AI Output & Research Disclaimer
Effective date: October 11, 2026
SynthGap uses artificial intelligence to find meta-analysis topics and to scan the evidence for the ones you choose. After you choose topics, that analysis runs automatically, without a person reviewing each step. This page explains what the AI does, what it can get wrong, and what you remain responsible for. It forms part of our Terms of Service.
In short: SynthGap drafts; you decide and you are accountable. Treat every AI-generated topic suggestion, refined topic, inclusion decision, study assessment and paragraph of text as a draft to verify against the source, never as a finished result. The deep analysis may reword or narrow the topic you chose, and it is a rapid evidence scan, with a preview meta-analysis, to help you decide where to invest your time. It is not a systematic review, and its preview is not a result you can cite.
1. What the AI does at each stage
| Stage | What AI does | What code does (no AI) | What you must verify |
|---|---|---|---|
| Topic discovery | Writes the search that bounds the literature the run works in (if your topic is a narrow question, also up to three searches for the wider research areas around it, which you are asked to choose from), proposes candidate meta-analysis topics from retrieved PubMed records, checks them against an index of existing systematic reviews, and writes a summary with a PICO framework | Counts what each search returns and offers the wider areas only when your topic alone matches too few articles to look for gaps in; the search that is run is the one you choose, and the result says if it covered a wider area; retrieval and similarity search over our PubMed index and review index | That the topic is genuinely novel and feasible. Search other databases and PROSPERO before you commit. If you chose a wider area, that it is the area you meant, and that a topic the agent calls "close to" or "elsewhere in" your area is one you want |
| Choosing topics | None. You choose which of the proposed topics to analyse. Each topic you choose uses part of your usage (and of your credit, once the month's usage is gone), and you can choose as many topics as your usage covers; deep analysis is part of the Plus and Pro plans, not of the Free plan | Nothing | That the topics suit your purpose |
| Refining the topic | One AI agent is given your topic, its PICO and the studies discovery picked for it, and the corpus that discovery searched. Its only job is to improve them: it may reword or narrow the topic, change its PICO or its scope, search the corpus by meaning, read abstracts, decide which studies belong and which of the studies discovery picked do not, check which systematic reviews already cover the subject, and say what they leave open. It is not given a number of studies to reach and it reads no full text at this stage | The searches of the pool, the lookup of reviews that cite the studies, and the search over review titles. The code refuses to include a study whose abstract the agent has not read | That the refined topic is still the question you want to ask: the page shows it with what was changed and why. Every included study and every study taken out (the downloadable list shows what became of each one). That the studies are all there is: the agent searches only the corpus that discovery searched, which is built from article titles and MeSH terms in a PubMed-based index, not abstracts, and that index is not a complete database. What it says is already reviewed, and what that leaves open: search other databases and PROSPERO |
| Reading the papers | Reads every included study at once. It chooses keywords; they are searched through the full text of each paper and it is given the short passages around them, paper by paper (a paper with no full text is given its abstract). It then assesses each paper: whether it has enough characteristics and data to be included. If no free full text can be found for a study, the analysis stops and asks you for its PDF | Retrieval of open-access full text and of any PDFs you upload, removal of the reference list, and finding the passages around the keywords. There is no limit on how many studies are read | Every assessment against the paper, especially numbers, units and what the study actually compared. A paper read from its abstract only is weaker evidence than one read from its text, and the report says so |
| Preview meta-analysis | An AI agent decides the primary outcome from your topic's question before it reads any numbers, reads the included studies (the abstract and, where there is one, short passages of the full text around keywords it chooses) and enters each study's numbers together with the sentence they came from. It may also fill in risk-of-bias forms and a certainty tally from what the papers say, decide which further analyses to run and which pictures to draw | All the statistics (effect sizes, pooling, measures of how much the studies differ, subgroup and sensitivity analyses, tests for small-study effects) and all the pictures (forest, funnel and others) are calculated by code, not by the AI. The code refuses a number that is not printed in the sentence the agent quotes and a sentence that is not in the text the agent was shown, keeps one report of each trial, and pools nothing from fewer than three studies | Every number against the paper, using the sentence it is shown with; which studies were used and which were left out, and why; that the outcome and the effect measure suit your question. The preview is an estimate of what a meta-analysis might show, built from abstracts and partial full text. The risk-of-bias forms and the certainty tally are an AI's reading of the papers, not a formal appraisal |
| Reports | Writes a critical report on the refined topic from what it read in the papers, the preview meta-analysis and the reviews it found, including whether it suggests pursuing, considering or skipping the topic, the strongest argument against doing it and what would change that. If you chose several topics, writes a closing comparison from the reports alone | Limits the suggestion by what was found: a topic with no included study is never suggested for pursuit, and one with fewer than three cannot be | Everything. The suggestion is a draft judgment, not a conclusion. Check each claim and number against the papers |
| Sweeps (Pro) | For every topic the 20 rounds found, a separate AI agent reads the topic and the studies picked for it, searches the literature (including a live search of PubMed and the index of existing systematic reviews), reads abstracts, and gives the topic a score out of 10 for the feasibility and quality of a systematic review or meta-analysis on it. | Asks for the 20 rounds, one after another; puts the topics in an order from the scores (equal scores are settled by the agent's own verdict, then by how many studies stand behind a topic, then by when it was found); stops when your usage runs low. The scores are used only for the order: the Service shows you the order and not the scores | That the order suits your purpose. A topic high in the order is not a recommendation to do it, and a low one is not a reason to leave it: each was assessed on its own by a model that can be wrong, from the studies our search found |
| Reserving a topic | When another user's discovery finds a topic that shares studies with one you reserved, an AI model reads the two topics and says whether they are the same or very similar | Counts the studies the two topics share, and leaves the topic out of the other user's result if the model says yes, or cannot answer | That a topic shown as taken really is the same as a reserved one, and that a topic still shown is not: the comparison can be wrong in both directions |
| PROSPERO draft | None | Fills in the fields of a PROSPERO registration from the topic's own words, adds standard wording for the methods and marks everything else for you to complete | Every field, before you submit it anywhere: SynthGap submits nothing, the standard wording is a suggestion and not your protocol, and PROSPERO does not accept a review that is already past data extraction |
2. What can go wrong
- The AI can be wrong. Language models can misread a paper, invent details, misclassify a study, or state something confidently that isn't in the source. The numbers in a study note are the highest-risk output.
- The topic can change, and not always for the better. The refinement agent is asked to improve your topic and its studies, so it may narrow, widen or reword the topic and change what it asks. That is by design, and the page shows the result next to what was changed and why, but it can also drift from what you had in mind, or settle on a topic that sounds stronger than the evidence supports. Compare the refined topic with the one you chose.
- Coverage is incomplete. Our indexes are built from PubMed and PubMed Central. They don't include every database, trial registry or the grey literature, so the Service can miss relevant studies and existing reviews. "Novel" or "still open" means "not found in our index," not "nobody has done this."
- Full text is partial. Open-access full text is available for only part of the literature, and a study with no free full text is read from its abstract, which often lacks the numbers a careful reading needs. Where the AI does read full text, it sees only short passages (about 200 words each) around keywords it chose, so it can miss something that sits elsewhere in the paper. Text extraction from PDFs can also drop tables and figures. You can add the PDF of a study that was read from its abstract only, remove a PDF you added by mistake, and re-run the analysis.
- Limits on how much is read. The agent works through a limited number of search rounds and reads the abstracts of the studies it considers, not every record in the pool, so it can stop early or miss studies. Up to 200 included studies get a note; studies beyond that are listed but not analysed.
- The preview meta-analysis is a preview. Its numbers are the ones an AI could find printed in the abstracts and the passages of full text it was shown, so a study that does not print its outcome in a form that can be pooled is left out (the preview lists which and why), and a mean and standard deviation that the code worked out from a median and quartiles, or a standard deviation from a standard error, are marked as estimates. It is not built from a systematic search, from screening, or from data extracted twice by independent reviewers, and a result that would change with a different choice of method says so. Subgroup, regression and over-time analyses are chosen after seeing the numbers, so they are hypotheses, and tests for small-study effects need about ten studies; with fewer, the preview says that publication bias could not be assessed. Where fewer than three studies could be pooled, nothing is pooled and the preview says so. The risk-of-bias forms and the certainty tally are an AI's reading of the papers and are not a formal appraisal or a GRADE assessment. Do not cite the preview as a result; use it to decide whether the topic is worth doing, and do the review properly if it is.
- This is not a systematic review. The analysis does not follow a registered protocol, search other databases, use independent human reviewers or meet PRISMA. Don't describe its output as a systematic review or meta-analysis in any publication.
- Older analyses. Analyses made before this version of the Service used earlier methods: the earliest screened abstracts, extracted data in two AI passes and pooled the results; the later ones widened the corpus around the topic and wrote a note for each of up to 50 studies. They are shown as they were made. Everything on this page about checking AI output applies to them too, and their pooled results rest on numbers an AI extracted, which you must verify against the papers. Analyses made before the preview meta-analysis was added have none; re-running one with a PDF builds it.
- The order of a sweep is only an order. A model's scores are noisy and are not comparable between sweeps or between topics assessed by different agents, so two topics close together in the order may differ in nothing that matters. Use it to decide where to look first, not whether to look.
- Results can vary. Running the same task twice may give different AI outputs.
- Earlier free trial. Until 28 October 2026 a free trial returned a single topic per run, chosen from the middle of what the run found: neither the strongest nor the weakest candidate. Runs made then still show only that topic; the others were not kept. Runs made since then, on every plan including Free, show every topic they find.
3. Your responsibilities
- Verify every AI output against the source before relying on it.
- Follow your field's methodological standards. Most (for example the Cochrane Handbook and JBI guidance) expect independent human reviewers for screening and data extraction. The Service's analysis runs automatically and does not replace them. If you go on to do a full review of a topic, do the screening and extraction yourself with independent reviewers.
- Check that every citation, PMID and number is real and says what the draft claims.
- Register your protocol where appropriate (for example PROSPERO or OSF) and follow the reporting guideline for your review type.
- Follow the AI policies of your journal, funder and institution, and disclose your use of AI tools as they require. Under ICMJE recommendations, for example, AI cannot be listed as an author and its use should be disclosed.
- Keep your own audit trail of decisions and changes.
4. Not medical advice
Nothing the Service produces is medical advice or a clinical recommendation for any patient, clinician or other person. Outputs are research-synthesis drafts intended for qualified researchers exercising independent professional judgment. Don't make or change treatment decisions based on them.
5. Not authorship; no guarantee of acceptance
SynthGap is a tool, not an author, and takes no responsibility for authorship criteria, journal or funder requirements, research-integrity standards, or whether a manuscript will be accepted. You alone are responsible for the accuracy, integrity and compliance of anything you publish or submit.
6. How your data is handled
To generate outputs, parts of your content are sent to the AI and infrastructure providers on our Subprocessors page. We don't use your content to train AI models. Don't enter information that identifies a patient or other individual. See our Privacy Policy.
7. Found a problem?
Please tell us at alikesen54@gmail.com: what you ran, what you expected, and what the Service produced. We can't fix what we don't know about.