How EBL uses AI

EBL is edited by one person, Arnold Sandlow, D.C., and much of the reading and building behind it is done with a language model, Claude, made by Anthropic. This page says where that model is used, where it is not, and how a reader can tell which records it read.

Where a language model is used

• Reading trials against a claim. For each graded record, the model reads the cited trials adversarially: what each trial set out to measure, whether it met that, and whether the letter is the highest the written A to D definitions allow. That reading is recorded per record, in a checklist kept with the site.
• Applying the written grade definitions. Where the written definitions and rules settle a letter, the model applies them and logs which rule decided it. A letter that the rules leave open goes to the editor.
• Searching the literature and drafting. PubMed searches, first drafts of record text, and summaries of trials are prepared with the model and checked against the papers.
• Building the site and its checks. The pages, the tools and the hundreds of automated checks that run before every publish are written with the model’s help.

How a reader can tell who read a record

Every graded record says, in its Sources section, who read its papers against its letter: the EBL editor, or a language model not yet reviewed by a clinician. The sign-off ledger lists every record that way, from the same files, and the count of each is published on evidence integrity. No record has yet been checked by a reviewer outside EBL.

What a language model does not do here

• It does not move a letter quietly. A letter changes only after a logged review, and every change is published in the Evidence Log with the trial that moved it.
• It does not override the citation checks. Every cited paper is matched to its PubMed record and checked for retraction on every build by code, not by a model’s memory. A citation that does not resolve stops the build.
• It does not answer readers. There is no chatbot on this site, and no page gives a model’s answer to a reader’s question about their own health.
• It is not paid to say anything. No company has any say over a letter or its wording, whether the reading was done by a person or a model.

What a reader types stays with the reader

Search, claim checking, the medication and supplement tools and the calculators run in the reader’s own browser against files this site publishes. Nothing typed into them is sent to EBL or to an AI service. Voice intake uses the speech service built into the reader’s browser; on Chrome that service is Google’s and on Safari it is Apple’s, as the voice intake page states, and EBL never receives the audio.

Why this page exists

A language model can read faster than a person and can be wrong with confidence. The safeguards above are the reason it can be used at all: written rules it must apply, a record of who read what, checks in code that do not trust its memory, and an editor who signs off. Where a record has not yet had that sign-off, it says so.