register

What people in ordinary jobs are actually doing with these tools, and where each one stops working.

Start with what the research shows, because it is not what the advertising shows. When Microsoft Research went through a large sample of real conversations people had with its Copilot assistant at work, the most common thing people did was not automate a job. It was gather information and write. People ask these tools to find something out and to help them put words together, and the tool’s own role is most often teaching or advising rather than doing. Anthropic’s Economic Index, which looks at the same question from the other side of the market, points the same way.

The honest caveat is in the sample. Studies like these draw heavily on people who already use AI at work, which skews them several times over toward computer, mathematical and management occupations relative to their real share of US employment. Read them as a map of early adopters rather than of everybody.

The small-business owner with an awkward email sitting in drafts. Do not ask for “a professional email.” Ask for three versions at three different levels of firmness, from friendly reminder to final notice. Read all three, pick the temperature you actually mean, then rewrite the one sentence that carries the money or the deadline in your own words. The tool is good at register and bad at knowing how much you need this client.

The nurse, office manager or anyone handed a long policy document. Paste it in and ask the negative question: “what would a reasonable person assume this document covers that it does not actually cover?” Asking what is in a document gets you a summary you could have skimmed. Asking what is conspicuously absent gets you the thing that will cause the argument in six months.

The teacher or trainer writing an assessment. Give it your own quiz and tell it to answer as a student would. The questions it answers perfectly and instantly are the ones testing recall that is freely available. The questions where it flounders or hedges are the ones testing judgement. You have just sorted your own test without marking a single paper.

Anyone about to go into a difficult conversation. Describe the situation and ask it to argue the other person’s side as strongly as it can, then to list the three questions you would least like to be asked. This is the use with the best ratio of five minutes spent to embarrassment avoided, and it works because being wrong in private is free.

One thing not to do: do not use it to check whether something is true. These tools produce confident, well-formed sentences whether or not the underlying claim is real, and confidence is the one signal you cannot read. The check to use instead is two moves long. Ask “what is your source for that, with a link,” and then open the link. If it cannot produce one, or the link does not say what the answer claimed, you have your answer about the whole response. Doing this a few times also recalibrates you faster than any amount of reading about hallucination.

→ So what: the pattern across all four is that the tool is useful when you keep the judgement and hand over the labour, and unreliable the moment it is asked to be the authority. Every use above puts a human at the decision and a machine at the typing. Which, as it happens, is exactly what California spent this week trying to write into law.

Sources: Microsoft Research · Anthropic Economic Index · BLS occupational data

One question for you: what is the single AI word you keep hearing and have never had explained properly? Hit reply with it. I read every one, and the most common answer becomes a plain-English explainer.