He scored "tier one" for AI use. His actual pipeline has at least seven jobs in it

A reader in Techdirt's Insider Chat flagged a five-tier "REAL Rating" scale built to replace the blunt and misleading question "did you use AI on that?", and the site's author responded with an itemized account of where AI actually sits in his production pipeline ^[1]. He ran his own workflow through the rating wizard, came out with a "one" for automation, and said the label was technically accurate while telling you very little about how he actually uses the tools ^[2]. That gap is the useful part for anyone writing a policy.

The account is a history of substitutions, not a single decision. He first described his usage in April 2024, when he used a tool called Lex, trained on his own writing, to suggest edits on articles he had already written; some suggestions prompted a rewrite, some he ignored, and the value was in the forcing function ^[3][4]. In June 2025 he wrote about a personal task management tool built with the vibe coding tool Lovable, used partly to track what he wanted to write about each day ^[5]. He has since moved off Lex, which he says has gone largely without updates for a year or so as its team moved on, and built his own editing tool inside the task tracker, now self-hosted and entirely under his control ^[8][9].

What that tool does is specific enough to audit. Any task converts to a writing project with one click ^[10]. Sources go in as PDFs, URLs or pasted text, which he says pays off during edits ^[11]. While finishing the post he had the tool build a feature that automatically ingests every source linked in a story, so the editor checks them even when he does not add them by hand ^[12]. It can also go looking for additional sources, which he has not found better than searching himself ^[13]. After the draft is done, an "AI review" button returns a critical read against a style guide and a system prompt that spells out the help he wants, such as challenging his assumptions and facts, and the help he does not, such as rewriting his prose ^[14][15]. He defaults to Claude Opus 5 or Sonnet 5, swaps models periodically, and is increasingly running Gemma 4 locally ^[16].

Counted up, that is at least seven distinct AI-touched tasks sitting behind one tier label ^[18]. Two of them he tested and declined: automated source research ^[13], and drafting itself. At the start of the year he tried, with heavy scaffolding and detailed instructions, to get AI to produce a passable Techdirt article and concluded it could not, calling the attempts weak facsimiles with iffy language and cliched phrases ^[6]. He has re-run the test as models improved and says the conclusion holds, because the rewriting and fact-checking would cost more than writing the piece ^[7]. He still writes his own articles ^[20]. So the rejections are load-bearing too ^[19].

The operator lesson is that a usable disclosure regime has to be task-level: capture, source ingestion, critique, and generation are different risks with different failure modes, and a single yes or no flattens all of them. The second lesson is about dependency. The most consequential change in this pipeline was not a model upgrade but a vendor going quiet, which forced a rebuild ^[8][9].

Worth watching: whether tiered scales like the REAL Rating get adopted by anyone with a compliance obligation, whether self-hosted tooling that grows features on demand stays maintainable ^[12], and whether local models displace the hosted defaults in workflows like this one ^[16].