Signals

Anthropic publishes its own junior line: juniors fell from 27% to 9% of its technical hires

Anthropic published data on its own workforce: the junior share of its technical staff fell from 21% to 14% since early 2025, and the junior share of technical hires from 27% to 9%. Around a third of its machine-learning managers and a fifth of its software engineering managers said they were hiring less because of AI, and the median employee spends 20% more time in meetings than a year ago. It is one company and gives no absolute headcounts, but this is the junior headcount line the Ledger asks every employer to publish.

Caveat. A single company reporting on itself.

California requires employers to disclose AI-caused layoffs by job type and system, a first

California’s SB 951, signed 30 September 2026 and taking effect 1 January 2027, extends the state’s WARN Act layoff notice law to cover AI. When a layoff is substantially due to AI adoption, the notice must say so, naming the number and type of jobs displaced and the specific AI system involved. The state’s Employment Development Department will post notice summaries and a quarterly statewide report, making it the first US law to require public disclosure of AI-driven job displacement by employer and by system.

California bans firing by AI alone, the first state law requiring a human sign-off

California’s SB 947, the “No Robo Bosses Act,” signed 30 September 2026 and taking effect 1 July 2027, bars employers from relying solely on an automated decision system to discipline or fire a worker. A human reviewer must independently corroborate the finding using other sources such as personnel files and manager evaluations, and the worker must get written notice of the AI data used and a named human contact. It is the first US state law to require a human in the loop for this specific decision, by name rather than by implication.

California bans AI tools that read workers’ emotions or nervous-system data

California’s AB 1883, signed 30 September 2026 and effective immediately, stops employers using AI-powered workplace surveillance tools to recognise or predict a worker’s emotional state, or to collect neural data generated by measuring nervous-system activity. Violations carry a penalty of up to $500 each, enforceable by the Labor Commissioner, a public prosecutor, or the worker themselves under the state’s Private Attorneys General Act.

California requires hospitals to protect clinicians’ judgement whenever AI informs care

California’s AB 1979, signed on 30 September 2026 and in force from 1 January 2027, requires hospitals, clinics and medical practices to take reasonable steps to ensure that a licensed health care professional can still exercise independent professional judgement whenever an AI clinical decision support system informs a patient’s care. It also bars AI from independently performing clinical functions that require a licence. The authority to override the machine becomes something the organisation has to provide.

AI-adopting firms grow senior hiring 6.7% while junior hiring falls 2.5%, a 41-country study finds

A working paper by Bharat Chandar (Stanford Digital Economy Lab) and Bouke Klein Teeselink (King’s College London), released 21 September 2026, tracked 1.25 billion job postings and 154 million employment records across 41 countries since ChatGPT’s November 2022 launch. At AI-adopting multinational affiliates, overall employment rose 3.3%, made up of a 6.7% rise in senior hiring alongside a 2.5% fall in junior hiring: what the authors call “labour-saving for junior workers and labour-expanding for seniors.” The junior share of the workforce fell in roughly three-quarters of the countries studied.

Caveat. Working paper, not yet peer-reviewed.

Workers spend 42% of their AI time fixing its output

BambooHR’s survey of 1,608 US desk workers, published on 1 September 2026, found they use AI for an average of 87 minutes a day. 42% of that time goes on troubleshooting errors and re-prompting, against 35% on productive work: roughly 20 working days a year spent fixing the tool. It is the second survey this year to find that checking AI is a job nobody is counting.

Caveat. Vendor-run and self-reported: US desk workers only.

The second independent measurement of the same hidden job, after Glean’s botsitting figure of 10 June 2026.

HR trusts AI to run an interview twice as much as the people it would interview

BambooHR’s September 2026 survey found 67% of HR professionals would trust AI to conduct a candidate interview on its own, against 34% of employees: a two-to-one gap between the people who run the system and the people it is run on. 80% of those same HR professionals reported finding bias problems in the AI tools they use for hiring, and 33% had observed a pattern of demographic bias specifically. Confidence is highest among the people closest to the evidence of failure.

Caveat. Vendor-run and self-reported: US desk workers only.

From the same BambooHR release as the 42% rework figure, so not an independent source.

EU AI Act transparency duties apply; human oversight deferred

Article 50 transparency obligations came into application on 2 August 2026: people must be told when they are interacting with an AI system, and AI-generated content must be marked as such. The Article 14 human-oversight duties for high-risk systems, which name automation bias in statute, did not: the Digital Omnibus on AI, in force since 27 July 2026, deferred them to 2 December 2027.

Corrected 22 September 2026. This entry originally said the Article 14 human-oversight duties also applied from 2 August 2026. They were deferred before that date.

Workers spend 6.4 hours a week “botsitting”

Glean’s Work AI Institute published the inaugural Work AI Index on 10 June 2026. Workers report saving 11 hours a week through AI, then spend 6.4 hours a week “botsitting” (37% of their AI time) feeding context, supervising outputs and correcting wrong answers. 69% admit shipping work they haven’t verified, don’t fully understand, or can’t confidently stand behind.

Caveat. Vendor-run and self-reported: 6,000 workers in the US, UK and Australia.

Measured again, independently, by BambooHR in September 2026.

Connecticut enacts a comprehensive AI law covering employment, health and online safety

Governor Lamont signed Substitute Senate Bill 5 into law as Public Act 26-15 on 2 June 2026. Alongside employment, healthcare and online-safety provisions, it requires state evaluation and reporting on the adoption of AI by businesses and its effect on the workforce, with reporting due by 1 January 2027. Provisions phase in from 1 July and 1 October 2026.

Colorado strips impact assessments from its AI Act and delays to 2027

Colorado’s AI Act was due to take effect on 30 June 2026 with mandatory annual impact assessments. Governor Polis signed SB 189 on 14 May 2026, moving the effective date to 1 January 2027 and shifting the law’s focus from risk assessment to transparency. Disclosure survives; assessment gets negotiated away.

UN proposes a global dashboard to sit alongside GDP

The UN Secretary-General’s High-Level Expert Group published Counting What Counts: A Compass of Progress for People and Planet, proposing a global indicator dashboard to run alongside GDP. The General Assembly is expected to act on it. Sixty years after Kennedy’s Kansas speech, the macro argument is finally institutional.

Biased AI cut clinician diagnostic accuracy by 11.3 points, and explanations didn’t help

A randomised study of 457 clinicians found accurate AI predictions raised diagnostic accuracy by 4.4 points, while systematically biased predictions cut it by 11.3. Showing clinicians explanations that explicitly highlighted the model keying off irrelevant features did not rescue them.