While using GPT-4 Pro’s Deep Research feature for academic purposes, I encountered repeated and serious factual errors in author attribution. Specifically, it assigned research findings to the wrong individuals — which is a significant issue

ChatGPT
, , ,
1

I asked about a well-known orthopedic surgeon’s preferred implant stem and bearing combinations. Deep Research responded with:

“Dr. A reported a 95% bone integration rate in 1,000 cases using a specific titanium stem over 5 years…”

This is factually incorrect.

Actual paper reporting that result:

  • Title: Hip Arthroplasty Using [Stem X]: Evaluation of 1,000 Consecutive Cases
  • Authors: Author B, Author C, Author D (Dr. A was not involved)
  • PMCID: [PMC6284079]

In another instance, Deep Research claimed Dr. A led a long-term ceramic-on-ceramic bearing study, when in fact this study was published by a different surgeon (Dr. E) from another institution.

Why This Is Serious:

  • Misattributing clinical outcomes undermines academic integrity.
  • Affects citation accuracy and researcher reputation.
  • Makes Deep Research unreliable for author-level fact-checking.

Likely Cause:
Deep Research appears to use entity-based summarization that blends context around author names, institutions, and keywords, resulting in incorrect attributions.

This issue has become significantly more frequent since recent system changes (e.g., introduction of “lite” fallback modes and the use of smaller models like o4-mini).

Requested Improvements:

  • Improve author disambiguation logic to verify author-level attribution against actual publications.
  • Allow Pro users to clearly toggle between fast (lite) and accurate (fulltext) modes.
  • Provide an option to disable summarization, displaying instead the raw, accurately attributed source text.

I can provide detailed screenshots and further evidence if needed.

Thank you.

2 Likes