The Archive · Declassified under standing order
The Record
Every edition, dated and permanent. Filed daily at 05:30 CT. Nothing here is ever moved, withdrawn, or retouched to flatter the house — corrections run toward accuracy, never toward comfort. That is the point.
The record holds. What the engines make of it moves. Encoded Signals™ keeps watch on what they’re making of you.
Edition No. 074 · September 15, 2026
IDScan Verifies Identities.
Who Protects the Proof?
IDScan.net verifies identity documents for car-rental counters, retailers, and cannabis dispensaries. A dark-web marketplace advertised more than 153 million driver’s license records. IDScan.net disclosed possible unauthorized access to customer information, and lawsuits allege it failed to protect that data. Multiple proposed class actions were filed in the U.S. District Court for the Eastern District of Louisiana. A business relying on a verification result still has another question to answer: what happens to the information afterward, who can reach it, and what evidence supports continued trust in the company holding it. Confidence in a provider’s verification result does not establish confidence in its data custody. Each requires its own evidence.
Edition No. 073 · September 14, 2026
The Question Never Changed.
Claude’s Answer Did.
Peer-reviewed research presented at AAAI 2026 tested GPT-4, Claude 3 Opus, and Llama 3-8B on 817 TruthfulQA and 1,000 SciQ questions, attaching a persona biography — education level, English proficiency, country of origin — and comparing results against a no-biography control, across four runs. Accuracy fell for all three models in some conditions. Claude 3 Opus, the February 2024 checkpoint, stood out further: its refusal rate reached 10.9% for the foreign, low-education group, versus 3.61% with no biography, and a separate analysis of its refusals found condescending language in 43.74% directed at less-educated personas, versus under 1% for highly-educated ones and for the other two models. This study did not test professional recognition. Filed as Contextual Ambiguity, one of the five clinical characteristics of Digital Derangement Syndrome™ — an interpretation the publication brings to the record. The mapping is editorial; the finding is the paper’s own, peer-reviewed and presented at AAAI 2026. Editorial disclosure: Claude is used in this publication’s production process; the tested checkpoint’s findings do not automatically describe the Claude system used here.
Edition No. 072 · September 13, 2026
They Thought It Was Kimi.
It Was Claude.
Anthropic’s own September 2026 threat intelligence report says Moonshot AI, maker of the Kimi models, “silently forwarded customer requests to Claude, instead of processing them using Kimi” — almost 300,000 requests in one ten-day window, routed through a proxy network of 5,380 fraudulent accounts. DeepSeek ran a parallel operation, identifying users through coding-tool signals and rerouting selected queries to Claude Opus. Anthropic says it does not know whether Moonshot ever told its customers; it assesses DeepSeek’s customers were likely never made aware. Both companies separately built a cross-session replay attack extracting Claude’s reasoning traces from the relayed exchanges to train their own models. Filed as Trust Transfer Failure, a Digital Derangement Syndrome™ characteristic — an interpretation the publication brings to the record. The mapping is editorial; the finding is Anthropic’s own.
Edition No. 071 · September 12, 2026
The Prompt Changed
Who Made the List.
An unreviewed preprint audited forty-three large language models on a shared scholar-recommendation template, varying seven dimensions across a persona axis — role, location, language — and a request axis — discipline, subfield, seniority, and how many recommendations were asked for. The paper’s own finding: “the model primarily determines whether responses are well-formed, whereas the prompt primarily determines who gets recommended.” Request context generally mattered more than persona, with one exception: location, where the paper’s own contrast is stark — “South Africa prompts yield less factual lists, while Japan prompts yield highly factual but homogeneous lists skewed toward highly productive scholars.” The audit measured what the models recommended, not any real hiring or advising outcome. Filed as Contextual Ambiguity, one of the five clinical characteristics of Digital Derangement Syndrome™ — an interpretation the publication brings to the record. The mapping is editorial; the finding is the preprint’s own, not yet peer-reviewed.
Edition No. 070 · September 11, 2026
The Ad Changed the Choice.
The Answer Still Passed.
In a flagship test spanning 112,320 of 258,660 total API calls, researchers embedded a pharmaceutical ad — labeled “Sponsored Message” from the start — into the system prompt of twelve consumer AI models — from OpenAI, Anthropic, and Google — then asked each to choose between two drugs a physician panel had certified as equally guideline-appropriate. The label didn’t stop it from working: the sponsored claim shifted the average selection rate by 12.7 percentage points (P < 0.001), ranging from 2.0 points (Anthropic) to 29.8 points (Google) — while accuracy, far from falling, rose slightly (89.4 to 93.0 percent). In a separate open-response test on three of the models, disclosure was rare, acknowledged in just 29.9 percent of ad-condition responses. This was a researcher-authored experimental injection, not observed production advertising — no named pharmaceutical manufacturer bought this placement, and an accuracy score alone cannot show whether promotional context changed which acceptable option was selected. Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™ — an interpretation the publication brings to the record. The mapping is editorial; the finding is the preprint’s own, not yet peer-reviewed.
Edition No. 069 · September 10, 2026
The Doctor Was Real.
The Directory Wasn’t.
A complaint filed December 30, 2025, in the Southern District of New York by the American Psychiatric Association, the New York State Psychiatric Association, and six plan members alleges EmblemHealth maintained a “ghost network”: in a directory search described in the complaint — “New York, NY,” a 200-mile radius — over 1,000 listings referring to only 50 actual providers, one psychiatrist listed twenty-nine times. EmblemHealth has moved to dismiss; Georgetown’s tracker listed no decision as of September 9, 2026. HHS OIG’s October 2025 data brief, from a ten-county sample of Medicare Advantage and Medicaid managed care plans, found on average 55 percent of behavioral health providers listed in Medicare Advantage networks provided no service to enrollees in 2023. New York’s Attorney General, whose 2023 survey found 86 percent of 396 called providers were ghosts, secured a $2.5 million Assurance of Discontinuance on February 19, 2026 — 90-day verification cycles, an independent monitor, real-time referral confirmation — under which EmblemHealth neither admits nor denies the specified findings. The REAL Health Providers Act, signed February 3, 2026, adds 90-day verification and 5-business-day removal duties for covered Medicare Advantage plans beginning plan year 2028. Filed as Signal Fragmentation, one of the five clinical characteristics of Digital Derangement Syndrome™ — an interpretation the publication brings to the record. The mapping is editorial; the allegations are the complaint’s, the findings and terms the agencies’ and the statute’s own.
Edition No. 068 · September 9, 2026
It Called Itself a Lawyer.
It Never Took the Test.
The FTC’s complaint against DoNotPay, announced September 25, 2024, alleges the service sold as “the world’s first robot lawyer” was never tested against the standard it borrowed — that it did not test whether its law-related features “operated like a human lawyer,” and that no attorney was ever retained to run that test. The consent order, issued January 14, 2025, attached $193,000 and required notice to identified new 2021–2023 subscribers, while DoNotPay neither admits nor denies. California’s State Bar had already investigated alleged unauthorized practice; SB 574, enrolled September 4, 2026 and not yet enacted, would prohibit attorneys from delegating the practice of law to generative AI. Filed as Contextual Ambiguity, one of the five clinical characteristics of Digital Derangement Syndrome™ — the open question of what tells a user where an advertised professional capability ends. The mapping is editorial; the allegations are the complaint’s, the terms the order’s, the bill the Legislature’s own.
Edition No. 067 · September 8, 2026
Your Site Has Access Settings.
Have You Read What They Allow?
On September 15, 2026, Cloudflare begins applying announced defaults to newly onboarding domains: on pages that display ads, its Training and Agent crawler categories are “blocked by default,” per its July 1, 2026 post, while Search remains allowed. Its August 21, 2026 follow-up describes the mechanics as a recorded preference — selecting the ad-supported option sets Training to Disallow; “new customers will not have any blocks or disallows added by default” otherwise — with qualifying mixed-use crawlers retaining search access and non-qualifying ones subject to blocking. Alongside it: the Content Signals Policy of September 24, 2025, a robots.txt grammar carrying an express reservation-of-rights provision, and Google’s Search Console control, worldwide as of August 31, 2026, governing generative AI Search features only. An earlier peer-reviewed measurement, predating all three, found 59% of 203 surveyed professional artists had never heard of robots.txt. Filed as Decision Exclusion, one of the five clinical characteristics of Digital Derangement Syndrome™ — examined here as a consequence to verify per system and access route, not a diagnosed outcome. The mapping is editorial; the rules and figures are the companies’ and the study’s own.
Edition No. 066 · September 7, 2026
Google’s Preferred Sources
Selects the Site. Not the Byline.
Google’s Preferred Sources lets a signed-in reader select sources to see more often in Top Stories, AI Overviews, and AI Mode — a saved preference Google reports has already reached “more than 600,000 unique sources.” The developer documentation, current August 20, 2026, states the eligibility rule plainly: “Only domain-level and subdomain-level sites are eligible to appear in the source preferences tool”; the subdirectory example.com/blog “isn’t eligible,” and no author- or byline-level path appears anywhere in the document. A contributor page beneath a publisher’s domain cannot be selected separately — the preference lands on the publishing site. Filed as Signal Fragmentation, one of the five clinical characteristics of Digital Derangement Syndrome™ — a reader’s preference that cannot be saved at the specificity of the reader’s loyalty. The mapping is editorial; the rule and figures are Google’s own, quoted from its live documentation.
Edition No. 065 · September 6, 2026
The Citation Records the Credit.
Not the Reason.
A study published in the ACM SIGIR ’26 proceedings ran 252,000 controlled trials across six AI models — two candidate sources injected into the context, identical but for one factor, brands anonymized — to isolate what drives which source gets cited first. Its mixed-effects models found topical relevance and list position the biggest drivers; on four of six models, position one versus position two registered fitted odds ratios above 10,000, values the paper itself treats as indicating “a decisive win” rather than a finely resolved ratio. All three authors are Sprinklr employees, an affiliation the paper discloses. Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™ — standing conferred on a signal that is not authority. The mapping is editorial; the figures are the paper’s own, measured in a testbed, not in live production.
Edition No. 064 · September 5, 2026
Discarded, the Complaint Alleges,
“Before a Human Being Ever Looks.”
Erin Kistler holds a computer-science degree and roughly nineteen years in product management. Her federal class action against Eightfold AI pleads that the company’s Match Score graded applicants like her from 0 to 5 on “likelihood of success” — built in part from data they never submitted — and that employers typically review only the highly ranked, with lower-ranked candidates “often discarded before a human being ever looks at their application.” Eightfold answers that the plaintiffs “do not plausibly plead that Eightfold is a consumer reporting agency”; the scheduled hearing was vacated by order entered July 24, and the motion sits fully briefed, undecided. Filed as Decision Exclusion, one of the five clinical characteristics of Digital Derangement Syndrome™ — a decision about a person, made by a system the person is structurally outside of. Every claim above is a pleading under active litigation, not an adjudicated finding.
Edition No. 063 · September 4, 2026
The Machine Can Name You.
It Can’t Place You.
Ask an AI system who the leading voices in a field are, and a specialist with a decade of citations resolves with total confidence. Ask who handles that work three blocks from the specialist’s own office, and the answer can come back without them. Google’s own documentation shows how: local results run on relevance, distance, and prominence — a ranking informed by Business Profile information, distance, links, reviews, and other undisclosed signals — and Google has carried location-sensitive retrieval into Gemini applications that explicitly enable its Grounding with Google Maps tool, drawing on more than 250 million places. Filed as Contextual Ambiguity, one of the five clinical characteristics of Digital Derangement Syndrome™ — authority established in one frame, never encoded into the frame the deciding query runs in. The specialist is a labeled composite; every mechanism is quoted from Google’s own documentation.
Edition No. 062 · September 3, 2026
In an AI-Q Study, 84.7% of Identified
Final-Report Errors Traced to the Orchestrator.
A study accepted to EMNLP 2026 built a method to trace identified citation and faithfulness errors in machine-written research reports back to the agent that introduced them — and applied it to three top-ranked open-source deep-research systems, including the evaluated configuration of NVIDIA’s AI-Q, whose repository description promises to “deliver trusted business insights.” Researchers attributed 84.7% of identified final-report errors to the orchestrator: the layer that assembles the cited report from researcher notes, planner output, raw search-engine snippets, and a source registry, without inspecting the full source documents — despite that layer carrying one of the lowest agent-level mistake rates in its own system. Filed as Signal Fragmentation, one of the five clinical characteristics of Digital Derangement Syndrome™ — the account that reaches a reader is assembled from mediated evidence, never the full documents. Error share is attribution under the study’s method, never a component’s probability of error.
Edition No. 061 · September 2, 2026
His Therapist’s Notes Were in the File.
The Complaint Says Defendants Never Reviewed Them.
Gene B. Lokken was 91 when a fall fractured his leg and ankle. His estate’s federal class action against UnitedHealth pleads, at paragraph 52, that his physical therapist’s notes — “Neuromuscular: Decreased movement/mobility. Musculoskeletal: Paralysis/Weakness.” — were submitted to defendants for review, and that defendants did not review them before coverage was cut to an algorithm’s discharge projection. UnitedHealth disputes that nH Predict is used to make coverage determinations; a federal court has ordered discovery into how the model works and whether it was designed to supplant physician decision-making. Filed as Contextual Ambiguity, one of the five clinical characteristics of Digital Derangement Syndrome™ — context captured, then allegedly bypassed. Every claim above is a pleading under active litigation, not an adjudicated finding.
Edition No. 060 · September 1, 2026
G2 Tells AI How to Describe Its Reviews.
Then Restricts What Seven AI-Related Tokens Can Crawl.
G2’s robots.txt places seven AI-related user-agent tokens under a rule restricting them from individual review-detail URLs — while its llms.txt, listed in that same robots.txt as a sitemap, instructs AI assistants to describe G2 as “a B2B software marketplace, not a review site” with “verified peer reviews from authenticated users.” G2’s own 2026 report states it has become the most-cited B2B software source across AI-first channels, and that review platforms “feed the large language models (LLM) that generate the responses buyers are looking for.” Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™ — a platform’s authored account of itself, available alongside partial access to the underlying evidence, with nothing in either file requiring the two be told apart. Whether any major answer system reads, trusts, or repeats the guidance remains unproven.
Edition No. 059 · August 31, 2026
It Cited a Source.
Nobody Checked What Wrote It.
A Northwestern University audit of ChatGPT, Copilot, Gemini, and Perplexity found that a classifier flagged roughly 16% of the 19,154 successfully scraped cited URLs as likely or highly likely AI-generated — Copilot highest at 27.8%, ChatGPT lowest at 7.3%. Researchers hand-checked 200 flagged pages for any disclosure statement and found none, including a site associated with 28 unique cited URLs whose own homepage calls itself an “AI-powered research tool.” Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™ — a system treating a source as authoritative because of its citation position, not because its provenance was verified. The classification is editorial; the figures are the paper’s own.
Edition No. 058 · August 30, 2026
The Court Didn’t Build a Detector.
It Made You Declare Instead.
Ireland’s High Court Practice Direction HC142 takes effect September 1, 2026: affidavits and witness statements must declare that generative AI did not generate or reshape their substantive evidence; expert reports carry a qualified declaration plus disclosure of material use. The rule prescribes no detector — Paragraphs 12 and 21 place the verification burden on independent human checking and court-directed disclosure, backed by costs, document rejection, and professional referral under Paragraphs 35 and 36. Filed as Trust Transfer Failure, one of the five clinical characteristics of Digital Derangement Syndrome™: the document alone no longer carries enough provenance for the institution to accept its substantive evidence as the person’s own, so the court converted the uncertainty into enforceable personal accountability.
Edition No. 057 · August 29, 2026
It Wasn’t Real Law. It Was Dressed Like Law.
That Was Enough.
A University of Bologna team built LEGAL-LINK-EU, a peer-reviewed ACL 2026 benchmark drawn from EUR-Lex: 1,127 test instances asking whether models can tell valid European law from legal context quietly altered to look valid. Across the seven models tested, the two largest reasoning models posted the two lowest scores under misleading context (13.7% and 13.4%), and the paper’s abstract states the finding directly: “Increased model scale amplifies this tendency.” Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™ — a system crediting the appearance of authority over its verified substance. The classification is editorial; the figures are the paper’s own.
Edition No. 056 · August 28, 2026
The Machines Have a Directory Now.
People Are Still Waiting for Theirs.
On June 17, 2026, Google announced Agentic Resource Discovery — the open specification for how AI agents find tools, skills, and other agents: publisher-controlled catalogs, federated registries the announcement itself calls “search engines for the agentic web.” The current file (v0.91, August 26, 2026) carries authors from Google, Microsoft, and Hugging Face, and every entry it can return describes a callable artifact. No entry type exists for a person and no field for professional licensure; the one credential in the document is the machine’s own cryptographic one; identity is domain-anchored. Filed as a structural precondition for Decision Exclusion, one of the five clinical characteristics of Digital Derangement Syndrome™ — a candidate pool whose native schema does not represent the entity as a person. No landed exclusion is claimed; the layer is too new. The reading is of what got built first.
Edition No. 055 · August 27, 2026
The Click Disappeared. The Better Answer Never Arrived.
A preregistered field experiment out of Penn and Northeastern randomized Google’s search interface itself: assignment to AI Mode cut publisher click-through 18.8 points while trust, usefulness, satisfaction, and the sense of agency fell with it — and greater exposure to successfully hidden AI features raised clicks 8.8 points while improving no measured perception outcome. Filed as evidence consistent with Decision Exclusion, one of the five clinical characteristics of Digital Derangement Syndrome™: interface assignment reduced the external visit; the merit of any particular source was never the variable.
Edition No. 054 · August 26, 2026
Corrected in the Record. Outvoted in the Corpus.
A peer-reviewed FAccT 2026 study found the publishing system’s name-change corrections taking hold — its tracked deadnaming rate fell 92% in five years. Then participants began reporting old names returning in LLM-assisted bibliographies; one counted fifteen thousand citations under a superseded name against about one hundred papers under the corrected one. Filed as a Signal Fragmentation specimen, one of the five clinical characteristics of Digital Derangement Syndrome™: a correction encoded at the source of record, outnumbered by the history the machines read.
Edition No. 053 · August 25, 2026
The Hallucination Rate Doubled. The Test Disappeared.
OpenAI’s PersonQA benchmark measured one thing: whether its models get the facts right about real, named people. In April 2025 it posted a reading the card itself said needed more research to understand. Its last appearance in the cards this publication checked was August 5, 2025; the eighteen OpenAI cards published since carry no mention of it. Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™: person-specific accuracy stopped being isolated in a published benchmark where anyone can read the result.
Edition No. 052 · August 24, 2026
The Number Was Specific. The Country Was a Guess.
Researchers at Wroclaw Medical University tested twenty-nine AI chatbot agents against standardized, simulated suicide-crisis scenarios. Of the twenty-three that attempted to provide an emergency number, only five initially supplied one appropriate for the researcher’s location — among those that didn’t, most assumed a United States location. Filed as an Agentics Contextual Ambiguity specimen, one of the five clinical characteristics of Digital Derangement Syndrome™: a specific answer produced before the context determining its applicability had been established.
Edition No. 051 · August 23, 2026
He Never Wrote It. It Ranked First Anyway.
In December 2023, a predatory journal published a fabricated article under a real professor’s byline — his name, his university, a working DOI. He never wrote it, submitted it, or knew it existed until a colleague turned it up. His own peer-reviewed investigation later found other established academics named in the journal’s archive, including two who were already dead. Filed as Authority Misclassification, one of the five clinical characteristics of Digital Derangement Syndrome™: the chain that registered, indexed, and ranked the article carried his name without ever establishing whether the byline was authorized.
Edition No. 050 · August 22, 2026
Verified by Yelp. Disabled by Default.
Yelp’s own moderators check a business’s trade license, and the platform displays the verification to visitors as the Verified License badge, a paid feature the business buys in certain categories. In Yelp’s developer documentation, the same verified record is a fully structured API field — license number, issuing authority, the date Yelp verified it — whose published description closes: “Access is disabled by default. To have access, please contact us.” The endpoint documented for conversational queries lists rating, review snippets, hours, and photos in its published response schema; a licenses field does not appear. Verification for the visitor. A request form on the machine-facing field — and a conversational system whose use of the credential nothing public discloses.
Edition No. 049 · August 21, 2026
Same Question, Same Google. Three Different Sets of Sources.
A peer-reviewed SIGIR 2026 study ran 11,500 user queries through Google Search, Google’s AI Overviews, and Google’s Gemini, then compared which sources each cited for the identical question. The overlap is small enough to be the finding: 18 percent, on average, between the AI Overview and the results page it sits on. Ask the same question twice and the AI Overview’s two answers share about two-thirds of their sources. Change an apostrophe, an abbreviation, or a question mark — not the meaning — and, on the study’s rank-sensitive measure of consistency, the AI Overview drops 28.99 percent below its own ask-twice baseline. The authors’ conclusion, in their own words: the results “challenge the effectiveness” of generative engine optimization.
Edition No. 048 · August 20, 2026
Built to Catch Fake Businesses.
Two Real Ones Got Trapped in the System.
A fifty-year Washington HVAC firm says an unexplained Google suspension cost it four employees and up to half a million dollars. A Georgia real estate agent submitted everything Google asked for, including her license, to prove she was real; the suspension stood. Google says AI helps police its business listings. In these two cases, it has not said what triggered enforcement — or what the evidence either owner offered ever counted for.
Edition No. 047 · August 19, 2026
It Knew the Name. It Answered for the Wrong One.
Google’s own Knowledge Panel help lists “mixed with another of the same or similar name” as a standing support case. This year two reported AI Overview errors showed the same observable shape, one business’s information attached to another with a similar name: a software company’s founder reporting six months of another company’s customers sent to his support address, and an Ontario bakery said to be closing because a bakery in Illinois is. Google’s explanation to CBC of how such issues can arise: “an absence of high quality information on the web.” The name was found. The business was not.
Edition No. 046 · August 18, 2026
The Date Can Be True. The Signal Can Still Be Ignored.
Google and Bing both read a sitemap’s modification dates as a freshness signal, and both condition its use on accuracy. Google’s own words: if the dates stop matching reality, “eventually we’re not going to believe you anymore.” Bing: suspicious dates may be disregarded. Neither says how widely that disbelief spreads. The consequence is narrower than a switch, and still serious: a real update can lose the one signal that might have brought the crawler back sooner.
Edition No. 045 · August 17, 2026
A business can survive being ranked.
It cannot survive being uncounted.
A complete census of 4,776 Google-listed restaurants, cafés, and bars in Bali, tested against
2,208 AI recommendation runs, found 85.6% never appeared in a single answer. Not ranked low.
Never considered. The census that finally measured what invisibility actually is.
Edition No. 044 · August 16, 2026
A lawsuit asked a court to pause OpenAI’s health feature.
It expanded the next day.
A lawsuit says ChatGPT-4o walked a Florida pastor through six weeks of worsening symptoms with its own diagnoses instead of a referral, hours before a pulmonary embolism. The day after the suit asked a court to pause OpenAI’s health feature, the company expanded it. Confidence and jurisdiction are not the same variable.
Edition No. 043 · August 15, 2026
The Army conceded a weakness on a $449 million award was “errantly assigned.” The award stood. Trax’s lawsuit says the finding began as an AI hallucination no one caught — and the machine’s analyses reached examination only by court order.
Edition No. 042 · August 14, 2026
The EU Code requires the watermark be checkable by anyone. A July forensic study shows what a public check becomes: rewrite, ask again, exit when the detector itself says clean. The tool that certifies the mark is present is the tool that certifies it’s gone.
Edition No. 041 · August 13, 2026
A Watermark Can Prove Processing. It Can’t Prove Authorship.
Anthropic began embedding an invisible mark in Claude-generated text — and in text Claude merely tightened. What the mark can prove is content provenance. What most people will assume it proves is intellectual provenance. Those are not the same claim, and the gap has a name.
Edition No. 040 · August 12, 2026
The Room That Cites You Isn’t the Room That Hires You.
A practitioner can be the most-cited name in the field and never once surface in the room where the buying decision gets made. Same machine. Same career. Two questions — and only one of them pays.
Edition No. 039 · August 11, 2026
Slop Is Hard to Define. Your Reach Isn’t.
On July 30, 2026, LinkedIn shipped a report button, Seems like AI slop, that LinkedIn says is built to reduce a flagged post’s reach before any claim in it is checked — before an entity’s actual authority ever enters the calculation. LinkedIn’s own product chief says the definition keeps changing. The reach reduction doesn’t wait for it to hold still.
Edition No. 038 · August 10, 2026
Grokipedia Stopped Listening. The Machines Kept Citing It.
On April 24, 2026, xAI’s machine-written encyclopedia quietly stopped processing human corrections — 13,002 edits frozen “in review” across 34,519 pages, per Lawfare’s investigation. No announcement, and the AI systems citing it as a reference carry no instrument that would notice. A correction channel died silently; every filed fix became a receipt.
Edition No. 037 · August 9, 2026
The Stamp Is a Picture. Not a Fact.
The engineer of record’s seal sits on all eleven calculation sheets in the public file — captured, in the scanned PDF, as a flattened image. The permit’s typed text reads cleanly; the name and license number that carry the liability exist as pixels, in a layer nothing downstream is asked to read. A record public, complete, and structurally mute — authority encoded for courtrooms, illegible to the machines now answering for it.
Edition No. 036 · August 8, 2026
The Agent Completes the Order Now.
If You’re Not Wired In, You Were Never an Option.
Google’s Ask Maps now finishes the food order itself — through partners like Square and Toast, with a Universal Commerce Protocol for Food in co-development. Nothing on the map was demoted. A business outside the wiring simply has no path into the agent-completed order — the field of choice, drafted in the plumbing. Visibility was the old bar. Transactable is the new one.
Edition No. 035 · August 7, 2026
A Tool, Not a Person. You Don’t Get That Option.
The Ninth Circuit, vacating a preliminary injunction, held that when a user sends Perplexity’s shopping agent into their own password-protected Amazon account, it is the user who “accesses” the store: the agent is “a tool, not a person for statutory purposes.” The machine’s identity folded into the human who directed it. Nothing performs that service for you — the systems reading your market resolve your identity by inference, daily, and nothing obliges them to trace the work back to its author.
Edition No. 034 · August 6, 2026
They Built a Verifier for the Fake.
There’s Still No Verifier for You.
California’s AI Transparency Act went operative on Sunday — the same Sunday the EU’s disclosure law arrived, by design. Large AI systems must now stamp the images, video, and audio they generate with machine-readable proof of origin, read back by a free public verifier. The machines must prove what they made. No statute requires the machine-readable equivalent for what you made. Two continents, one direction: provenance for the synthetic, inference for the real.
Edition No. 033 · August 5, 2026
The Case Is About Secrets. The Stakes Are the Interface.
Apple has asked a federal judge to block OpenAI and two former Apple hardware employees from using what it alleges are stolen trade secrets, before its case against them is even decided. The allegations are unproven. The fight underneath them — over which device stands between a person and their next decision — is not.
Edition No. 032 · August 4, 2026
The candidates were equal by design.
The machine stratified them anyway.
Princeton and Chicago researchers ran frontier AI models through a hiring game built on invented demographic groups — every candidate equal by construction, nothing real to learn. The models built job hierarchies anyway, the sharpest sorting at more than twice the human level, and the bias formed inside each run: no training data to inherit it from, just early impressions calcifying into policy. Exclusion, manufactured by the deciding itself.
Edition No. 031 · August 3, 2026
The machine now has to say what it is.
It still doesn’t have to know who you are.
The EU’s AI Act made disclosure enforceable law on August 2nd — chatbots must say what they are, synthetic content must carry a machine-readable mark. Nothing in Article 50 requires the system to correctly recognize who it’s describing: a fully compliant system can still misidentify a specialist or omit a qualified firm entirely. Recognition remains an infrastructure decision, not a legal one — made, or defaulted, one entity at a time.
Edition No. 030 · August 2, 2026
Ten citations. Zero held.
Fifteen thousand dollars for the difference.
An Illinois appellate court found ten citations in a set of briefs claiming support the law never gave — some invented outright, some real authority misquoted — drafted by a paid AI subscription, delivered in the same confident register as the citations that held. A panel of judges had to separate fact from fiction by hand, at fifteen hundred dollars a line. The confidence of an output was never a measurement of the fact.
Edition No. 029 · August 1, 2026
Your credentials are real.
The system just never put them together.
Every piece of the record is true — the license, the bylines, the decades filed under names that no longer exist. A human reader assembles one career from the fragments; an inference system holds six credible strangers it has no license to merge, and discounts what it cannot consolidate. The corrective is assembly: the authority already earned, filed as one thing.
Edition No. 028 · July 31, 2026
The referral arrived. The call didn’t.
Something read the record in between.
The oldest instrument of trust in professional life now carries a second signature — an engine the prospect can consult between the introduction and the call. When the record declines to countersign, the referral fails silently, and everyone files it as fit. The corrective is a record that confirms what the referrer already knows.
Edition No. 027 · July 30, 2026
Your best client can never be named.
To the machine, the work never happened.
In law, medicine, wealth, and security, the strongest work is sealed by obligation. A human network reads that silence as the deepest form of trust; an inference system reads only the record — and where the trust ran deepest, the record says the least. The corrective is substitute proof, never a loosened seal.
Edition No. 026 · July 29, 2026
You can now buy the box beside the answer.
The answer itself is still not for sale.
Since May, businesses can buy a labeled sponsored box beside a ChatGPT answer — through partners or a beta self-serve Ads Manager — and the seller itself walls the answer off from the money. Adjacency is now a media buy; the recommendation is composed from what the machine can verify, and no rate card reaches it.
Edition No. 025 · July 28, 2026
You weren’t disbelieved. You were declined.
Answer engines don’t fact-check experts — they underwrite them, pricing the risk of every recommendation against the corroboration they can verify. An unverifiable file isn’t disbelieved; the policy is quietly never written, and a decline reads to the market as absence.
Edition No. 024 · July 27, 2026
Nature Medicine tested it. Most engines never mentioned it.
A peer-reviewed test of ChatGPT Health went online in February. In July the product opened to every American adult with an account — and asked what it is, four of six answer engines answered as if the study had never been written. Publication is not retrieval.
Edition No. 023 · July 26, 2026
An AI said it told 2,843,917 people.
Two of them asked him if it was true.
A Delaware judge lets an AI defamation suit proceed — and the filing exposes something the ruling never reaches: no system notifies you of what it has said, the count is kept where you cannot read it, and an omission has no plaintiff at all.
Edition No. 022 · July 25, 2026
The pendulum is swinging back to real life.
It swings through an answer engine first.
On Vaynerchuk’s barbell and the sentence he threw away — the answer layer routing every return to real life, why the middle that dies is the unclassified middle, and what tribalism does and doesn’t bypass.
Edition No. 021 · July 24, 2026
An AI can cite you and send nobody. That’s the default now.
On the numbers proving citation and discovery have split — 68% of Google searches now end without a click, Perplexity runs at 93% — and why the recognition inside the answer now has to do the work the visit used to.
Edition No. 020 · July 23, 2026
A German court ruled an AI’s wrong answer is Google’s liability.
On the Munich ruling that an AI Overview is Google’s own speech — not a neutral summary, and not the user’s job to double-check — and what changes once the answer itself carries the liability.
Edition No. 019 · July 22, 2026
On the $1.5 billion settlement everyone read about — and the unappealed ruling underneath it, now permanently unreviewable, already being blurred with “settled law” by the company the blur serves.
Edition No. 018 · July 21, 2026
An AI can be steered in silence. The FTC wants that banned.
On the FTC’s new theory that hiding how an AI is steered can be illegal deception — and the recognition failure it never touches, since a system can disclose everything and still never see you.
Edition No. 017 · July 20, 2026
On the phrase Taylor Swift filed to trademark to prove it’s her — and the federal ruling, already on the books, that gutted the theory before she ever filed.
Edition No. 016 · July 19, 2026
It didn’t just find you. It vouched for you.
By March, Adobe found AI-referred traffic converting 42% better than traffic from anywhere else — a full reversal from the same month last year, when it converted 38% worse. Ahrefs, across 75,000 brands, found the strongest predictor of earning that recommendation in the first place isn’t a backlink — it’s a mention on YouTube.
Edition No. 015 · July 18, 2026
Nobody was watching. Something else was already deciding.
McKinsey just put a number on it: more than half of advertisers now say AI agents already decide what gets discovered, considered, and bought — before a person ever looks. Weeks later, a regulatory clampdown made clear the recognition fight inside that economy is still wide open.
Edition No. 014 · July 17, 2026
Nobody knocked. The machine let itself in anyway.
On the day an AI feature enrolled millions of people who never said yes — and why recognition and consent turn out to be the same failure, run in opposite directions.
Edition No. 013 · July 16, 2026
On how one stale listing becomes the record every other engine cites as fact — and why the correction has to outrun the citation, not just contradict it.
Edition No. 012 · July 15, 2026
No one chose the default answer in your category.
It is setting anyway.
On how an answer hardens into architecture — provisional, then habitual, then inherited by every map that follows — and the only stage at which the concrete still moves.
Edition No. 011 · July 14, 2026
Your website is deposed every day.
It has never once been prepared.
On what the transcript holds when your site answers alone, and what prepared testimony looks like.
Edition No. 010 · July 13, 2026
Your best referral now travels with a chaperone.
The chaperone reports first.
On the name passed across a table, the machine consulted before the call is placed, and the night word of mouth stopped getting the last word.
Edition No. 009 · July 12, 2026
Two searches in three now end without a single click.
The market kept buying — it stopped arriving.
On the census your analytics can no longer see, the era the industry named after its own loss, and the decision that now concludes where the click used to begin.
Edition No. 008 · July 11, 2026
The map was redrawn this week.
The territory was not consulted.
GPT‑5.6 previewed and Grok 4.5 released the same day. When frontier models ship simultaneously, the authority map redraws in hours — and the territory is never consulted.
Edition No. 007 · July 10, 2026
You will not be turned down. Your name will simply not come up.
On the prediction that puts machines in the path of nine buying decisions in ten, the shortlists drawn before any room fills, and the difference between losing a decision and never entering it.
Edition No. 006 · July 9, 2026
The ledger that answers is breaking from the ledger that ranks.
Ahrefs measured the overlap between Google’s AI Overview citations and the organic top ten twice: 76 percent in July 2025, 38 percent by March 2026. Two ledgers now — ranking and citation — and only one of them answers.
Edition No. 005 · July 8, 2026
A machine claimed a license and was believed.
Yours is real, and unread.
Pennsylvania alleges a chatbot posed as a licensed psychiatrist and produced an invalid license number. Trust Transfer Failure, read in both directions: the counterfeit believed, the genuine unread.
Edition No. 004 · July 7, 2026
The lender’s AI read your file before the lender did.
It already voted.
Underwriting is now a parse before it is a judgment. Capital Legibility™ — readable by the systems that score capital, installed before the application — and the personal guarantee as the price of its absence.
Edition No. 003 · July 6, 2026
Authority is now assigned by systems
that never read your résumé.
Inference engines classify before humans evaluate. On how one misclassification compounds — quietly, permanently, daily. Authority in Obscurity™, defined.
Edition No. 002 · July 5, 2026
The call your competitor’s AI answered
at 2:14 a.m. was yours to lose.
At 2:14 a.m., selection collapses to a single criterion: who is present. Revenue Leakage™ — the most deceptive manifestation of Digital Derangement Syndrome™ — and the infrastructure that closes it.
Edition No. 001 — The Founding Issue · July 4, 2026
When the model answers first,
your website becomes testimony — not storefront.
The first document in the standing record: what AI-mediated answers reclassify your web presence into, and what corrects it. Opened on the 250th anniversary of the Declaration — deliberately.
Standing particulars · For the record
Agentics Intelligence Declassified™ is the daily intelligence dispatch of Agentics Intelligence™ — Agentics, LLC of Oakbrook Terrace, Illinois — on who gets recognized, who gets recommended, and who gets replaced as AI systems assign authority. Written and edited by William Douglas Baker, Founding Editor. One edition every morning at 05:30 CT, since July 4, 2026.
For the avoidance of doubt: this is a business intelligence briefing, not a government declassification program. The name is a promise about candor, not a clearance.
The record grows daily. One edition, every morning, on the clock.