The three model jobs in a research pipeline
Research is the clearest case for a multi-model workspace, because the pipeline has three genuinely different demands. You need a model that can hold an enormous amount of text at once, a model that writes carefully enough to survive peer review, and a model that returns rigid structured output without drifting. No single product is best at all three, and the switching cost between three products is paid in re-uploading the same corpus. The marginal cost of switching inside one workspace is small and knowable: on Whizi's model cost index (OpenRouter list rates, fetched 2026-08-20), Gemini 3.7 Flash works out at $0.001313 per standard answer, Claude Sonnet 5 at $0.007, and GPT-5.6 Terra at $0.008, all under a cent per answer.
| Stage | Model | What it is doing |
|---|---|---|
| Reading a corpus of papers | Gemini | 1M token context window, so 15 to 30 PDFs sit in one conversation |
| Building an evidence table | GPT | Strict schemas, consistent column values, no editorial drift |
| Drafting prose and synthesis | Claude | Careful hedged academic register, faithful to the source |
| Finding recent work and checking references | A web connected model | Retrieval against live sources rather than memory |
| Adversarial review of your own argument | Any model that did not draft it | Independent reader, no sunk cost |
The single most important thing to understand before you start: a language model asked about a paper from memory will invent details, and it will do so fluently. A model reading a PDF you uploaded is working from the actual text. Those are different activities with different reliability, and almost every horror story about AI in research comes from the first being mistaken for the second. Upload the source. Always.
Reading 20 papers without losing the thread
The mistake is asking for a summary. A summary of 20 papers is a paragraph that could describe any 20 papers in the field. What you actually want is an evidence table, because a table forces the model to commit to a specific value per paper and makes the gaps visible.
Prompt: the evidence table
I have uploaded [n] papers. Build a table with one row per paper and these columns: citation (authors, year), research question, study design, sample size and population, primary outcome measured, headline finding with the effect size or key statistic, stated limitations, and funding source if disclosed. If a field is not reported in a paper, write NOT REPORTED. Do not infer values. Do not include any paper I did not upload.
Prompt: the synthesis pass
Using only the table you just built, answer: where do these studies agree, where do they disagree, and what explains the disagreement (design, population, measurement, or period)? Identify the questions in this literature that no uploaded study addresses. For every claim, cite the specific rows it rests on.
Prompt: methodological critique
For each study in the table, note the most serious threat to validity: selection, measurement, confounding, statistical power, or generalisability. Be specific to what the paper reports, not generic. If a study is methodologically strong, say so rather than inventing a criticism.
That last instruction is doing real work. Ask a model to find problems and it will always find problems, including in a well conducted study, because the request implies they exist. Giving it explicit permission to say a paper is sound keeps the critique honest.
For the mechanics of getting long documents into a chat cleanly, see how to chat with a PDF and the guide to summarizing PDFs with AI.
Extraction: getting structured data out of unstructured text
Systematic extraction is the task where AI saves the most defensible time, because the output is verifiable. You can check a number against page 7. You cannot check a vibe.
Prompt: strict schema extraction
Extract every reported statistic from this paper as JSON. Schema: {"value": number, "unit": string, "measure": string, "population": string, "location_in_paper": string, "confidence_interval": string or null, "p_value": string or null}. Include only figures that appear in the text, tables, or figure captions. Do not compute derived values. Do not round. If a figure appears in both a table and the text with different values, return both entries and flag the discrepancy.
The location_in_paper field is the part people leave out, and it is the part that makes the output usable. It turns a verification pass from re-reading the paper into a targeted check, which means the verification actually gets done.
Prompt: comparing across papers
Across the uploaded papers, find every reported estimate of [the quantity you care about]. Return a table of: paper, estimate, unit, population, method of estimation, and year of data collection. Do not average them. Note explicitly where the estimates are not comparable because of differing definitions or populations.
Instructing the model not to average is not pedantry. Averaging incomparable estimates is the single most common way an AI assisted literature review produces a number that is confidently wrong, and it is the kind of error a reviewer catches immediately.
Drafting the review without borrowing anyone else's sentences
Claude is the best drafter here, but the prompt has to enforce the two rules that matter in academic writing: claims stay attached to sources, and hedging is preserved. Models trained to be helpful tend to strengthen careful language, turning "was associated with" into "caused" without being asked.
Prompt: section draft
Draft the [section name] of a literature review using only the evidence table above. Rules: every claim must be attributable to a specific row, preserve the original hedging (do not upgrade an association to a causal claim), use [citation style] format, and mark any sentence where you were tempted to generalise beyond the evidence with [CHECK]. Audience: [field] researchers. Length: approximately [n] words.
Prompt: the hedging audit
Review this draft against the source material. List every sentence that states more certainty than the underlying study supports, and every sentence that attributes a finding to the wrong source. Quote the sentence and the conflicting source text side by side. Do not rewrite.
Run the hedging audit on your own writing too, not just on model output. It catches the same drift in a human draft written at 11pm.
Citation verification, which is not optional
This is the section that matters most, so it is deliberately blunt. Language models fabricate references. They produce citations with real author names, plausible journal titles, correct-looking volume and page numbers, and DOIs that resolve to nothing or to an unrelated paper. Retracted papers have entered the literature this way. Reputations have been damaged by it.
The rules that prevent it:
- Never ask a model to "find sources" for a claim from memory. Ask a web connected model to retrieve them, then open every link yourself.
- Verify every DOI by resolving it. A DOI that does not resolve is a fabrication, not a typo.
- Check that the cited paper actually says what the citation claims it says. A real paper attached to a claim it does not support is the harder error to catch, and it is more common than an invented paper.
- Never cite something you have not opened. This rule existed before AI and AI has made it load bearing.
- Keep the uploaded PDFs as the source of truth. Anything the model tells you about a paper you did not upload is unverified by construction.
Prompt: reference audit
For each reference in this list, state: whether you can confirm it from a source in this conversation, or whether it comes from your training data and is therefore unverified. Do not guess. Mark every unverified entry clearly. I will check them manually.
That prompt does not make the model reliable. It makes it explicit about which claims are anchored to something you provided, which is the distinction you need in order to know what to check.
What this does not replace
It does not replace a systematic review protocol. If your work requires PRISMA, the search strategy, screening, and inclusion decisions remain yours to document and defend. AI can help you draft the protocol and speed up extraction from included studies; it cannot be the search.
It does not replace your reference manager. Keep Zotero, Mendeley, or EndNote as the record of what you have actually read. The chat is working memory, not an archive.
It does not replace statistical analysis. A model can explain a method, write the R or Python for it, and sanity check your interpretation, but running your analysis inside a chat window and trusting the arithmetic is not defensible. Generate the code, run it in your own environment, and check the output.
And it does not replace disclosure. Journals and institutions increasingly require a statement of how AI tools were used. Write down what you used at each stage as you go, because reconstructing it at submission is unpleasant.
- Upload the papers rather than asking the model what it remembers about them
- Build an evidence table with one row per paper before asking for any synthesis
- Require a location_in_paper field on every extracted figure so verification is targeted
- Instruct the model to preserve hedging and never upgrade association to causation
- Resolve every DOI and open every source before it enters your reference list
- Ask the model to flag which claims are anchored to uploaded text and which are from memory
- Keep a running note of how AI was used at each stage for your disclosure statement
Frequently asked questions
Can Whizi handle very long papers?
Yes. Gemini in Whizi lists a 1M token context window, which in practice means a long dissertation, a full regulatory filing, or a set of 15 to 30 papers can sit in one conversation and be queried as a whole. That is different from summarizing a document a chunk at a time, because cross-document questions such as "where do these disagree" only work when everything is present at once.
Does Whizi cite sources?
When you use a web connected model it returns links, and when you upload documents it can point to where in your source a claim came from. Neither is a substitute for verification. Resolve every DOI, open every link, and confirm that the cited work supports the claim attached to it. Fabricated and misattributed references are the most serious failure mode of AI assisted research, and the only reliable defence is opening the source yourself.
Why not just use one model for everything?
Because the three stages want opposite behaviour. Reading a large corpus needs an enormous context window. Extraction needs rigid, boring adherence to a schema. Drafting needs careful prose that preserves academic hedging. Running all three inside one workspace means the corpus is uploaded once and the context carries across the switch, instead of being re-explained to three separate products.
Is it acceptable to use AI in academic work?
Policy varies by journal, funder, and institution, and it is changing quickly, so check the specific requirements that apply to you. The broad pattern is that using AI for extraction, drafting, and editing is generally acceptable with disclosure, while listing a model as an author is not, and you remain fully accountable for every claim and citation in the finished work. Record what you used as you go.
Can I upload unpublished or confidential data?
Whizi does not train on your conversations and each provider’s policy is reviewable before you enable that model, but your ethics approval and any data sharing agreement are the binding constraints. Human subjects data almost always needs to stay out unless your approval explicitly covers third party processing. De-identified extracts and your own drafts are the normal safe cases.