Consensus AI Review: Are Its Answers Safe to Cite in a Thesis?

“Yes 100%” reads the bar under a Consensus answer about CRISPR delivery, with N equals 9 printed beside it.

The thread had read 20 papers. Nine carry a yes badge in the reference list and eleven are marked unknown, and one of those unknown papers is cited twice in the answer itself.

So the figure is 100% of the nine it could classify, not of the twenty it read.

<CTA title="Run the checks before your supervisor does" description="Five reviews read the draft itself, covering claim support, source quality, tone, proofreading and a peer review pass" buttonLabel="Try Jenni Free" link="https://app.jenni.ai/register" />

TLDR

What Consensus does

Consensus is a search engine that answers a research question in prose and shows the studies underneath, rather than returning a list of links for you to read.

The box asks you to ask the research, with a corpus picker, a Deep toggle, a filter, a microphone and the submit arrow set along its edge. Two of those are worth knowing before the first question:

• Deep - the mode Consensus describes as the more comprehensive and thorough literature review of many studies.

• Filter - sits beside the box, so a question can be narrowed before it runs rather than after.

A left rail holds New Thread, Home, My Library and History, with Paper search and a Citation Graph listed under Tools. The thread is auto-titled and lands in that history, so an answer can be found again.

The stage this fits

Consensus earns its place at the moment a supervisor asks whether anyone has actually shown a thing. It returns a direct answer and the studies under it in well under a minute.

It is less suited to the moment after that, when the claim has to be written into a chapter with a reference that will survive a reference check. That gap is what the rest of this page measures.

What one question returned

The question put to it asked whether lipid nanoparticles can carry gene editing to tissues outside the liver, taken from the hedge in the CRISPR passage that runs through these reviews.

Consensus does not hide its pipeline. Above the answer it printed the funnel it ran, from the number retrieved down to the number it actually read.

The pipeline printed above the answer, with the two searches it ran and the number of papers it read.

Two searches went out, returning 79.3 thousand and 105.8 thousand hits. From 185.1 thousand retrieved, 58 were judged eligible and 20 were included and read as abstracts and PDFs.

The answer arrived about 38 seconds after the question. It opens with a yes, naming the lungs, the spleen and muscle as the tissues reached.

Below that sits a short table headed Evidence for Extrahepatic CRISPR-Cas9 Delivery, which grades its own claims. One is marked strong and two are marked moderate, each with the author names and years that support it.

The prose also pulls figures straight out of the papers, captioned and credited to the study they came from, so a flow cytometry panel from one 2026 paper appears inside the answer.

<ProTip title="📉 Read the funnel first:" description="The retrieved number is a search count, not an evidence count. The number that matters is the one it says it read, which was 20 here" />

What the meter counts

Under the answer sits the Consensus Meter, a bar reading Yes 100 percent with Possibly, Mixed and No all at zero, and a small N equals 9 beside the heading.

The answer and the meter, with the N value that says how many papers the percentage covers.

That N is the whole story of the number. The reference list holds 20 entries, and each one carries a badge: nine read yes and eleven read unknown.

Unknown is not a quiet way of saying no. It appears to mean the paper was read but could not be sorted into a yes or a no for this question, and at least one of them still feeds the prose.

A reference marked unknown, which keeps it out of the meter while its content stays in the answer.

The fourth reference is a 2022 paper marked unknown with seven supporting quotes, and it is cited twice in the body of the answer. So the meter and the text are drawing on different sets.

A thesis sentence that said 100 percent of studies agree would therefore be wrong twice over, once on the denominator and once on what the unknown papers contributed.

<ProTip title="🔢 Quote the N, never the percentage alone:" description="Write the number of classified papers into your own sentence. A percentage with no denominator is the single easiest thing for an examiner to pull apart" />

Where the qualifier sits

The headline sentence says lipid nanoparticles can deliver gene editing to those tissues. It does not say where that evidence comes from.

One line further down, the first figure caption describes the evidence as extensive and preclinical. That word is the difference between a claim about laboratory models and a claim about treating people.

Anyone lifting the headline into a chapter without reading the caption would overstate the field, and the correction is one adjective rather than a rewrite.

<ProTip title="🔬 Carry the qualifier across:" description="If the evidence is preclinical or in animal models, put that word in your own sentence. The qualifier often sits in a caption rather than the headline" />

What a reference gives you

Each entry in the list is a card rather than a line. It carries a one-sentence key takeaway, the badges, the year, a citation count, the first author, the journal and a link to the PDF.

The key takeaway is a summary line rather than a quoted sentence, so it reads cleanly and still needs checking against the paper it stands for.

Two of the badges are about study quality rather than the answer. Four of the 20 entries were tagged as being in a rigorous journal, and eight were tagged as non-randomised experiments.

Opening a card gives five tabs, and the useful one is Evidence. It prints each supporting quote in full and labels where it came from, which on the top paper was the abstract in both cases.

That labelling is the most useful thing on the card, because a quote drawn from an abstract carries less weight in a chapter than one drawn from a results section.

The quote count differs from paper to paper. The top paper is supported by two, while that unknown fourth entry carries seven.

The Metadata tab lists the name, type, year, authors, journal, volume and DOI. It does not list the issue number or the page range, which matters at the next step.

The reference it generated, checked line by line

A quote icon in the row of actions at the foot of the card opens Cite this finding, which offers APA, MLA, Chicago, Harvard, BibTeX and AMA with a copy button.

The generated APA entry for the top paper, as Consensus produced it.

The APA entry it produced for the top paper was checked against the publisher page, and four things would come back marked:

• The journal name - printed as Nature nanotechnology, where APA 7th wants the journal title in title case.

• The issue number - missing. The article is volume 15, issue 4.

• The page range - printed with a spaced hyphen rather than an en dash.

• The last author - given one initial where the author has two.

None of these change which paper is cited, and the DOI in the entry is correct. They are the kind of thing a supervisor marks in a margin, and they have to be fixed by hand once the entry is in a document.

<ProTip title="✍️ Fix the entry where it lands:" description="Paste a generated reference into your reference manager or editor and correct the journal case, the issue and the dashes there, not in six places later" />

What Consensus costs

Three individual tiers are listed on its pricing page, with monthly and annual figures for each.

October 2026 pricing, read on 2 October from consensus.app/pricing.

The Pro tier is the one a thesis writer would weigh, because the free listing meters the AI answers at ten a month while Pro lists them as unlimited.

Team and enterprise pricing sits behind a separate tab, and both paid tiers add API and MCP usage with a per-use charge above the included allowance.

Deep reviews are metered separately from messages on every tier, at three, fifteen and two hundred a month as the tiers rise.

Where Consensus sits against the others

Jenni - the two tools answer different questions. Consensus tells you what the literature says, while Jenni works on the sentence you are writing and the sources under it.

On the same paragraph, Jenni's five reviews run on the draft itself, and Claim confidence returned four suggestions with a source and a quoted passage attached to each one.

Its citation search also files the reference as the citation goes in, so the entry does not have to be pasted and repaired separately.

Elicit - it returns a table of papers and extracted figures rather than a single answer, and the Elicit review covers that test.

One number differs between the two in a way worth knowing. The same 2020 paper showed 1,992 citations in Consensus and 1,602 in Elicit, so a citation count is a database figure rather than a fact about the paper.

Are its answers safe to cite in a thesis?

The findings are, once the N is read and the paper is opened. The quotes are verbatim, they are labelled with where they came from, and the DOI resolves to the right article.

What is not safe is the summary sentence on its own. The percentage needs its denominator, the headline needs the preclinical qualifier, and a key takeaway is a summary line rather than something to quote.

The generated reference is not, not without a pass by hand. Keep the DOI and correct the entry before it goes near a reference list.

<CTA title="Set the style once and every entry follows" description="Cite in APA, MLA, Chicago, IEEE or Harvard, or search the Zotero style repository, and the reference list matches throughout" buttonLabel="Start Writing Free" link="https://app.jenni.ai/register" />

Use Consensus to find out whether the evidence is there, then write the chapter in Jenni, where the claim, the source and the review sit in the same document. More comparisons are in the tool reviews.