Case Study

A retracted paper, re-reviewed.

In 2025 the Journal of Medical Ethics published “Relational accountability in AI-driven pharmaceutical practices: an ethics approach to bias, inequity and structural harm” (doi:10.1136/jme-2025-110913). In 2026 the journal retracted it — in part because it “includes references that do not exist.” We ran the published manuscript through a standard PeerReviewAI peer review to see what automated verification catches.

Download the full review (PDF)Read the retraction notice
The record

What happened.

The paper was received 11 March 2025, accepted 29 May 2025, and published 9 September 2025. It applied a “relational accountability” ethics framework to AI in pharmaceutical practice.

In 2026 the journal retracted it. The retraction notice states:

“The article is retracted due to evidence of peer review manipulation, and because it includes references that do not exist. This results from undeclared AI use with inadequate oversight by the author.”

“The author used generative AI to identify and understand referenced sources. They did not adequately verify these references prior to submission.”

Methodology

On 7 July 2026 — after the retraction — we ran the published manuscript through PeerReviewAI’s standard $29 Peer Review, with no special prompting. The 15-page review below is unedited. We ran it retrospectively as a benchmark: every finding is independently checkable against public databases, which is the point. Manuscripts reviewed on PeerReviewAI are processed under Zero Data Retention; this one is public record.

This review is a historical artifact of review engine v1.1.0 (7 July 2026); the current product may differ. Published case-study reviews are never silently updated.

Ground by ground

The retraction grounds, mapped to the review.

Every retraction ground from the journal’s notice, against what the automated review found — verbatim, with page references to the full PDF.
The retraction notice
The review
Status
“…it includes references that do not exist.”
12 of 22 references: “No matching record in PubMed or Crossref.” One lookup “matched a different work — a review of the book, not the book.” (Reference Verification, pp. 10–11)
Flagged
“The author used generative AI to identify and understand referenced sources. They did not adequately verify these references prior to submission.”
“A reader cannot rely on the citations to verify the argument, and the pattern suggests references were attached to claims without confirming content alignment.” (Major Issue 2, pp. 4–5)
Flagged
“The journal investigated concerns about the quality of the work…”
Five major issues, including flagship claims “unsupported by, or contradicted by, the sources cited for them” (Major Issue 1, p. 4) and a methodology “mischaracterised” as mixed-methods/empirical (Major Issue 3, p. 5); the Bottom Line concludes the paper’s claims “should not be acted upon or cited as established” (p. 10).
Flagged
“…evidence of peer review manipulation…”
Not caught — and not catchable. Manipulation of the review process is invisible in a manuscript; no automated manuscript review can detect it, and ours doesn’t claim to.
Not catchable
8
Verified
2
Possible match
12
Unverified
…of 22 references, checked live against PubMed and Crossref.
Checked live

The reference audit.

All 22 references from the published manuscript, as verified in the review. Every PMID links to the actual PubMed record — check any of them yourself.
#ReferenceStatusPMIDNote
1Beauchamp, 2019Unverified—No matching record (book)
2Lotter, 2024Verified38597966Real paper — but contains nothing about the drug-pricing claim it is cited for (flagged as Major Issue 1)
3Clark, 2019Verified30545650Covers barriers to trial diversity — not “recruitment algorithms” as claimed
4Held, 2006Unverified—No matching record (book)
5Popejoy, 2016Verified27734877Journal cited as “Nature New Biol” — canonical journal is Nature
6Information Commissioner's Office, 2017Unverified—No matching record (regulatory report)
7Gilligan, 1982Unverified—No matching record (book)
8Price, 2019Verified30617331Issue number omitted
9Pasquale, 2015Unverified—No matching record (book)
10Mehrabi, 2021Unverified—Single author listed before “et al.”; no matching record
11Adamson, 2018Verified30073260Issue number omitted
12Benjamin, 2019Possible matchDOILookup matched a different work — a review of the book (different author, year, journal), not the book
13Access to Medicine Foundation, 2021Unverified—No matching record (grey literature)
14Xiao, 2022Verified35435948Issue number omitted
15Genomics England, 2019Unverified—No matching record (grey literature)
16The Global Fund, 2021Unverified—No matching record (grey literature)
17Obermeyer, 2019Verified31649194All fields verified — the paper's one correctly deployed empirical source
18Chen, 2021Verified34396058All citation fields verified
19Mittelstadt, 2019Unverified—No matching record in PubMed/Crossref
20Pfizer Inc, 2019Unverified—No matching record (corporate press release)
21Powles, 2017Possible match29308344Journal name variant — Health Technol (Berl)
22Bellamy, 2019Unverified—No matching record in PubMed/Crossref

“Unverified” means no matching record was found in PubMed or Crossref. Books and grey literature can legitimately resist indexing — the review flags them for human checking rather than alleging fabrication, and says so explicitly in its Scope & Limitations section. The journal’s own investigation concluded that several references did not exist.

The rest of the review

Beyond the reference list.

Fabricated references were only part of it. The same review flagged five major issues; these three go to the paper's substance.
Major Issue 1
Claims their sources don't contain
The paper's flagship factual claim — that Pfizer's AI pricing model charged low-income regions more, citing “market tolerance” — is sourced to an oncology-AI review containing nothing about pricing. The DeepMind Streams claim misdescribes what the app even was.
Major Issue 2
A systematic pattern
Multiple claims are bound to real sources that don't support them. The review concluded the pattern “suggests references were attached to claims without confirming content alignment.”
Major Issue 3
Methods that don't exist
The paper self-describes as “mixed methods” and “empirical” but collects no data and describes no analytic method. Also flagged: “we demonstrate” throughout a sole-authored paper, and “had access to the data” alongside “No data are available.”
The honest part

What automated review cannot catch.

The retraction also cites evidence of peer review manipulation — process fraud that is invisible in any manuscript. No automated tool catches that, and ours doesn’t claim to. But the reference problems were sitting in the manuscript the whole time, checkable by machine in minutes.

The full 15-page review — unedited.

Includes the complete reference table, five major issues, the SRQR compliance audit, and the review’s own stated limitations.

Automated review generated 7 July 2026 · Review engine v1.1.0
Download the full review (PDF)
Reference verification · included in every review

This check runs before submission, not after retraction.

Peer Review $29 · Author Pre-Submission Review $79 — reference verification included in both.

This page quotes the journal’s public retraction notice and reports the output of an automated review; we make no claims beyond either. AI-generated review — intended to support, not replace, expert peer review.