Elicit AI Review: Does It Find the Right Papers for a PhD Thesis?

Elicit was given one sentence from a draft, the claim that a CRISPR trial cut serum transthyretin by 87%, and asked to find the papers behind it. The 2021 NEJM report came back first, and the 2022 reprint that carries the same title never appeared.

A column added to that table asked for the reduction at each dose. Elicit returned 52% against the 0.1 mg per kilogram group and 87% against 0.3, and wrote not reported, rather than a guess, where a paper carried no dose figures of its own.

<CTA title="Keep every paper you find in one library" description="Search your own papers by meaning or by keyword, across titles, authors, abstracts and the full PDF text" buttonLabel="Try Jenni Free" link="https://app.jenni.ai/register" />

TLDR

What Elicit is

Elicit is a research agent that works over a paper index, and its home page puts that index at over 138 million academic papers and 545,000 clinical trials.

The prompt box carries a mode button reading Research agent, and under it sit five chips that set what the agent will produce:

• Find papers - returns a sortable table of sources with a Summary and a Relevance column for each one.

• Draft report - writes a cited report and shows the sources it drew on.

• Build deck - the slide mode, which this test did not run.

• Strategy and Landscapes - two further modes, also left untested here.

The left rail holds New, Search, Recents and Library, so sessions and saved papers stay in one account rather than in a document.

Who Elicit is for

Elicit belongs to the reading stage rather than the writing stage. It fits the point in a thesis where the question is settled and the papers behind it are not.

It also suits anyone who needs the same figure pulled out of every paper in a set, because that is what its table columns do. It is not an editor, and the draft still has to be written somewhere else.

For a researcher with a manuscript already drafted, the use is narrower. Elicit will tell you what the field says about a claim you have made, but it will not mark up the sentence that makes it.

The test, run by run

Two sentences from the CRISPR paragraph used across these reviews went through three Elicit modes in a single session.

• The claim - serum transthyretin down a mean 87% at the higher dose after one infusion (Gillmore et al., NEJM 2021).

• The trap - a 2022 reprint carries the same title as that paper, so a search can return the wrong one.

• The figures to protect - the PubMed abstract gives 52% at 0.1 mg per kilogram and 87% at 0.3.

• How it was logged - every count and quoted line here was taken from the page, and every timing from a check against the clock.

Find papers: one claim in, four papers out

The run - the claim went in as a single sentence with the Find papers chip selected. Rows began appearing about 50 seconds after the click, and the table was complete at the 80 second check.

The result was a four-row table under the heading Papers on first-in-human systemic LNP-CRISPR therapy for transthyretin, with Source, Summary and Relevance columns.

Row one was the 2021 NEJM paper, marked Full text available, with the Relevance cell reading "Direct primary evidence for the trial and the 87% figure in the prompt." The 2022 reprint was not in the table.

The Find papers table, with the 2021 NEJM report in the first row and its relevance note alongside.

The other three rows were a 2021 commentary, a 2023 review of a cardiomyopathy cohort and a 2018 preclinical study. Every row carried the journal, the year, a citation count, a DOI chip and a full-text badge.

Beside the table, Elicit wrote out what it had found in prose, naming the trial, the six participants, the dose and the day the measurement was taken, with the citation attached to the sentence.

A Methods note under that answer says the results were curated into a source-linked collection, prioritising the primary study and adding adjacent interpretation. It is one short line about how the four rows were chosen.

Three follow-up questions were offered under the answer, including one that would look for the 12-month results and one that would export the collection as a bibliography.

<ProTip title="🔎 One claim per search:" description="Paste the single sentence you need to support rather than the whole paragraph. The narrower the claim, the more exactly the top row matches it" />

Add a column: the doses it got right

The run - the Add a column button opens a panel with a Name field and an Instructions field. The column was named TTR reduction by dose, and the instructions asked for the mean reduction at each dose with not reported where a paper gives no figure.

The column filled in under 40 seconds. Against the NEJM paper it read "0.1 mg/kg: 52%; 0.3 mg/kg: 87%.", which matches the abstract on both doses.

The added column, with the two dose figures extracted from the trial paper.

The commentary and the preclinical study both came back as not reported, which is the instruction being followed rather than a gap being filled with a guess.

<ProTip title="🧪 Name the unit in the column:" description="Write the unit and an example into the instructions field, such as a dose in mg per kg. Columns asked in plain words come back in mixed formats" />

Draft report: where it argued back

The run - the Draft report chip took the hedge from the same paragraph, which asserts that ionizable lipid chemistry is the main lever for targeting outside the liver. The report arrived about two minutes after the click.

Its progress lines are specific about scale. Elicit reported comparing evidence across 56 screened sources, then comparing direct results against delivery proxies across 10 studies.

The report answered the first half of the question yes, then declined the second half. The report says the evidence does not support calling lipid chemistry the universal or sole lever, and names the other variables that move the result.

The opening of the generated report, with a citation chip closing the summary.

It also separates three things that a careless reader would merge, which are particles reaching an organ, expression after delivery, and editing in the intended cells. Where a cited result was the weaker kind, the report says so.

Every claim in the report carries an inline citation chip naming an author and a year, and a chip that covers several sources prints a count after the name. There is no numbered reference list under the report itself.

The last section is headed Overall assessment and remaining gap, and it closes on what the evidence cannot yet settle rather than on a restatement.

<ProTip title="🧭 Give it the claim you doubt:" description="Put your own hedge into the prompt and see whether the report defends it. A tool that pushes back is doing the job a supervisor would" />

Where Elicit asks for a second look

The report panel carries a Sources view, and that view is where its limits show. It listed 13 sources for a report that ran to roughly 9,400 characters.

Of those 13, 8 were marked Full text available and 5 were marked Abstract only. So 5 of the sources behind the report were read at abstract level.

The sources view, with the badge on the first source showing it was read at abstract level.

The menu item that opens the view prints the source count beside it, so the 13 is visible without opening anything. The split between full text and abstract is not, and that takes a count.

None of this makes the report unusable. It means the badge is worth reading before a sentence from the report becomes a sentence in a chapter.

The Relevance column does similar work in the paper table. On the 2023 review it reads "Related cardiac-amyloidosis cohort context; distinct from the initial six-patient polyneuropathy report.", which is Elicit marking a row as adjacent rather than equivalent.

<ProTip title="📄 Check the badge, then the DOI:" description="Open the sources view and look for the abstract-only label. Anything you plan to quote should be read in full, from the DOI rather than the summary" />

What Elicit costs

Elicit lists two price sets on its pricing page, one for academic users and one for industry, with monthly and yearly figures for each.

October 2026 pricing, read on 2 October from elicit.com/pricing.

The academic Plus tier is the one that matters for a thesis, because Elicit lists export to RIS, CSV, BIB, PDF and DOCX there, along with five table columns at a time and search across more than 500 thousand clinical trials.

Basic covers unlimited search, unlimited summaries and chat with papers, with agent and report usage described as limited. Pro adds the systematic review workflow that screens 5,000 papers, 20 columns at a time, custom extractions and API access.

Access to Routines is a Pro line too. Elicit describes them as the agent finding new evidence, updating your work and reporting back, which is the feature for a review that has to stay current. Extracting from figures sits a tier higher again, on Scale.

Elicit against the alternatives

Jenni - the difference is where the work lands. Elicit builds a table in its own workspace, while Jenni puts the source into the document being written.

On the same paragraph, Claim confidence returned 4 suggestions in about 47 seconds, each one an inline card naming the proposed source, its citation count and its impact factor, with the quoted passage attached and accept or reject in place.

Jenni also keeps a Library that searches titles, authors, abstracts and full PDF text, and builds the reference list under the draft as citations go in.

Consensus - it answers a research question from published studies rather than handing back a table, and the Consensus review covers what its answers are worth. The same paragraph has been through several other tools in the tool reviews.

Should a PhD student pay for Elicit?

For the reading stage, yes. It found the right paper from a claim rather than a title, it pulled two dose figures out correctly, and it refused to endorse a hedge that the draft had overstated.

Keep the DOI of anything that will be quoted, and check the abstract-only badge before a report sentence becomes a chapter sentence.

<CTA title="Watch the reference list write itself" description="Cite while you draft and the reference list assembles underneath, in APA, MLA, Chicago, IEEE or Harvard" buttonLabel="Start Writing Free" link="https://app.jenni.ai/register" />

For the chapters themselves, draft the thesis in Jenni, which keeps the writing and the sources in the same file. Bring Elicit in at the start of a chapter, while the reading is still being done.