What Deep Review is
Deep Review is an AI document reviewer from Google Research. According to the launch post, it checks logical consistency and mathematical statements, grounds claims against evidence, compares the work with existing literature and suggests improvements, while leaving higher-level judgements to people. The Deep Review page also describes feedback on weak arguments, missing context and overall clarity, together with a broad literature search.
It was piloted as pre-submission feedback with computer science conferences before public release, and the paper describing an earlier version also covers computer science. Neither the launch post nor the Deep Review page mentions medical or clinical manuscripts, or reporting guidelines such as CARE, STROBE or STARD. That coverage is unconfirmed in the public description, not proven absent.
Access and data handling, as Google states them
- Access: two reviews a week are free. For more, connect your own Gemini API key and pay for the Gemini API calls; Google says there is no additional charge for Deep Review itself. The Deep Review page estimates about $3 (Low with 3.8 Flash) to $20 (4 Argon on Extra High) for a 40-page document, and says actual charges vary by document and the estimate is only a guide.
- Input: the launch post lists PDF and plain text; the Deep Review page lists PDF, text, Google Docs and Google Slides. Word files are not mentioned.
- Training: the sources word this differently. The launch post says uploads are never used to train, fine-tune or improve Google models. The Deep Review page says documents are not used for training without your permission and offers an opt-in to help improve Deep Review.
- Retention: documents are kept for up to one year for display.
- Web search: the Deep Review page says the tool may run web searches for related work.
Before uploading an unpublished manuscript, check your institution’s rules and your co-authors’ agreement, and remove patient identifiers. Also check whether you have opted in to sharing data for improvement.
How we compared
The manuscript
One unpublished original article co-authored by a member of our team: a single-center study of an imaging AI’s diagnostic performance using retrospective data, with several tables and figures.
Deep Review
We used the report the author received after uploading the manuscript PDF to Deep Review, unchanged.
Paper Buddy Health Check
We imported the Word file of the same manuscript into Paper Buddy and ran Health Check. The first result we compared came from an earlier version; we ran it again after improving it.
Checking each comment
We checked every comment from both tools against the original Word file and the figure images, sorting them into correct comments, incorrect comments and issues both missed. Ratios and counts were recalculated with code.
This is one manuscript, the comparison was not blinded, and the Paper Buddy team did the checking. Which comments matter most also involved our judgement. Results may differ for other manuscripts and fields.
What Deep Review caught
Checked against the manuscript, Deep Review found most of the errors a reviewer would spot at once. Each takes minutes to fix but is easy for authors to miss.
- A reversed comparison: the Results said one value exceeded another when it was actually smaller.
- Table values and counts: the accuracy in a table did not match the count of misclassified cases. On a closer look, the figure could be a mean across cross-validation folds, so this was better treated as a question for the authors than as an error.
- A figure that contradicted the Methods: the training approach shown in a model diagram differed from the Methods text.
- A threshold chosen after the fact: a threshold picked by looking at the evaluation data was presented as a clinical use, with no note that the analysis was exploratory.
- Reproducibility details: training settings such as the optimizer, learning rate, number of epochs and data augmentation were missing, as was how scores were converted to probabilities, and a visualization analysis had a data-splitting problem.
- Small inconsistencies: a table footnote that confused patients with images, bootstrap counts that differed between figures, and typed superscript characters (instead of Word superscript formatting) that break in some fonts.
- Related work: it accurately suggested recent studies the manuscript did not discuss.
Where it fell short on a clinical manuscript
- False alarms from PDF conversion: it reported a missing first reference, sentences without citations, citations not in superscript and a repeated table heading. All were correct in the original Word file; they appear to come from extracting text from the PDF.
- Requests we could not act on now: new experiments such as comparisons with other models, an expert-reader baseline and multicenter data were mixed in with the revision requests. This looks like design rather than a flaw: the Deep Review page presents it as a strength that its feedback led 60% of authors to run entirely new experiments. That suits computer science papers, where experiments can be rerun, but in clinical studies limited to approved retrospective data such requests are often not possible now. They are better handled as limitations prepared for reviewer questions.
- Volume: there were dozens of detailed comments. Many were accurate, but the authors had to decide what to fix first.
- Reporting guidelines: there was no item-by-item check against a reporting guideline such as STARD for diagnostic accuracy studies.
- Clinical methods questions: it did not ask whether the clinician who produced the test inputs knew the final diagnosis (assessor blinding), or which diagnosis each repeat-visit image was labelled with and whether treatment happened in between. A clinical reviewer would likely ask both.
Paper Buddy Health Check on the same manuscript
The first result we compared was weaker than Deep Review. It missed the main numerical errors, the figure contradiction and the threshold wording, and it judged the manuscript too leniently. It did raise reporting items that Deep Review did not, such as assessor blinding, the recruitment period and whether enrollment was consecutive, but some items had judgements that contradicted their reasons, and some suggested related work was off-topic.
We traced the causes and made these changes by October 21, 2026:
- Automatic number checks: whether counts and percentages agree, whether a sentence saying A was larger than B matches the actual values, whether a value attributed to a table is really in that table, and whether author–year citations in the text appear in the reference list are now checked by calculation, not by the AI’s judgement. It calls something an error only when it is certainly wrong; a value that may have been calculated another way, such as a cross-validation mean, is raised as a question to confirm.
- Comments that disappeared for lack of evidence: Health Check does not show a comment unless it can find the supporting passage in the manuscript. Comments that cited a single table row or a single number used to fail this check and drop out; we fixed that.
- Order and voice: the result opens with an overall assessment of the study’s core answer and why it matters, then its strengths, then the few concerns to address first, with remaining detail checks folded. It is written as a colleague’s suggestions.
- Clinical methods questions: it now asks about blinding, participant selection, time and treatment between the index test and reference standard, and the unit of analysis.
When we ran it again after these changes, the concerns to address first were the overinterpreted conclusion, the threshold wording, assessor blinding, and the link between visit labels and treatment timing; the numerical mismatches appeared in the folded detail checks. The reversed comparison was still missed in that run, because in the manuscript as loaded into the editor that sentence ran together with the one before it and was read as one. Health Check now separates sentences joined this way, and after the fix it flagged that sentence as an error in our test.
Limits remain. Health Check does not read figure images, so it cannot catch contradictions inside figures, and it also missed the visualization data-splitting problem. It does not search new literature during the check, so it describes where the study sits in the field from research saved in the project or from the works the manuscript cites.
Side by side
| Aspect | Google Deep Review | Paper Buddy Health Check |
|---|---|---|
| Purpose | General review of technical documents; released after computer science conference pilots | Pre-submission checks inside a writing workspace for clinical, dental and nursing manuscripts |
| Input | PDF and plain text; the Deep Review page also lists Google Docs and Slides. Word not mentioned | A Word manuscript imported into the workspace, or the manuscript you are drafting |
| Output | A detailed list of comments with severity and location | Overall assessment and strengths, a few concerns to address first, folded detail checks |
| Numbers, tables and text | Its strongest area on our manuscript | Automatic number checks added; certain errors separated from items to confirm |
| Figures | Flagged a diagram that contradicted the text | Reads captions, not figure images |
| Related work | May search the web for related studies | No new search during the check; uses research saved in the project or the manuscript’s own citations |
| Reporting guidelines | Not mentioned in the public description; no item-by-item check in our run | Checks key items of the guideline that fits the design; not the full checklist |
| Clinical methods questions | No blinding or label-timing questions in our run | Raised both as main concerns after the changes |
| Revision requests | Included new experiments and extra comparisons | Focuses on fixes possible with current data; points harder items to the limitations |
| Where you fix | Read the report, then return to your file | Jump from a concern to the passage and fix it in the same editor |
| Access | Two free reviews a week; more with your own Gemini API key at API rates (about $3–$20 for a 40-page document in Google’s example) | Plan-based usage limits; see plans |
| Data | Training wording differs between the launch post and the Deep Review page; kept up to one year; may search the web | Manuscripts, research data and AI conversations are not used to train AI models; content needed for AI features is sent to AI providers for processing. See the Privacy Policy |
The Deep Review column’s features and access come from Google; everything else reflects what we saw on one manuscript. No tool guarantees that a manuscript is correct or that it complies with a reporting guideline.
Using both, step by step
The two tools answer different questions at different times rather than replacing each other. This is the order we suggest.
While writing: major flaws and reporting items
Once a draft comes together, use Health Check for the overall assessment, the concerns to address first, key reporting items and numerical consistency, and fix them in the same manuscript.
Before submission: a second reader
If your institution allows it, upload the final version to Deep Review for another pass over numbers, tables, figures and reproducibility details.
Sort comments into three groups
Fix now (numbers, tables, formatting); possible with existing data (explaining an analysis, a table of training settings); not possible now (a new cohort, an expert baseline). Turn the last group into limitations and prepared replies to reviewers, not revision tasks.
Finish with the original and the official checklist
Recheck every comment in your original Word file, complete the official reporting checklist yourself, and disclose the AI tools you used as your journal requires.
See our comparison of CARE and STROBE checkers, the pre-submission manuscript check and the AI disclosure guide.
Frequently asked questions
Is Google Deep Review free?
As Google stated on October 21, 2026, two reviews a week are free. Beyond that, you can connect your own Gemini API key and pay API rates. The Deep Review page gives about $3 to $20 for a 40-page document as an example, with actual charges depending on the document and the review level you choose. Free limits and prices may change, so check the official page.
Can I use Deep Review on a medical or clinical manuscript?
Neither the launch post nor the Deep Review page mentions medical or clinical manuscripts or reporting guidelines such as CARE, STROBE or STARD. That coverage is unconfirmed in the public description, not proven absent. On our one manuscript it caught numerical and figure mismatches well, but it did not check reporting-guideline items or raise clinical methods questions.
Can I upload a Word manuscript?
The launch post lists PDF and plain text; the Deep Review page lists PDF, text, Google Docs and Google Slides. Neither mentions Word files. Converting to PDF can disturb superscript citations and table structure and cause false alarms, so recheck any comment about citations or tables in your original Word file.
Is it safe to upload an unpublished manuscript?
The launch post says uploads are never used to train, fine-tune or improve Google models. The Deep Review page says documents are not used for training without your permission and offers an opt-in to help improve Deep Review. Documents are kept for up to one year for display, and the tool may run web searches for related work. Check your institution’s rules and your co-authors’ agreement, and remove patient identifiers before uploading.
If I use Deep Review, do I still need Paper Buddy Health Check?
They do different jobs. If you want a broad consistency review of a finished manuscript, Deep Review alone can give you a lot. Health Check is built for the manuscript you are still writing: an overall assessment, the few concerns to address first, key reporting-guideline items and clinical methods questions, fixed in the same place. Neither tool guarantees that a manuscript is correct or will be accepted.
Sources and how we checked them
We checked Google’s sources on October 21, 2026. The comparison reflects one manuscript, not an accuracy study or an independent evaluation. Check the current terms and data policy before you upload.
- Google Research — Analyzing technical documents with Deep Review (October 14, 2026)
- Google Research — Deep Review page and FAQ
- arXiv:2606.28277 — Towards Automating Scientific Review with Google’s Paper Assistant Tool
- EQUATOR Network — STARD reporting guideline for diagnostic accuracy studies
- Paper Buddy — what Health Check covers and its limits