Why ChatGPT Keeps Making Up Sources, and How to Get Verifiable Answers

Share

13 min

The short version

ChatGPT invents citations because it predicts plausible text rather than checking real sources. This guide explains why, walks through documented cases of fake references, and shows three workflows. ChatGPT Search, Deep Research, and file uploads, that produce answers you can actually verify.

Summarize with AI

ChatGPT

OpenAI

Gemini

Google

Perplexity

AI Search

Editorial Disclosure & Affiliate Notice

This content is published for informational and educational purposes only. It is not intended as a substitute for professional, legal, financial, or medical advice. AIUnpacker is funded through clearly labeled sponsored articles and reviews, relevant paid link placements, and affiliate commissions where disclosed. Commercial partnerships help fund the free editorial library. Sponsored work is labeled, and a payment does not change the verdict.

  • For educational purposes only. Nothing here should be taken as a guarantee, recommendation, or professional recommendation.
  • AI-assisted editing. Drafts are produced with AI assistance and get a human pass before they go up.
  • Opinions are our own. Also, we are not affiliated with most tools we cover unless explicitly stated.
  • Information may be outdated. Verify pricing, features, and policies directly with the vendor.
  • Last reviewed: . Published .

Read more on our About page, Terms and Editorial Policy.

ChatGPT makes up sources because it generates answers by predicting likely text, not by checking a database of real papers, cases, or links. Ask it for references from memory and it will produce author names, titles, and page numbers that look perfect and don’t exist. The fix is to stop asking it to recall and make it look things up instead: ChatGPT Search for quick cited answers, Deep Research for full citation-backed reports, and file uploads or connectors when the answer must come from your own documents.

Why does ChatGPT make up sources in the first place?

Because a language model’s training rewards giving an answer (any plausible answer) over admitting it doesn’t have one.

Hallucination is OpenAI’s own term for when “the model produces responses that are not factually accurate,” and the help center’s examples include “fabricated quotes, studies, citations or references to non-existent sources.” The same page admits ChatGPT “can sound confident, even when it’s wrong,” and its advice is blunt: “Confidence isn’t reliability.”

The mechanics are simple once you see them. ChatGPT predicts the next word based on patterns in its training data, so when you ask for “three peer-reviewed studies on X,” it doesn’t search a library: it generates text shaped like the citations it has seen. The citation format is easy; the citation existence is not something a next-word predictor checks.

OpenAI’s researchers said this out loud in the September 2025 paper Why Language Models Hallucinate. Training and benchmarks reward guessing and penalize “I don’t know,” so models learn to bluff. The paper ties unavoidable errors to what it calls the “singleton rate”: facts that appear only once in training data, like obscure papers and birthdays, which a model can’t reliably learn and will happily confabulate. The honest answer costs points; the invented one might slide through.

And without tools, there is no ground truth anywhere in the loop. The model has no database of DOIs to check against. A 2023 study in Scientific Reports by Walters and Wilder made the scale concrete: across 636 citations ChatGPT generated for 42 topics, GPT-3.5 fabricated 55% and GPT-4 about 18%, and 43% of GPT-3.5’s real citations contained substantive errors.

Fake citations have already caused real damage

This isn’t a theoretical flaw: it has sanctioned lawyers, embarrassed researchers, and seeded fabricated references through the scientific literature.

The landmark case is Mata v. Avianca in New York. Lawyers filed a brief citing six court cases that ChatGPT had invented, complete with quotes. In June 2023, Judge P. Kevin Castel fined the lawyers and their firm $5,000 and found they had acted in bad faith, per the New York Times and Legal Dive. In September 2025, California’s Second District Court of Appeal went further: it fined attorney Amir Mostafavi $10,000 after finding 21 of 23 citations in his brief were fabricated: CalMatters called it a historic sanction.

The academic literature shows the aggregate effect. A study led by Maxim Topaz of Columbia University and published in The Lancet analyzed more than 2 million papers and 97 million citations, per STAT News’s May 7, 2026 report. It found roughly 4,000 fabricated citations across about 2,800 papers, with the rate climbing from 1 in 2,828 papers in 2023 to 1 in 458 in 2025. The Scientist’s coverage confirmed the early-2026 figure: 1 in 277.

Fabricated references in biomedical papers went from 1 in 2,828 in 2023 to 1 in 277 by early 2026: a more than tenfold rise in three years (The Lancet, per STAT News).

The overall picture isn’t improving fast either. The Stanford HAI 2026 AI Index ran a new accuracy benchmark across 26 top models and found hallucination rates ranging from 22% to 94%, with GPT-4o’s accuracy collapsing from 98.2% to 64.4% under the stricter test. The report also found that models “struggle to tell the difference between knowledge and belief”: performance tanks when a false statement is framed as something the user believes.

Newer models help. OpenAI reported GPT-5 made roughly 45% fewer factual errors than GPT-4o and about 80% fewer than o3 with browsing enabled, and the GPT-6 announcement claims Sol makes “about half as many mistakes as its predecessor.” But “fewer” is doing a lot of work in both sentences. The fix isn’t hoping the model guesses better; it’s changing what the model does.

Plain chat vs. Search vs. Deep Research vs. file uploads

The single biggest predictor of whether your citations are real is which mode you used. Here’s how the four options differ, per OpenAI’s documentation:

Plain chatChatGPT SearchDeep ResearchFile upload / connectors
Where facts come fromTraining memoryLive web pagesLive web, files, connected appsDocuments you provide
CitationsOften inventedLinked sources you can openInline citation links plus a source listAnswers tied to your text
SpeedSecondsSecondsMinutes to tens of minutesSeconds to minutes
AvailabilityAll plansAll plans, even logged-outLimited on Free; full on Plus/ProFree: 3 uploads/day; audio needs paid plan
Best forBrainstorming, draftsQuick current factsReports, comparisons, literature reviews“What does my document say?”
Main riskConfident fabricationsWrong, stale, or missing linksSlow; still needs quota; misreads possibleScan and table errors

Plain chat is where fabrication lives, because there’s nothing to check against. The other three modes all trade speed for grounding. OpenAI’s own guidance on Deep Research says it plainly: for urgent answers, use search or standard chat: deep research is slower but more thorough.

How to stop ChatGPT making up sources: three workflows that work

Each workflow below fits a different job. Pick by stakes, not by habit.

Workflow 1: Force a cited answer in a normal chat (2 minutes)

Best for quick factual questions where a couple of real links settle it.

  1. Start a chat and pick Search from the tools menu, or type ”/” and choose it. On any plan. Search is included on Free and works even logged out.
  2. Say what kind of source you’ll accept: “Answer using news outlets or official sources published in 2026.”
  3. Ask for receipts up front: “For each claim, give the exact URL, publication date, and a one-line quote.”
  4. Add the escape hatch: “If you can’t verify something, say so instead of guessing.”
  5. Open every link. Confirm the quote is really there and says what ChatGPT claims. This step is non-negotiable. OpenAI’s help center warns that search citations “can be incomplete, outdated, or incorrect.”

Workflow 2: Use Deep Research for a report you can defend (15-30 minutes)

Best for decisions that need many sources: a market scan, a literature review, a buy-versus-build comparison.

  1. Select Deep research from the tools menu. It’s on every plan with limits: Free gets a small monthly allowance, Plus gets more, Pro gets the maximum. (OpenAI stopped publishing exact counts; the last published figures in 2025 were roughly 25 tasks a month on Plus and 250 on Pro.)
  2. Describe the outcome you need, the timeframe, and what sources count. Specificity here shapes everything downstream.
  3. Answer its clarifying questions, then edit the proposed plan before it runs. This is your one cheap moment to redirect it.
  4. Optionally connect Google Drive or SharePoint so it can also read your files, or restrict it to specific domains under Sites.
  5. Let it browse: you can watch the activity and interrupt. When the report lands, check the sources list at the bottom, spot-check the inline citations, then export to Markdown, Word, or PDF.

The reports open fullscreen with a table of contents and a dedicated sources section, so verification is built into the format. It can still misread a source (real links, wrong interpretation) so read the citations against the claims, not just for existence.

Workflow 3: Ground the answer in your own documents

Best when the question is “what does this contract/report/paper actually say?”: the setup where hallucination is easiest to eliminate.

  1. Upload the document (up to 512 MB for most files, per the file uploads FAQ) or connect Google Drive or SharePoint under connectors. Free users get 3 file uploads a day.
  2. Ask it to quote before it summarizes: “Quote the exact passage that supports each point, with the page or section.”
  3. Ask targeted questions (“What does section 4 say about termination?”) rather than “what is this document about?”: vague prompts invite the model to fill gaps from memory.
  4. Spot-check the quotes against the original. For scanned PDFs, remember the model is reading extracted text; garbled tables are a known weak spot.
  5. Keep the file in the chat while you iterate. Once the document scrolls out of context, the model drifts back toward general knowledge.

That’s the honest trade-off: grounded answers are only as good as the file and the extraction. The file uploads FAQ itself warns that transcripts “may contain errors, and speaker identification may be unreliable.”

Prompt rules that cut fake citations

The workflows set up the machinery; your wording decides whether ChatGPT uses it. Community testing and guides like Prompt Architects’ citation walkthrough converge on the same rules:

  • Tell it to search, not remember. “Search the web and answer only from results” beats “please include sources” every time.
  • Demand exact URLs and DOIs. Fabricated references can’t survive a link you can click. A DOI that doesn’t resolve is the clearest giveaway of an invented source.
  • Require direct quotes with each claim. A quote forces the model to have actually read the page.
  • Name acceptable sources. “Peer-reviewed journals, government agencies, or major news outlets only” filters the junk before it starts.
  • Give permission to be uncertain. “If you can’t verify a source, say ‘I couldn’t verify this’: do not invent one” sounds soft, but it gives the model an out that guessing doesn’t.

One example prompt that combines them: “Search the web for recent (2025-2026) data on freelance market growth. Only use official statistics agencies or major outlets. For each stat: exact URL, publication date, one-line quote. If a stat can’t be verified, list it as unverified. No invented sources.”

What ChatGPT’s citations still get wrong

Even with search on, verification is your job. OpenAI’s help pages repeat it, and the failure modes are worth knowing by name.

A real link isn’t proof of a real claim: the model can summarize a page it only skimmed or attach a citation to the wrong sentence. Search results depend on sites allowing OpenAI’s crawler, so relevant sources can be missing entirely, and the help center notes ranking “is not guaranteed.” Deep Research burns quota and takes minutes, which is why OpenAI recommends saving it for questions that deserve the wait. And pushing back on a confident wrong answer can backfire: the Stanford HAI knowledge-versus-belief finding explains why correction loops sometimes make answers worse.

If ChatGPT’s limits sting, the alternatives each trade differently. Perplexity is built around linked citations and suits quick source-finding. Google’s Gemini offers Grounding with Google Search, tying responses to live results. Anthropic’s Claude pairs a web search tool with inline citations. ChatGPT’s edge is that search, deep research, and file analysis live under one roof.

What to do next

Match the mode to the stakes. Quick fact? Search, with links opened. Decision that costs money or credibility? Deep Research, with the source list actually read. Question about your own documents? Upload and demand quotes. Adopt one rule everywhere: no citation gets used until you’ve clicked it and read the passage. That habit alone would have prevented every case in this article.

One caution worth stating clearly: for legal, medical, or financial work, AI verification isn’t a substitute for a qualified professional checking sources: the courts have already made that point the expensive way.

FAQ

Why does ChatGPT keep making up sources?

ChatGPT generates answers by predicting plausible text from training data, not by looking citations up in a database. OpenAI’s 2025 research paper “Why Language Models Hallucinate” shows that training and benchmarks reward guessing over admitting uncertainty, so the model fills citation-shaped gaps with realistic-looking but invented author names, titles, and page numbers. Without tools, it cannot check whether a source exists.

Which ChatGPT mode gives the most verifiable citations?

Deep Research gives the strongest citations: it browses the web in multiple steps, then produces a structured report with inline citation links, a source list, and export to Markdown, Word, or PDF. For quick questions, ChatGPT Search attaches clickable source links. Plain chat without tools is the least reliable, because the model is answering from memory.

How do I get ChatGPT to cite real sources?

Make it look sources up instead of recalling them. Select Search in the tools menu or type ”/” and ask it to answer from the web, then require exact URLs, publication dates, and direct quotes for each claim. Add “If you can’t verify a source, say so: don’t invent one.” Then open every link and confirm the quote actually appears in it.

How much does ChatGPT cost to get verifiable, cited answers?

Nothing. ChatGPT Search is included on the Free plan, and Free users get limited Deep Research access. Go costs $8/month, Plus $20/month, and Pro spans three usage tiers at $100, $200, and $500/month (as of November 2026). Higher plans mainly raise limits and unlock stronger models, not a fundamentally different citation system.

What are the main alternatives to ChatGPT for cited answers?

Perplexity is built around linked citations and suits quick research. Google Gemini offers Grounding with Google Search, which ties responses to live results. Anthropic’s Claude has a web search tool with inline citations plus a Citations API for grounding answers in your own documents. ChatGPT’s advantage is doing search, deep research, and file analysis in one place.

Is ChatGPT safe for academic citations?

Only as a starting point, never as the final source list. A 2023 Scientific Reports study found GPT-3.5 fabricated 55% of citations it generated, and a 2026 Lancet-published analysis traced a sharp rise in fake references in real papers to AI use. Use Search or Deep Research, then confirm every reference in Google Scholar or PubMed before it goes anywhere near a submission.

What are the main limitations of ChatGPT’s search citations?

OpenAI’s own help center warns that search results and citations “can be incomplete, outdated, or incorrect,” so a working link does not guarantee the claim comes from it. Deep Research is slow and quota-limited, and can still draw wrong conclusions from real sources. Search ranking is also not guaranteed: sites must allow OpenAI’s crawler, so relevant sources can be missing entirely.

Can I use ChatGPT’s cited answers commercially?

Yes. OpenAI’s Terms of Use state that you own the output, so using cited answers in commercial work is allowed on consumer plans as well as Business and Enterprise. But you own the risk too: publishing a fake citation in marketing, legal, or compliance material is your liability, not OpenAI’s. For anything legally sensitive, have a qualified professional verify the sources before you publish.

How can I spot a fake citation?

Check three things: does the source exist (paste the title into Google Scholar or a DOI resolver), do the details match (authors, year, page numbers), and does it actually say what ChatGPT claims. Perfect-looking formatting is not evidence. ChatGPT formats invented references flawlessly. A DOI that doesn’t resolve is the clearest giveaway.

What should I do next to get trustworthy AI answers?

Pick the tool to match the stakes: ChatGPT Search for quick facts, Deep Research for decisions that need a citation-backed report, and file uploads when the answer must come from your documents. Adopt one rule everywhere: no citation gets used until you click it and read the passage. If the stakes are legal, financial, or medical, have a professional verify the sources too.

Sources

Weekly digest

Get our weekly AI digest

The latest AI tools, prompts, and insights — short, useful, every week.

No spam. Unsubscribe anytime.

AIUnpacker Editorial

Articles are written for a specific question. Sponsored work is labeled.

About us ·More articles