Research tools
AI Research Paper Reader, Interpreter and Critical Reviewer
Go beyond a summary. The Reader explains a research paper, identifies its design, tests its methods and reasoning against the standards of its field, names its strengths and weaknesses, and shows the authors how the work could have been stronger.
Build your review prompt
Choose your settings, copy the prompt, then paste it into your preferred AI assistant together with your paper. The prompt carries the full reviewing framework, so the assistant reads and critiques the paper the same way every time.
Your prompt
- CopyCopy the prompt above.
- Add your paperOpen your AI assistant, paste the prompt, then attach the PDF or paste the text. If you paste a link, check the assistant can open it; if not, paste the text.
- Read and ask moreAsk follow-up questions, or paste a revised version for a point-by-point comparison.
Twelve domains of appraisal
Every paper is judged on the same twelve domains, with extra checkpoints for its specific design. Each domain gets a rating, evidence from the text, and a note on how any weakness affects the main conclusions.
- Problem and contributionIs the question clear, the gap real, and the claimed contribution proportionate?
- Literature and framingDoes the paper engage current and opposing work, and use its theory rather than only name it?
- Design fitDoes the design match the question, and do the claims stay within what it can support?
- Methods and measuresSampling, sample size, instruments, procedures, controls, preregistration.
- AnalysisFit of methods to data, assumptions, missing data, multiplicity, effect sizes and uncertainty.
- Results and consistencyCompleteness, flow of participants or data, and whether text, tables and figures agree.
- Validity and biasConfounding, selection, measurement, attrition, leakage, and generalizability.
- InterpretationWhether conclusions follow from results, and where claims overreach or spin.
- LimitationsWhether the real limits are acknowledged and uncertainty is stated honestly.
- Ethics and integrityApproval, consent, data protection, conflicts of interest, equity and dual-use.
- TransparencyData, code and materials access; replicability; adherence to reporting guidelines.
- Writing and presentationClarity, structure, tables and figures, and balance of referencing.
- Exemplary
- Strong
- Adequate with concerns
- Weak
- Critically flawed
- Not assessable
Constructive by design
The most useful part of a review is what to do next. The Reader sorts its recommendations so authors know where to start.
- Must fix: problems that undermine a main conclusion.
- Should fix: problems that limit confidence in an important result.
- Could improve: refinements that strengthen the work.
Each recommendation names the problem, the specific action, the expected benefit, and whether it needs only a reanalysis or rewrite, or additional data, or a new study. Each weakness is also labelled by severity: critical, major, moderate, minor or presentational.
Standards it draws on
| Randomized and experimental | CONSORT, SPIRIT, TIDieR; Cochrane RoB 2 |
|---|---|
| Observational | STROBE; Newcastle-Ottawa Scale, JBI tools |
| Non-randomized evaluations | TREND; ROBINS-I |
| Systematic reviews and meta-analyses | PRISMA 2020, MOOSE; AMSTAR 2, ROBIS, GRADE |
| Diagnostic and prediction studies | STARD, TRIPOD and TRIPOD+AI; QUADAS-2, PROBAST |
| Qualitative research | COREQ, SRQR; CASP |
| Surveys and mixed methods | AAPOR, CHERRIES; MMAT |
| Economics and evaluation | CHEERS; identification and robustness checks |
| Laboratory and animal research | ARRIVE 2.0; replication and pseudoreplication checks |
| Computational and AI research | Leakage, baselines, ablations, variance across runs, documentation |
| Case reports and quality improvement | CARE, SQUIRE 2.0 |
| Theory, engineering and humanities | Argument, proof and source-handling checks; evaluation against real baselines |
These are used as guides to what a well-reported study should contain. The Reader does not certify compliance with any standard.
Direct upload and link review
The next stage is a built-in reader: upload a PDF or Word file, paste a web address or DOI, or paste text, and receive the report here with downloads. It needs a secure server-side connection, daily limits per visitor, and a clear privacy design, so it is being prepared separately. Until it opens, the prompt builder above gives you the same reviewing framework in your own AI assistant.
Using it responsibly
- The review is only as good as the text it reads. A review of an abstract alone is limited, and the Reader says so.
- It cannot check raw data, images, statistical code or cited sources. A human expert should verify the points that matter.
- Do not upload confidential or embargoed manuscripts, including papers you are peer reviewing, without the permission of the journal and the authors.
- It cannot determine misconduct. Where numbers or figures look inconsistent, it describes what it sees and suggests the authors check.
- For clinical, legal or policy topics, it assesses research quality only. It is not medical, legal or policy advice.
Questions
Is this a peer review?
No. It is an AI-generated critical reading designed to help you understand a paper, prepare a review, or improve your own work. Real peer review is done by human experts chosen by a journal, who can ask the authors questions and examine the underlying data.
Which kinds of papers can it review?
Original studies of any design, systematic and narrative reviews, theoretical and conceptual papers, methods papers, case reports, preprints and thesis chapters, across the sciences, social sciences, engineering, humanities and the arts. It names the design it detects and applies the matching standards.
What is the difference between a quick and a full review?
A quick review is about one page: design, a short summary, overall confidence, top strengths, top concerns and top recommendations. A full review adds the twelve-domain appraisal, a claims-versus-evidence table, questions for the authors, a scorecard and a letter to the authors.
Can the Reader be wrong?
Yes. It can misread a table, miss context outside the paper, or judge too harshly or too leniently. It is built to anchor each point in the text, mark inference as inference, and say what it could not check, so you can verify quickly.
Can I use it on my own draft before submitting?
Yes. Choose “Author” as your role. You will receive prioritized revisions, likely reviewer questions, and an indication of whether the pattern of issues suggests minor revision, major revision or substantial rework.
Does Specialty Digest store my paper?
Not in this stage. The prompt builder runs in your browser and nothing you type is sent to Specialty Digest. What happens to your paper afterwards is governed by the AI assistant you choose to use.