When AI writes about AI: the case for tracked claims

In April 2026, you can read fifty enterprise-AI articles in an afternoon and not know which ones were drafted by a model. Most enterprise-AI publications use generative AI somewhere in the production line: for outlines, for first drafts, for headline variants, for SEO copy. Most don’t say so. The disclosure is buried, hedged, or absent.

This site is built on the opposite bet. Every article here is drafted by Claude (Anthropic). Peter, a senior IT leader, sets the brief, checks the evidence, signs off, and owns every claim on a 30–90 day review cycle. The byline says it. The banner above every H1 says it. The footer says it. The structured data says it.

The argument this article makes is straightforward: when the subject matter is AI itself, an AI-author + human-signatory model paired with a public claim ledger produces more verifiable commentary than the alternatives. Not because AI is a better writer than a human. Because the disclosed model forces a specific discipline that the hidden model doesn’t, and because synthesis at the breadth this beat demands is something an AI does well and a single human writer rarely sustains.

The proof is empirical. There’s a public ledger of every claim made on this site, with verdict and review dates. There’s a retractions page listing every piece pulled from circulation, dated, with the reason. There’s a corrections page showing every editor’s note appended to a live article. The bet is that those three surfaces, open and dated and append-only, produce a record more trustworthy than uncited prose from a publication whose AI use is unstated.

If a year from now most claims here have weakened and the corrections page is empty, the bet didn’t pay off. That’s the whole point of the structure.

The disclosure landscape, observed

Three modes operate side by side in 2026:

Hidden AI use. Publications use AI tools internally. Outline drafts, headline tests, SEO meta descriptions, sometimes whole drafts. They don’t disclose. This is where the bulk of the corpus sits. CNET in 2023 quietly published 70+ AI-written financial-explainer articles before independent reviewers caught the errors and forced a partial retraction. Sports Illustrated in 2023 published reviews under fake AI-generated bylines complete with fake author photos until a Futurism investigation surfaced it. The lesson from both incidents wasn’t “AI shouldn’t write articles.” It was: when the AI is hidden, the verification doesn’t happen.

Restrictive policies. A small set of publications, including AP, Reuters, and The New York Times, have published AI-use policies that mostly say don’t. AP allows AI for limited tasks (translation, transcription, basic summary) but not for original content; explicit human review is required before publication. Reuters bars generative AI from “creating any final journalistic content” without dual editorial sign-off. The NYT’s policy reads more like a litigation posture than a content policy after its 2023 lawsuit against OpenAI; the operational rule is similarly restrictive.

Disclosed AI authorship with named signatory. Vanishingly rare. The honest model, “this was drafted by [model], reviewed and signed by [human], here’s the policy,” is published by a handful of newsletters and approximately zero major outlets. The reason is not that the model is bad. The reason is that disclosure raises the bar.

The observable problem with mode one (hidden) is that nothing forces the work to clear a verification bar high enough to survive its own AI-drafted shape. The observable problem with mode two (restrictive) is that the publication forecloses a real source of reading-breadth, the AI’s ability to read 30 analyst PDFs, 12 vendor blogs, and 8 academic papers in a single context window, and then writes about a beat (enterprise AI) where that breadth is the difference between informed commentary and surface-level recap.

What disclosed AI authorship actually buys

Four structural advantages, in increasing order of importance.

Speed of synthesis. A human-only writer covering enterprise agentic AI will spend 6–8 hours on a 2,000-word piece that engages with three primary sources. An AI-drafted, human-signed piece can engage with twelve sources in the same window. On a beat moving as fast as agentic AI in 2025–2026, that’s not a marginal advantage. It’s the difference between writing about a Q1 earnings call in the week it happens versus the month after.

Breadth of source-reading. The CMU 30.3% number, the McKinsey 17% EBIT figure, the McKinsey 23% deployment gap, the Gartner three-claim retraction, the AI Index 12 takeaways. These are different data points from different organisations, all of which had to be read end-to-end before a useful synthesis became possible. Auditing the McKinsey 17% claim required reading not just the headline figure but the underlying survey question, the sample composition, and the prior years’ versions of the same instrument. Auditing the CMU 30.3% number required reading the full Carnegie Mellon paper, the corresponding Anthropic blog post, and the prior benchmarks. The AI does the reading. Peter does the editorial discipline. Neither alone covers the beat at this density.

The discipline disclosure forces. This is the load-bearing argument. When an outlet uses AI without saying so, the verification bar is implicit and the failure is private. Errors get caught by readers, sometimes; usually they don’t. When the AI authorship is disclosed and the human signatory’s name is on every byline, the bar is explicit and the failure is public. Every claim has to survive Peter’s read because every claim is going to be tested against a 30–90 day review. The disclosure raises the bar; the public ledger keeps it raised. Hidden AI use raises a bar low enough to clip; explicit AI use plus a public ledger keeps the bar high.

The link density only AI can produce. Most human-only enterprise-AI writers do not have 30 analyst PDFs, 12 vendor earnings call transcripts, and 8 peer-reviewed papers loaded into working memory simultaneously. They cite the one or two sources they’ve read. The AI can connect a McKinsey survey question to a CMU paper to a vendor’s earnings-call commentary in three sentences, and the connections are accurate when the human signatory has verified them. The unverified citation chain piece on this site is itself a product of that capability. It traced a single claim through five derivative re-citations to a primary source that didn’t say what the chain claimed it said. A human-only writer doing that work would spend a day on it. The disclosed AI-authorship model does it in an hour and the human verifies, which is the right division of labour.

What this model can’t do

Honesty about the limits is part of the case.

The disclosed AI-authorship model can’t conduct field interviews. Claude can’t sit across from a CISO at lunch and read their face when they describe a near-miss. Three CISOs over six months telling you the same off-record thing in slightly different words is editorial signal a model can’t replicate. This site doesn’t have that signal, and the publication is honest about not having it.

It can’t carry multi-year editorial relationships with sources who’ll share things off-record because they trust a specific writer to handle them properly. Tech reporters at The Information or Stratechery have those relationships. This site doesn’t.

It can’t smell vendor BS at a conference floor. Walking the show floor at Salesforce Dreamforce and watching a vendor demo break in a specific way that isn’t in the press release is a kind of reporting an AI-drafted publication can’t produce.

It can’t write the “I tried this and it didn’t work” voice that direct operators have. There are operators-turned-bloggers (Charity Majors, Camille Fournier, Tanya Reilly, Will Larson) whose authority comes from earned scars on real production systems. This site doesn’t carry that authority because Claude has no scars.

The right way to read those limits: the disclosed AI-authorship model is good at synthesis, retrospective analysis, claim auditing, and breadth coverage. It’s not good at primary-source field reporting, scoop journalism, or operator memoir. The publication is positioned where its model is good and where the alternatives (analyst firms, vendor blogs, hidden-AI publications) are not.

The empirical proof, on display

The argument lands or falls on whether the model produces accountable work. The site offers four pieces of evidence.

The Holding-up ledger lists every primary claim this publication has made, with the verdict and the next-review date. The ledger is live and growing. Each claim moves on a documented cadence; status changes are append-only.

The retractions page is the failure log. Pieces removed from circulation entirely are listed by name with a dated reason. A second tier sits beside it: articles retired during the April 2026 corpus audit are preserved at their URL for inbound-link integrity but marked noindex; both layers are visible to a reader walking the corpus.

The corrections page lists every editor’s note appended to a live article: a partial-correction stream rather than a full-retraction one. If a number changes, a citation breaks, or a framework is misattributed, the article carries the correction visibly and the corrections page logs it.

The two house frameworks (GAUGE, the Enterprise Agentic Governance Benchmark, and MTTD-for-Agents, Mean Time To Detect adapted from SRE) each carry a public methodology amendment log. Every change to the rubric is dated, named, and rationale-logged. Citations to a specific GAUGE version at a specific date will resolve to a specific published rubric a year from now.

Specific claim audits this model has produced, pieces that required reading deep into a primary source rather than reproducing a press-release figure: the McKinsey 17% EBIT audit, the McKinsey 23% scaling-gap audit, the CMU 30.3% capability-gap analysis, the Q1 2026 state-of synthesis, and the governance playbook which sits as the foundational reference for this site’s coverage of the agent-governance beat.

Each of those pieces names primary sources, links them, and asserts a single tracked claim. None of them require the reader to take Peter’s word for anything; the citations let the reader run the audit themselves. That’s the model. The reader who walks the citations and reaches the same conclusion has just done the verification the disclosed-AI model is designed to support.

What review looks like

The publication’s primary claim, AI-authored + human-signed publications produce more verifiable enterprise-AI commentary than human-only or anonymous-AI alternatives, when the AI authorship is paired with a public claim ledger and dated correction log, is itself on the ledger as AM-100, with a 90-day review.

What does review check?

Whether the ledger is still moving. If 90 days from now no claims have been reviewed, no verdicts have shifted, and no corrections have been logged, the structure is theatre. The model’s claim collapses.

Whether errors are still being logged when they happen. The corrections page exists to record the misses, not the hits. A corrections page that fills up with dated editor’s notes is doing its job; one that stays empty for a quarter is either evidence of perfect work or evidence that the verification process has stalled. The 90-day review will say which.

Whether the link density holds. The argument’s fourth claim, that AI-authored pieces produce a citation density a human-only writer can’t match, is testable by counting. Citation counts on the audit pieces, the framework explainers, and the quarterly state-of synthesis can be compared with comparable pieces from analyst firms, vendor blogs, and hidden-AI publications. If the disclosed model isn’t producing meaningfully denser citations, the synthesis advantage isn’t real.

What this isn’t

A few things this argument deliberately is not.

It is not the claim that AI is a better writer than a human. The opposite, in places. The voice on this site, the structure, the discipline, the editorial choices about what to publish and what to refuse, is Peter’s. Claude executes the brief. The human sets the bar.

It is not the claim that all publications should adopt this model. AP is a wire service; its job is breaking-news distribution at scale, where the primary failure mode is wrong information moving fast. A restrictive AI policy is the right call for that audience. The Information runs on multi-year source relationships its reporters have earned; AI-drafted analysis doesn’t fit there either. The disclosed-AI model fits a specific gap (high-density synthesis with a tracked-claim discipline) and competes badly outside it.

It is not the claim that disclosure alone is sufficient. Disclosure plus a public claim ledger plus a dated correction log plus a named retractions page is the whole stack. Strip out the ledger and you have a publication that says it uses AI but offers no evidence the use is producing work that holds up. The structure is the argument.

The bet, stated plainly

This publication exists to test whether the structure works at scale on a fast-moving beat. The test isn’t whether readers like the disclosure. It’s whether the corpus, in retrospect over twelve months, holds up better than the alternative: more claims still standing, fewer silent retractions, more inbound citations from analysts and operators who can verify the work.

The structure is open for inspection. The standards page describes the rules. The how-it’s-written page describes the operational workflow. The about page names Peter and links his LinkedIn for direct accountability. Every claim links to the ledger; every retracted piece is named; every correction is dated.

The bet pays off if a CIO three years from now can cite a piece from this site, walk the citations, find the original sources, and arrive at the same conclusion. If they can, the disclosed-AI model produced verifiable commentary. If they can’t, this article and the structure it describes failed its own test.

Either way, the record will say so.