An AI wrote this draft, and two independent reviews scored it 6.0 and 5.5 out of 10
This page gives the method, the fact-check result, the reader scores, the known weak points and the check of Anthony's ten practices.
Who did what
- Anthony Booth wrote the brief and the text under "Where I stand" on the start page. He replaced that text on 7 October 2026. His words are unchanged except for typing errors.
- Claude, an AI model made by Anthropic, did the rest on 7 October 2026: the research plan, the pages and the checks.
- Ten AI research agents collected 427 rows of evidence from UK and US sources. Each row has a source, a link, a date and a grade.
- Anthony has not yet edited this edition. Treat it as a draft.
The checks, and what they found
| Check | Result |
|---|---|
| Fact-check of the first draft | Four agents that did not write the text opened each linked source. 287 claims checked: 223 confirmed, 11 wrong, 37 overstated, 16 not on the linked page. Each one was corrected or cut. |
| Quotations | Each direct quotation matched its source word for word. One paraphrase had been set in quotation marks and was changed. |
| Corrections | The corrected sentences were not checked a second time against the sources. |
| Pattern scan | A script searched each page for machine-writing patterns. An independent editor then found more by eye: in the second draft, 16 three-part lists, 12 short verdict sentences and 22 places where the text talks about itself. |
| Reader scores, first draft | A UK copywriter persona gave 5.5 out of 10. A US account director persona gave 4.5. |
| Reader scores, second draft | Copywriter 6.0. Account director 5.5. |
The reviewers' main verdict on the second draft: it is a guide in structure and a dossier in voice. After Anthony's signature, no page speaks as a person.
Known weak points
- The example on Levels of quality is invented.
- The script for the client-saving question on Why AI has no figure in it.
- The voice profile is the guide's first recommendation, and no study has tested one.
- Disclosure is covered on four pages and not stated once as a rule.
- Several facts appear on more than one page.
- Some sources were read through a tool that extracts text, and some papers by abstract only.
What the research did not find
- A named UK brand marketer on the record about an agency's use of AI for writing.
- A public test of AI ad copy against human ad copy with sales as the measure, apart from the wine retailer trials.
- An effectiveness award case that states AI-written copy with results.
- A figure for the share of agency writing that is AI-assisted.
- A survey of agencies or clients that use AI detectors.
Anthony's twelve practices, checked against the evidence
| Practice | Evidence |
|---|---|
| Train the agent on a library of writing you love | No test found. The nearest result: a model trained on a company's most successful emails sold as much as staff writers, and so did a plain prompt. |
| Make it check and publish its sources | Supported. A label of AI use shown with the sources carried no trust penalty in one study. The error rates on The issues are the reason for the check. |
| Do not regenerate copy across drafts; version and edit | No test found. |
| Keep a negative list of AI constructions | Supported. The sentence patterns persist after the stock words fade. |
| Set gates that make the model read its own output | Partly supported. A generate, check and revise loop raised rule compliance in one company's own paper (Walmart, 2025). No independent test. |
| Make it declare its context before it writes | No test found. |
| Edit for length, repetition and factual density | Supported. Shorter and simpler text got more response. |
| Make it list the issues in its own writing | Partly supported. Model feedback on a person's own draft improved ad clicks for non-experts. A model's critique of its own draft has not been tested. |
| A final human edit | Mixed. One sales trial found no gain. Each editorial standard requires it, for accuracy. |
| Ask for a diff of what it changed | No test found. |
| Make it read as the intended reader and score the piece | No test found. This guide used it, and the scores are above. |
| Talk to it with voice notes | No test found. |
His opening point, that AI-assisted writing performs better in advertising tests, is partly supported. AI copy matched staff writers on sales in three email trials. A model tuned on past test results beat human experts on subject-line clicks by 32%, and a bare prompt did not. Experts who let the model write the first draft of a social ad did worse (What works).
These points of his were not checked against the research: the 7.5 seconds, the 20% of task time for editing, the 2,500 documents, the fall in traffic from organic rankings, and the link between rankings and citations.
How to read the evidence
Each claim links to its source. "Vendor" marks a company that sells a product in that field. The evidence covers the UK and the USA, and most of it is from 2025 and 2026.