A Tow Center for Digital Journalism study, reported by Nieman Lab in March 2025, tested eight AI search tools and found that more than 60 percent of queries failed to return the correct information, with Perplexity failing least at 37 percent and Grok 3 Search most at 94 percent. If you use AI for research, the question is not whether it will sometimes be wrong, but how you catch it when it is.
This guide gives you a practical workflow for using AI for research without hallucinated facts: why models guess according to OpenAI’s research, what the Tow Center found about citations, and a step by step method to verify every claim before you use it.
Why AI Hallucinates
OpenAI’s research on hallucinations defines them as plausible but false statements produced with confidence. It argues that models hallucinate because standard training and evaluation methods reward guessing over acknowledging uncertainty. It compares accuracy only scoring to a multiple choice test with no penalty for wrong answers, where a lucky guess can only help and a blank answer earns nothing.
OpenAI also says pretraining teaches a model to predict the next word from text with no true or false labels, so arbitrary low frequency facts, such as a person’s birthday, cannot be predicted from patterns alone. It concludes that accuracy will never reach 100 percent, but that hallucinations are not inevitable because models can abstain when uncertain.
For a plain language explanation of the problem, see our explainer on AI hallucinations. To understand one technical fix, read our guide to retrieval augmented generation as a fix for AI hallucinations.
How Accurate Are AI Search Tools? The Tow Center Findings
Nieman Lab reports that the Tow Center tested eight AI search tools: ChatGPT Search, Perplexity, Perplexity Pro, Gemini, DeepSeek Search, Grok 2 Search, Grok 3 Search and Copilot. More than 60 percent of queries failed to return the correct information.
| Finding | Detail from Nieman Lab |
|---|---|
| Overall failure | More than 60 percent of queries failed |
| Lowest failure rate | Perplexity at 37 percent |
| Highest failure rate | Grok 3 Search at 94 percent |
| Paid versus free | Grok 3 Search, a premium product, scored worse than the free Grok 2 |
| Fabricated links | Gemini and Grok 3 gave more fabricated links than correct ones |
| Dead links | Grok 3 pointed users to 404 error pages 154 times |
Nieman Lab’s article cites both 200 tests and 1,600 queries, so we have not quoted a test count. The practical lesson stands either way: AI answers and AI citations need checking.
A Step by Step Workflow for AI Research
Step 1: Use AI to Map the Topic, Not to Decide the Facts
Ask for an overview, a list of key terms and the questions experts argue about. Treat the output as a reading list to follow up, not as the final answer.
Step 2: Ask for Sources and Open Every One
Ask the tool to name sources, then click through each link. The Tow Center results show fabricated and dead links are real risks, so a source you have not opened is not a source.
Step 3: Check the Claim Against the Page
Find the sentence on the page that supports the claim. If you cannot find it, discard the claim. This takes a minute per fact and prevents most of the damage.
Step 4: Cross-Check Key Numbers
For numbers, dates and names, find a second source, ideally the original dataset, filing or study. If two sources disagree, report the disagreement and say which you trust.
Step 5: Record What You Verified
Keep a simple list of claim, source, link and date checked, which also makes it easy to update later.
Prompts That Reduce Guessing
OpenAI’s research says models are rewarded for guessing, so give the model explicit permission to say it does not know. Our guide to prompt engineering techniques covers the wider approach.
| Prompt habit | Example wording | Why it helps |
|---|---|---|
| Permit uncertainty | If you are not sure, say you do not know | Counters the pressure to guess |
| Ask for evidence | Quote the sentence from the source that supports this | Makes unsupported claims visible |
| Limit the sources | Use only the documents I have pasted | Reduces invented material |
| Separate fact from inference | Label what is stated in the source and what you are inferring | Shows where judgement entered |
| Ask for dates | Give the publication date of each source | Helps spot outdated material |
These habits reduce errors but do not remove them, so the verification steps still apply.
Red Flags That an Answer May Be Hallucinated
| Red flag | What to do |
|---|---|
| A citation you cannot find | Search for the title and author, and discard if it does not exist |
| A link that returns an error | Treat the claim as unverified |
| A very specific number with no source | Find the original source or leave it out |
| A quote you cannot locate | Do not use it |
| Confident tone about a niche topic | Check twice, since rare facts are where models guess |
Our guide to how AI content citations work explains why citations from AI tools can be unreliable.
Using AI for Research Responsibly
If you publish research, you are responsible for every fact in it. Disclose where AI helped, link to the primary sources you verified and keep a record of your checks. For academic work, follow your institution’s rules on AI use.
