AudioScholar

How a Specialty Weekly is made

From the week’s literature to a podcast, and where the human sits in it

Each specialty publishes one audio briefing per week, usually about 10 minutes and shorter in a quiet week. Here is exactly how it is assembled, what is automated, what is AI, and what a clinician decides.

“Where is the human–AI bridge: do you choose the articles, or does the AI?”

Neither extreme. A clinician sets the boundaries once: which journals, how recent, how often. Then every week, automatically, PubMed retrieves what is eligible inside those boundaries, and the AI selects and narrates from that retrieved set.

The AI never roams the open web, and no one hand-picks papers week to week. A citation guard rejects any paper the AI names that was not actually in that week’s PubMed results.

“What is the range of journals that gets scanned?”

~24 journals per specialty: the 14 elite generalist journals (NEJM, Lancet, JAMA, BMJ, Annals, Nature, Science…) scanned for every specialty, plus that specialty’s ~10 leading journals.

Drawn from a curated, impact-factor-ranked catalog of ~285 journals across 28 specialties, maintained by the clinician. Each run reads every paper added to PubMed since the previous run: normally 7 days, never more than 14.

Human (clinician) Automated · deterministic AI (Anthropic Claude) Guardrail
Once, at setup
0

The clinician defines the specialty’s “rails”

Human

For each specialty, a clinician sets the journal allow-list (from the catalog), the search scope, the weekly cadence, and the narration voice. This is a one-time decision per specialty, then revisited only when the journal catalog is updated.

This is the upstream half of the human–AI bridge: the human decides what is eligible, not which individual paper is featured.

Then every week, automatically
1

Retrieve the week’s candidate papers

Automated

A scheduled job searches PubMed / NCBI for every paper added since the specialty’s previous episode (normally 7 days, never more than 14) that appears in one of the allow-listed journals. Papers from the elite generalist journals count only when they are about that specialty. Papers without a usable abstract, and papers this feed has already covered, are removed. In the week to 13 October 2026 this left 21 to 86 candidates per specialty.

This is a plain database query against the U.S. National Library of Medicine, not AI. Conceptually: (allow-listed journals) AND (added since the last episode), with the specialty’s subject terms applied to the elite journals.

2

Select the most significant

AI

From that retrieved pool, the AI chooses up to 10 papers most likely to matter clinically this week, the editorial judgment of which of the eligible papers to feature. When more than 45 papers are eligible, it first reads every title plus the opening and closing of each abstract to shortlist 40, then reads the shortlist’s full abstracts to make the final choice.

The AI can only choose from what step 1 actually retrieved. It cannot add a paper from memory or the open web.

3

Write the spoken briefing

AI

The AI drafts the script from the selected papers’ abstracts: a connected roundup with the key findings and their citations, in plain spoken language.

4

Verify every citation

Guardrail

A citation guard checks each paper the script names against this week’s actual PubMed pool. Any citation that was not retrieved, or whose title doesn’t match, is dropped, and so is the script: it is rewritten, and if the rewrite still drops a citation, or more than 30% are dropped at once, that week’s episode is not published. A script that cites a reference number missing from the list fails the same way.

This is the safeguard against the AI inventing or mis-attributing references. Every drop is logged for audit.

5

Voice and publish

Automated

The verified script is converted to natural-sounding audio (Google Cloud text-to-speech) and added to that specialty’s podcast RSS feed: the same feed that appears in Apple Podcasts, Spotify, and other players.

Continuously
6

Clinician oversight

Human

Episodes publish without a human listening first. Afterwards the clinician listens to a selection, can remove or archive any episode, maintains the journal catalog, and reviews a per-episode “report an inaccuracy” channel. This is the downstream half of the bridge, human judgment closes the loop.


The human–AI bridge, in one line

The clinician lays the rails; PubMed retrieves within them; the AI selects and narrates inside the retrieved set; a guardrail keeps it honest.

No step lets the AI reach outside a clinician-defined, PubMed-only boundary. The human’s choices are structural and upstream (which journals, how recent, how often) and supervisory and downstream (listen, correct, remove), not week-to-week paper-picking. The AI’s job is the editorial middle: choose the most significant of what was retrieved, and explain it.


The journals scanned

A curated, impact-factor-ranked catalog

Every specialty briefing scans the same 14 elite generalist journals; most also scan a curated set of that specialty’s leading journals. Pick a specialty to see exactly which journals are scanned for it.

14
elite generalist journals
(every specialty)
 
specialty journals
for this specialty
 
journals scanned
for this specialty
~285
journals in the catalog
across the 28 specialties

Elite generalist: scanned for every specialty

Specialty journals

Papers from the elite generalist journals are included only when they are about the chosen specialty. The catalog is clinician-maintained and expanding.


Feedback

Suggest a change or share a thought

Spotted a journal we should add, a specialty to cover, or anything else worth knowing? Tell us.

Submitting opens a pre-filled email to the AudioScholar team. Nothing is sent until you press send in your email app.