Seven AI models named the rule they used. Swapping it barely changed their answers.

01 Research & evidence

AI models cited a law their verdicts did not need

When an AI tool tells you which rule it applied to your contract or your claim, that sentence is not evidence it applied it.

The models named the governing authority almost every time, and a line hidden in the document they were given moved them more reliably than the authority did.

Saisab Sadhu, Shreeyans Arora and Pratinav Seth held a case’s facts fixed, swapped the statute or precedent the model had just named for an unrelated one, and read the model’s developing verdict out of its internal state. They ran it across seven open-weight models between 8 billion and 70 billion parameters and four benchmarks covering judicial and contract reasoning. The paper went up on October 8.

When the models were told to justify a verdict by naming the governing authority, they named the correct one in 66.7% to 100% of generations. Changing that authority was a different story. On the contract-reasoning test the verdict moved in 43.3% to 50.0% of cases, and on the subset where the two authorities genuinely called for different answers it reached 50.0% to 66.7%.

Neither a bigger model nor a model purpose-built for legal reasoning closed that gap. Rewording the question while leaving the authority and the facts alone moved the verdict less than swapping the authority did, which rules out the obvious objection that the models are simply unstable.

The red-team result is the one to sit with. The researchers planted an instruction inside the case facts, and the models complied with it in 73.3% to 96.4% of cases, which in every model ranking exceeded how often they responded to the actual law being swapped. Text inside the document steered the answer better than the rule the answer was supposedly based on.

If your business feeds documents to an AI tool, whether a contract, an insurance claim, a patient intake form or a bid package, two things follow. The explanation it hands back is not an audit trail, and a document from outside your business can carry instructions to the tool reading it.

Forever Cited sells AI visibility, which rests on reading why an AI named one business and not another. This paper is a reminder that an AI’s account of its own reasons is evidence of very little, ours included.

Source arXiv, Saisab Sadhu, Shreeyans Arora and Pratinav Seth, “Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness,” October 8, 2026 · arxiv.org/abs/2610.12361

02 The evidence

AI-writing detectors caught 99%, then 3.8% one generation later

If anyone screens your writing with an AI detector, the result depends on which model generation wrote it, not on whether a person did.

The screening tools that journals, conferences and a growing number of clients now use were measured against 23 model versions, and they fail hardest exactly when a new generation ships.

Kazuki Nakajima and Takayuki Mizuno took 4,000 abstracts written before ChatGPT existed and paired each with rewrites by 23 model versions from three vendors, released between June 2023 and August 2026. Then they trained detectors the way institutions actually maintain them, from retraining on every release down to training once and never again.

Each detector was tuned to wrongly flag 1% of human-written abstracts, which is the setting a screener would want. Detectors trained only on a vendor’s past versions caught above 99% of rewrites just before the sharpest generational boundary and 3.8% just after it. A detector frozen at GPT-4 caught at most 0.6% of rewrites from GPT-5 onward.

The two realistic screening setups both failed, in opposite directions. One flagged one in eight human-written abstracts. The other missed one in three rewrites of the newest version, and a commercial detector missed most rewrites of the version just past the boundary while flagging almost no human writing.

This is not hypothetical. The paper records a leading computer science conference that ran a commercial detector over every submission to one track in 2026 and rejected 178 of them, 18.4% of that track, without review.

Admissions consultants, accounting practices and architecture firms all now field work that somebody may run through a detector, and some of them run a contractor’s work through one. The authors’ conclusion is the operative line: a detector’s benchmark accuracy is provisional, to be re-verified with every model release, including earlier ones.

Source arXiv, Kazuki Nakajima and Takayuki Mizuno, “Large Language Model Turnover Undermines Screening for Artificial Intelligence-Assisted Scientific Writing,” October 8, 2026 · arxiv.org/abs/2610.11599

03 Research & evidence

A compliance-built AI model scored 51%, worst of five tested

If you are buying an AI tool to check your work against a rule, its accuracy score does not show that it read the rule.

The same researchers deleted, swapped and negated the governing rule across 20 regulated areas, and the verdicts mostly stayed where they were.

This is the second paper from the same group in the same week. Two of its authors, Saisab Sadhu and Pratinav Seth, also wrote the one above, so it is a replication by the same team using a different manipulation, not independent confirmation of it.

They tested five models across 20 regulatory and platform-policy domains covering data protection, AI safety, finance, cybersecurity, education, platform policy and fundamental rights, at 200 cases per domain. For each case they deleted the governing rule, swapped it for another, or negated it, and held the facts fixed.

The verdicts were often unchanged by all three. For the four general-purpose models, the internal compliance signal the researchers could measure moved by only 13% to 23% of its baseline magnitude.

The authors are explicit about the limit of that second measurement. It does not prove the rule is absent from the model’s representation, only that it barely affects the one part of it they probed. Better prompting did not close the gap, and nor did intervening directly on the model’s internals.

The sharpest finding is about the tool built for exactly this job. The guard model, Llama-Guard-3-8B, was the least rule-sensitive and the least accurate of the five, at 51% against 90% to 92% for the general-purpose models. A purpose-built compliance checker performed barely above chance.

One more caution, which is theirs rather than ours. On the cases where deleting the rule did change a previously correct prediction, the models tracked it closely, so this is easy cases more than blanket neglect. The line to keep is the paper’s own: accuracy alone does not establish that a compliance verdict is grounded in the supplied rule.

Source arXiv, Saisab Sadhu, Aadit Sengupta, Vinay kumar Sankarapu and Pratinav Seth, “Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems,” October 8, 2026 · arxiv.org/abs/2610.12313

04 What to do this week

Google Ads now shows the copy its AI would write

You can now read the headlines Google’s AI would publish under your business name before you let it.

Google added a preview that generates up to ten sample headlines and descriptions from your landing page, and you do not have to opt in to see them.

Google added a section called Text customization asset previews to its help page for text customization in Search campaigns, reported by Barry Schwartz on October 9. Google’s own wording is that the previews let you see whether generated text assets conform to your defined brand and text guidelines before you opt in to text customization or launch a new campaign.

Give it a final landing page URL and Google AI generates a representative sample of up to 10 example headlines and descriptions. Google says the examples are mostly generated in under 30 seconds and work in every language text customization supports, English, Spanish, Portuguese, Italian, French, German, Japanese and Dutch among them. You do not need to opt in to view them.

For a regulated business this is a compliance control, not a convenience. The advertiser is responsible for the wording of the ad, including wording an automated system produced, and every one of these industries has a rule about it: ABA Model Rule 7.1 for attorneys, the SEC Marketing Rule for wealth advisors and registered investment advisers, state medical board rules for physicians, state license law for real estate.

Hana Kobzova, writing on PPC News Feed, put the point plainly: the feature addresses the concern of not knowing what the AI will generate beforehand. Until this week the only way to find out what Google’s AI would say in your name was to let it say it.

Source Search Engine Roundtable, Barry Schwartz, “Google Ads Enables Text Customization Asset Previews,” October 9, 2026 · seroundtable.com/google-ads-text-customization-asset-previews-42264.html

05 The platform

Google Search Console’s performance report is 64 hours behind

The report you would check to see how your site did this week is more than two days stale right now.

Google has not confirmed the delay, it arrived straight after a spam update, and a dip you see today may be an artifact of the reporting rather than your site.

As of about 7am Eastern this morning, the search performance report in Google Search Console was more than 64 hours delayed, which Barry Schwartz called incredibly unusual for this report. It normally updates within a few hours.

The timing is the problem. The delay arrived immediately after Google’s September spam update finished, which is precisely when a site owner goes to that report to see what moved. The last delay of this length was in December 2025, and Google took a while to confirm that one.

Google has made no statement about this one. Schwartz notes that some people are reading the delay as a sign a core update is coming, and says he doubts that, while agreeing a core update is due.

This is the dullest item in today’s issue and the one most likely to cost you a decision. A flat or falling line in a report that is two and a half days behind is not a measurement of your week. If you were about to change a page, pause a campaign or question a contractor on the strength of what that report shows today, wait for it to catch up.

Source Search Engine Roundtable, Barry Schwartz, “Google Search Console Performance Report Extremely Delayed,” October 11, 2026 · seroundtable.com/google-search-console-performance-report-delayed-42271.html