Using Perplexity and Deep Research for Academic Work — How to Avoid Fake Citations

You open Perplexity, type in your research question, and get a perfectly formatted answer with clickable citations. It looks amazing. You paste a few sources into your bibliography, hit submit, and then — months later — your professor asks how you could have cited a paper that doesn’t exist.

This isn’t a hypothetical scenario anymore. A 2026 study published in The Lancet found that 1 in every 277 academic papers now cites at least one fabricated reference. That number is up from 1 in 2,828 just three years ago. The surge is being driven, in part, by students and researchers using AI search tools as primary research assistants — tools that can hallucinate citations with terrifying confidence.

The good news: You can use these tools safely. You just need to follow a verification workflow. This article shows you exactly how — with tool-specific data, concrete steps, and a comparison of which AI search tools are actually safe for academic work.

The Fake Citation Problem Is Real — and It’s Getting Worse

Before we talk about how to use AI tools for research, let’s look at why verification matters.

What Does “Fake Citation” Actually Look Like?

Fake citations fall into four distinct patterns, each with a different difficulty to detect:

(Source: Sourcely, “AI Hallucinated Citations: How to Spot Fake Sources Before You Submit,” 2026.)

The Fully Invented pattern is the most dangerous for academic integrity because it’s nearly indistinguishable from a real citation at a glance. That’s exactly the kind of citation that can trigger a plagiarism or academic misconduct report — and it’s the kind of citation that has led to book retractions (Springer Nature retracted a machine-learning book after verifying only 21 of 46 references).

How Bad Is It Really?

(Source: DRBench study, arXiv:2604.03173, April 2026)

The surprising finding from the most comprehensive study on this topic (DRBench, DARPA-funded, 2026): multi-step deep research agents hallucinate URLs at a higher rate than standard search-augmented chatbots. Gemini Deep Research produced 113 URLs per query with 13.3% hallucination — that’s roughly 15 fake URLs per single prompt. More citations don’t mean better sources; they mean more errors.

Why This Matters Right Now

arXiv introduced a one-year submission ban in June 2026 for authors who submit work containing “incontrovertible evidence” of unchecked AI-generated content — including hallucinated references. The ban applies to both first-time and repeat submissions. This isn’t theoretical guidance; it’s actively enforced.

Chatbots vs. Academic Search Tools: The Critical Distinction

Not all AI research tools are created equal. Before you pick one, you need to understand the fundamental difference between two categories of tools.

Category 1: General-Purpose Chatbots (ChatGPT, Claude, Gemini)

These tools predict text. When you ask them for citations, they generate plausible-looking references from their training data patterns — they do not look up databases. This is the root cause of hallucinated citations. The model has never seen the paper you’re asking for; it’s making up a citation that sounds like it should exist based on the statistical patterns in its training data.

Hallucination rate: 3–13% of URLs (DRBench).

Category 2: Dedicated Academic Search Tools (Elicit, SciSpace, Consensus, Scite)

These tools use Retrieval-Augmented Generation (RAG). They first search a scholarly database (OpenAlex, Semantic Scholar, etc.), then use the retrieved papers to generate an answer, and finally verify every citation with code (not the model). A script confirms that each cited paper matches a retrieved record. This deterministic, code-based verification prevents the model from inventing non-existent references.

A 2026 test by the Singapore Management University Libraries found that 93 out of 100 references from academic search tools resolved to confident matches — and none were fabricated. The 7 flagged records were all false positives caused by minor metadata errors (wrong year, different title version, missing DOI), not hallucinated papers.

In short: academic search tools do not fabricate references. Chatbots do.

Perplexity AI: Strengths, Weaknesses, and How to Verify Its Sources

Perplexity sits somewhere between the two categories. It’s not a pure chatbot (it cites sources), and it’s not a pure RAG academic tool (it sources from the open web, not from a curated scholarly index).

Where Perplexity Excels

• Multi-source synthesis: Pulls from hundreds of documents to create balanced summaries
• Deep Research mode: Autonomous multi-step searches across hundreds of documents — great for regulatory documents and policy analysis
• Transparent citations: Every claim is linked to a numbered source you can click

Where Perplexity Struggles

• Citation pointing errors: Sometimes links to a homepage instead of the specific article
• Over-confident synthesis: Produces speculative summaries when dealing with forward-looking topics
• Paywalled content: Relies on secondary reporting when the actual academic source is behind a paywall
• 37% error rate in answers (Tow Center study)

The 5-Step Perplexity Verification Workflow

Step 1: Click every single citation

Don’t skim. Click each numbered link and confirm the author, title, and date match what you need. A surprisingly common error: the citation links to a blog summary or mirror rather than the original academic source.

Step 2: Verify the claim is supported

A real paper doesn’t mean the paper supports your claim. Read the specific paragraph or section Perplexity cited. If the source doesn’t actually say what Perplexity claims, it’s a citation faithfulness error — the paper exists, but it’s being misrepresented.

Step 3: Prefer primary sources over secondary summaries

If Perplexity cited a news article about a research paper, try to find the original paper. Perplexity’s summaries of academic research can introduce interpretation errors, especially on complex topics.

Step 4: Record permanent identifiers

Copy the DOI or a stable library permalink from the original source. Never use Perplexity’s generated link as your permanent identifier — it could break, change, or point to the wrong page.

Step 5: Keep a provenance log

Document: the query you ran, the timestamp, which sources you used, and when you verified them. This is critical if your professor or journal ever asks about your research process.

Which AI Search Tool Should You Use?

If you’re going to use AI for academic research, you should choose the tool that minimizes fake citations while maximizing source quality. Here’s how the major options compare:

(Data: DRBench 2026, SMU Libraries test 2026, Tow Center study, independent tool analysis)

The clear recommendation: If your goal is finding real academic sources without fake citations, use a dedicated academic search tool (Elicit, Consensus, SciSpace, or Scite). Use Perplexity for supplementary research on recent developments, but always follow the 5-step verification workflow above. Avoid Gemini Deep Research for citation-heavy research — its hallucination rate is more than double ChatGPT’s and nearly triple Perplexity’s.

The arXiv Policy: Why Verification Is No Longer Optional

In June 2026, arXiv introduced a policy that makes AI citation verification a practical requirement, not just an academic best practice.

The rule: Authors face a one-year ban if their submission contains “incontrovertible evidence” of unchecked AI-generated content. The three specific triggers are:

1. Hallucinated references (non-existent or fabricated citations)
2. AI-generated meta-comments left in the text (the model added editorial notes the author didn’t write)
3. Placeholder data never replaced with real results

After a ban, any further arXiv submission must first be accepted at a peer-reviewed venue — essentially forcing the author to prove their work is legitimate.

Why this matters for students: Even if you’re not submitting to arXiv, many university-level research assignments and thesis requirements now explicitly reference the same standards. If a paper gets flagged for AI-hallucinated citations at the undergraduate level, the consequences include failing grades, academic misconduct hearings, and — in severe cases — expulsion.

Practical Tips for Students Using AI for Research

Here’s what I’d recommend if you’re already using (or planning to use) AI search tools for your academic work:

Do

• ✅ Use AI tools as a discovery and scoping tool, not a citation source
• ✅ Always verify citations before adding them to your bibliography
• ✅ Prefer dedicated academic search tools (Elicit, Consensus, Scite) over general chatbots
• ✅ Record DOIs and stable links, not AI-generated URLs
• ✅ Keep a research log for transparency
• ✅ Use AI for brainstorming and drafting, then verify with primary sources

Don’t

• ❌ Never copy-paste an AI-generated bibliography without verification
• ❌ Don’t trust high citation counts — more citations = more errors (Gemini Deep Research averages 113 URLs per query)
• ❌ Don’t use AI tools for topics where the source material is paywalled
• ❌ Don’t cite an AI tool itself as a source — always attribute to the underlying original material
• ❌ Don’t assume a clickable link means the source is real or relevant

What We Recommend

If you’re using AI for academic research, here’s my honest assessment:

For literature review and source discovery: Use Elicit or Consensus. They’re the safest option — essentially zero hallucinated references, backed by scholarly indexes, and designed for academic workflows.

For understanding recent developments or policy: Perplexity is genuinely useful for this — just run the 5-step verification workflow I outlined above. Don’t skip steps.

For deep research with citations: Avoid Gemini Deep Research and be cautious with ChatGPT’s Deep Research. The DRBench study shows deep research agents hallucinate URLs at a higher rate than standard chatbots, because the multi-step synthesis process “blends patterns from retrieved pages, yielding plausible but nonexistent addresses.”

For your bibliography: Never, ever use an AI-generated reference list without verification. Even the safest tools can return minor metadata errors (wrong year, different title version). A quick CrossRef or Google Scholar search will catch every error.

FAQ

Is Perplexity AI detectable by Turnitin? Yes, Turnitin can flag content generated by Perplexity, especially if you use it to draft text or rewrite sections. The tool is designed to detect AI writing patterns.

Does Perplexity cite sources? Yes, Perplexity provides numbered, clickable citations with every answer. However, as documented in multiple studies, these citations can sometimes point to the wrong source or to secondary summaries rather than the original material.

Is Perplexity a reliable source of information? It’s a useful discovery tool, but not a primary source. You should never cite Perplexity directly — always trace citations back to the original academic material and verify them independently.

Is Perplexity a good AI for academic research? Yes, with caveats. It excels at policy review, literature scoping, and finding recent developments. For pure source discovery, dedicated academic tools (Elicit, Consensus) are safer. For all uses, always verify citations.