Does Gemini or ChatGPT Hallucinate More?

Both Gemini and ChatGPT hallucinate — they sometimes generate claims that are plausible-sounding but not supported by their sources or your documents. Which one does it more depends on the specific task, the information type and how you prompt. Neither is reliably hallucination-free, and independent benchmarks show that results vary by task category. The useful question is not which is better overall, but how to test both on the work you actually need to do.

When you find a verified source or confirm a fact, saving it outside the conversation protects the finding. Tools like fenbs let you record accepted sources and the decisions behind them, so the next session — in Gemini, ChatGPT or any other assistant — starts from confirmed ground rather than from what the model happens to produce fresh.

What hallucination means for AI assistants

Hallucination is when an AI model generates text that is presented as factual but is not grounded in its training data or the documents you provided. Common forms include incorrect dates and statistics, invented citations, names that do not exist, and confident explanations of events that did not happen.

The word "hallucination" covers a range of errors. A model might confuse two similar people, add a plausible-sounding number that is wrong, or simply fill in a gap with a confident guess. These errors are harder to catch precisely because the tone and format of a hallucinated claim looks the same as a correct one.

How to test both models on the same task

The most reliable comparison is to run the same prompt on both models and check each output against a primary source. Choose a task where you already know the correct answer, or where the source is easy to verify: a specific statute, a published study, a known timeline of events.

1. Write the same prompt for both Gemini and ChatGPT. Do not adjust the wording for each model.
2. Ask each model to list its sources or explain where its claim comes from.
3. Open each cited source and check whether the claim appears in it.
4. Note the specific failure mode: wrong date, wrong name, invented citation, or correct claim from an unverifiable source.
5. Repeat on three to five tasks relevant to your actual work before drawing a conclusion.

Running this comparison once on an easy question tells you little. Running it on tasks that mirror your real workflow — legal research, financial analysis, technical documentation — gives you a useful signal about which model is less risky for your specific context.

Citation accuracy and source grounding

Gemini can browse the web in real time on supported plans, which gives it access to current information. It cites sources in some formats, but the quality of those citations varies. A link in a Gemini response is not a guarantee the cited page says what the response claims — always open the link and check the relevant section.

ChatGPT can also browse the web on supported plans and provides citations in its browsing mode. Its base model (without browsing) relies on training data with a knowledge cutoff, which creates a specific hallucination risk for recent events or changed information.

Gemini Notebook (formerly NotebookLM) takes a different approach: it answers only from the sources you upload and cites exactly where in those documents each answer comes from. For high-stakes research on a defined set of materials, this reduces hallucination risk significantly compared to open-ended chat. See the Gemini Notebook overview.

Where each model tends to struggle

Model hallucination patterns differ by domain and update frequency, and they change as models are updated. Some patterns that have been noted in testing:

• Both models struggle with precise numbers: dates, statistics and figures are common hallucination points regardless of which model you use.
• Legal and medical specifics are high-risk areas for both, where a plausible but wrong detail can cause real harm. Always verify against the primary source.
• Recent events are a known risk for any model with a knowledge cutoff; live-web browsing reduces but does not eliminate this.
• Citations to academic papers can be invented or partially wrong even when the surrounding claim is broadly correct.

These are tendencies based on common reports, not controlled benchmarks. Your specific tasks may show a different pattern. Test before relying on either model for consequential facts.

Reducing hallucination risk in practice

Prompt for sources explicitly: ask the model to list the documents or web pages it drew on for each claim. Then check them. Treat any claim the model cannot source as unverified. For domain-specific questions — tax law, clinical guidelines, financial regulations — verify against the primary source even when the model provides a citation, since models can mis-attribute claims.

Use grounded tools where they exist. Gemini Notebook, when fed your own documents, is less likely to invent a fact about those documents than a general chat model is. If your research is bounded — a set of case files, a collection of reports — uploading them to a grounded tool is a practical risk reduction step.

Domain-specific hallucination risks

Hallucination risk is not uniform across subjects. Some domains are harder for both models, and knowing where to be especially careful helps you target your verification effort rather than checking everything equally.

Legal questions are a high-risk area. Both models can confidently describe a law, a case or a regulation that does not exist, or get the name of a real case right while attributing the wrong holding to it. Never rely on either model for legal research without verifying every cited authority at the primary source.

Medical and clinical content carries similar risk. Dosages, contraindications, diagnostic criteria and treatment protocols can be stated confidently and incorrectly. Treat AI outputs in this domain as a starting point for research, not an authoritative reference.

Financial figures — stock prices, interest rates, revenue numbers, market statistics — change constantly. Even with live web access, a model can cite a stale figure or misread a data table. Always verify financial claims against a primary source that carries a clear timestamp.

Mathematical calculations and structured data extraction tend to produce fewer hallucinations than open-ended factual questions, because the output is easy to check. If you ask both models to extract a table of numbers from a document you have provided, errors are quickly visible. If you ask both models to describe a historical event from memory, errors are far harder to spot without independent verification.

The pattern to watch for in any domain is confident specificity: the more precisely and confidently a model states something that you cannot immediately verify, the more carefully it deserves checking. Vague answers are often a signal that the model is uncertain; precise-sounding answers are not a signal that it is certain. Confidence in tone is a style choice, not an accuracy guarantee. Apply the same verification standard to a confident-sounding claim that you would to a hedged one.

Building a verification habit into your workflow

For work where accuracy matters — legal drafts, financial analysis, technical documentation — treat every AI output as a first draft that needs checking rather than a finished product. The model's job is to produce a draft efficiently; your job is to verify the facts before you rely on them.

A practical workflow: after each session that produces factual claims, note which claims you verified, which remain unverified, and the sources you found. Save those notes with the task so the next person (or session) knows what has been confirmed and what has not. An unverified claim clearly labeled as such is far more useful than one that someone assumed was checked.

How to do this with fenbs

1. Create a free account and add your research project to fenbs.
2. Write a Decisions and rules entry for your fact-checking standard — which types of claims must be verified against a primary source before the team acts on them.
3. As you verify sources, record the confirmed link and what it supports as a context note in the relevant task.
4. Connect your AI app over MCP from Integrations. It reads the verification rules at the start of each session.
5. When a new session picks up the research, the confirmed sources and standing rules are already there.