Instruments that changed

Every instrument on this page was found in circulation in September 2026, on a conference stage, in a briefing document or in a competitive analysis. Every one has moved. This page names none of the people who cited them, because the point is not that somebody was careless. It is that AI instruments are changing faster than the slides quoting them, and the only defence is to open the document.

1. OMB M-24-10 is rescinded, and the term it named is gone#

Usually said: that United States federal AI guidance requires agencies to guard against automation bias, citing M-24-10.

What is true: M-24-10, of 28 March 2024, did exactly that. It defined automation bias at section 6 as the propensity for humans to inordinately favor suggestions from automated decision-making systems and to ignore or fail to seek out contradictory information made without automation, and required at section 5(c)(iv)(G) that agencies ensure sufficient training, assessment and oversight for operators to combat any human-machine teaming issues (such as automation bias).

It was rescinded and replaced by M-25-21 on 3 April 2025, which says so in its own overview. The successor retains a requirement for human oversight, intervention and accountability for high-impact uses, and a route to timely human review and appeal. The phrase automation bias appears nowhere in it. The categories changed too: M-24-10 distinguished rights-impacting from safety-impacting AI, each with its own definition; M-25-21 collapses both into a single high-impact class.

One limit belongs with this, because it is the sort of thing that gets dropped in retelling. The words over-reliance, deskilling and complacency appear in neither document. The finding concerns one term, not a vocabulary. Both memoranda were read in full. Graded entry.

2. Canada's AI bill died, and has no successor#

Usually said: that Canada's Artificial Intelligence and Data Act is forthcoming, in progress, or a framework organisations should be preparing for.

What is true: AIDA fell with Bill C-27 on prorogation on 6 January 2025 and was never reintroduced. Canada has no federal AI statute and no AI bill before Parliament. The binding Canadian instruments on automated decision-making are the federal Treasury Board Directive, which applies to government departments rather than to employers, and provincial law. Quebec's is the strongest. It is set out in full on the Montreal page.

3. A New York human-review prohibition that is not in the statute#

Usually said: quoting New York State Technology Law as prohibiting agencies from using automated decision-making tools without meaningful human review, citing section 402 or section 502.

What is true: the codified statute does not contain that prohibition. Articles 4 and 5 as consolidated are disclosure and inventory regimes, and neither section 401 nor section 501 defines meaningful human review. The phrase survives at section 503, and what it says there is arguably stronger: an impact assessment must be bearing the signature of one or more individuals responsible for meaningful human review, and where an assessment finds the tool produces discriminatory or biased outcomes the agency shall cease using it and any information produced with it.

Section 503 cross-references a permission condition in section 502 that is no longer there, which is the visible seam where the prohibition was amended out. Both articles carry a note repealing them on 1 July 2028. Cite section 503, and say it binds government agencies rather than private employers.

4. Illinois has the duty and not the rules#

Usually said: that Illinois employers must now comply with new AI hiring regulations.

What is true: the duty is live and the regulations do not exist. Public Act 103-0804 added subsection (L) to the Illinois Human Rights Act with effect from 1 January 2026, and directs the Illinois Department of Human Rights to adopt the rules needed to implement and enforce it. The administrative code contains no reference to artificial intelligence and the department's employer compliance material does not mention it. Eight months in, there is an obligation and no instruction manual. The full position is on the Chicago page.

5. New York City's AI hiring law, and its own state auditor#

Usually said: that New York City's Local Law 144 shows AI hiring is now regulated.

What is true: the law is in force and is barely enforced, on the state's own account. Audit 2024-N-6, issued by the Office of the New York State Comptroller on 2 December 2025 and covering July 2023 to June 2025, found that the enforcing department surveyed 32 companies and identified one issue of non-compliance, while the auditors reviewed the same companies and identified at least seventeen instances of potential non-compliance. Two complaints were received in two years. The audit says potential, and the department disputes elements of it, and both of those belong in any retelling. Graded entry.

6. The 93 per cent zero-click figure describes about one per cent of Google#

Usually said: that 93 per cent of AI searches end without a click, presented as a fact about AI assistants or about search generally.

What is true: the study is real and the claim is not. Semrush analysed roughly 69 million Google sessions, United States desktop, May to July 2025, and found 92 to 94 per cent of Google AI Mode sessions were zero-click. Semrush states in the same piece that AI Mode was 0.25 per cent rising to about 1 per cent of Google search sessions in that window. So the figure describes roughly one per cent of search activity, not the category.

Two better sources exist if the argument is worth making. Pew Research, using browsing data from 900 United States adults across 68,879 searches in March 2025, found users clicked a traditional result on 8 per cent of visits where an AI summary appeared against 15 per cent where it did not, and clicked a link inside the summary on 1 per cent. And for Google search overall, SparkToro and Similarweb put United States zero-click at 68 per cent for early 2026. Neither is 93.

7. The 90 per cent of event planners figure rests on 92 people#

Usually said: that 90 per cent of meeting planners now use AI.

What is true: the underlying survey reported 91 per cent, from a sample of 92 self-selected respondents, fielded in December 2024 and conducted jointly with the vendor of the surveying body's own AI product, and explicitly labelled a preliminary pulse check. Only 15 per cent were classed as strategic users. Larger surveys give materially different numbers: one industry forecast puts it at 50 per cent and another at 65 per cent, with only 16 per cent saying it had materially improved planning. The word now is also doing work the data cannot support, since the fieldwork is from 2024.

8. Three figures with no source in existence#

Usually said: that speaker bureaus account for 65 per cent of bookings, directories 25 per cent and direct search 10 per cent; and that outcome-specific search queries have 70 per cent lower competition and three times higher conversion.

What is true: no industry body publishes a channel breakdown of speaker bookings. Not the Events Industry Council, PCMA, MPI, IAPCO or the trade press, each checked by name. Two tells make the number worth rejecting outright rather than caveating: the three figures sum to exactly 100, which real multi-select sourcing surveys never do, and the only literature discussing these channels is published by companies selling services to speakers. The 70 per cent and three times claim has no source either, and appears assembled from two unrelated pieces of search folklore, one of which is a traffic-share figure being repurposed as a competition figure. That is a category error rather than a rounding one.

9. Two Spanish claims, one of them geographic#

Usually said: that AESIA, Spain's AI supervision agency, is based in Madrid, and that it is the first national AI supervisory agency in the European Union.

What is true: the royal decree creating AESIA names its seat as A Coruña, in a building ceded by the municipality. The first-in-the-EU claim could not be verified at any primary source and appears only in secondary briefings, so it is not repeated here. Spain's genuinely quotable instrument is elsewhere and is stronger: CGPJ Instrucción 2/2026, binding on every judge in the country since January 2026.

Why this page exists, and what it is careful not to be#

Nine entries, from one week of competitive research. That rate is the finding. AI instruments are being written, replaced and amended faster than the material citing them is refreshed, and a claim that was accurate when a deck was built can be wrong by the time it is delivered.

This page names no speaker, no consultancy and no publication. It would be easy to write the version that does, and it would be worth less: the useful object is a checkable list of documents, not a list of people who were behind on their reading. Everyone doing this work, including the author of this page, has cited something that had moved. The corrections log records the ones found here.

The practical test is one question, and it takes about two minutes. Open the document. If a claim rests on a memorandum, find out whether it is still the operative memorandum. If it rests on a statute, read the consolidated text rather than the bill. If it rests on a percentage, find the sample and the population. Most of what is above would have been caught by that alone.