How to reduce AI hallucinations and check fake citations
- AI
- Guides
- AI assistants
- Research
- Fact checking
- Written by
- Bhargav Tibadiya
- Published
- Reading time
- 8 min read
You reduce AI hallucinations by changing two things: how you ask, and how you check. On the asking side, give the model permission to say it does not know, ground its answers in documents you supply, and require a quote behind every claim[1]. On the checking side, treat every citation as unverified until you have opened it, because a confident reference can point to a paper that does not exist[2]. No technique eliminates hallucinations entirely, so important claims still need your own validation[1].
This guide is a practical routine for people using AI at work, based on Anthropic's hallucination-reduction documentation and a university library's verification guide, both checked on November 5, 2026. It covers citations, document-based answers and uncertain claims. It does not promise a hallucination-free setup, because Anthropic's own guidance says its techniques do not eliminate hallucinations entirely[1].
Why does AI invent references?
Neither source explains the mechanism in detail, so treat this paragraph as general background rather than a sourced claim. A language model generates text that fits a pattern; it does not look a reference up in a record. A citation has a recognizable shape: author names, a year, a title, a page range. Text with that shape can be produced whether or not the paper exists.
That is why invented citations can look convincing. The BU library's guidance is blunt about the consequence: generative AI tools can produce incorrect or entirely made-up information, including fake citations, and if you cannot find a source the AI gave you, it may not exist[2].
The same mechanism produces wrong dates, misattributed quotes and confident numbers with no origin. Treat all of these as the same failure in different clothing: fluent text that was never tied to a record.
In my judgment, some categories deserve extra suspicion because they are easy to state and hard to eyeball: exact statistics, direct quotations, publication dates, version numbers, legal clauses and named studies. A sentence shaped like "a 2025 study found that 73 percent of teams..." is just as easy to write for an invented study as for a real one. The more specific and convenient a detail is, the more it earns a check.
Does asking for sources fix it?
Not by itself. If you ask a model to add sources, it can invent sources with the same fluency it used for the claims. A link-shaped answer is still generated text until you open it.
Even tools that search the web and show real links need checking. The link may be real while failing to support the sentence attached to it, and the BU guide warns exactly that: visible sources do not mean the sources are good or say what the AI claims[2]. Open the link, find the claim, and check that the page actually says it.
Asking for sources is still worth doing, because it converts a vague answer into a list of specific, checkable promises. The fix is not the request. The fix is the verification you do after it.
A checking routine that catches fake citations
The BU library's routine for a source you cannot find has three steps[2]. Search the exact title in a scholarly search tool. If that fails, search by author and keywords, because AI often gets titles or years slightly wrong. If nothing turns up, do not cite it - use a source you can confirm.
For general claims, the same guide recommends the SIFT sequence: Stop and consider whether the information seems credible; Investigate the source behind the claim; Find better coverage in trusted sources; Trace the claim to its original record, such as a peer-reviewed article or dataset[2].
Notice what this routine assumes: the checking happens outside the AI. Each step moves you from generated text toward an independent record. A routine that stays inside the chat window cannot catch a confident invention, because the same process that invented the claim will invent its confirmation.
Here is the routine applied to one claim. An assistant tells you a competitor raised $40 million "according to TechCrunch, March 2026." Stop: is a funding round the kind of thing that gets reported? Yes, so it is plausible, which is exactly why it needs checking. Investigate: search the exact claim and look for the original article, not a summary of a summary. Find better coverage: do two independent outlets report the same number with the same date? Trace: find the primary record, the company's own announcement or the actual filing. If the trail ends at "the AI said so," the claim does not make it into your document. Thirty seconds for a throwaway number, a few minutes for one your decision depends on.
Prompt-side habits that reduce invention
Anthropic's documentation lists several strategies to minimize hallucinations, starting with the simplest: explicitly give the model permission to admit uncertainty, with a line such as "I don't have enough information to confidently assess this"[1]. Anthropic says this technique can drastically reduce false information[1].
For tasks with long documents (Anthropic's example threshold is over 20,000 tokens), ask for word-for-word quotes first, then have the model base its analysis only on those quotes[1]. This grounds the answer in text you can inspect instead of the model's impression of the document.
You can also make answers auditable. Require a citation for each claim, then have the model verify each claim against a supporting quote and retract any claim it cannot support[1]. A retraction is a good outcome: it is the system showing you exactly where not to trust it.
Two more habits come from the advanced section of the same page. Instruct the model to use only information from the documents you provided, not its general knowledge[1]. And when an answer looks wrong, use thinking with summarized output and review the reasoning for faulty logic or assumptions[1].
Ground answers in your own evidence
The strongest version of this workflow is an evidence packet: a small set of documents or URLs you approve, a clear question, and an instruction to distinguish what the sources say from the model's interpretation.
This pattern shows up twice in the source material. Anthropic's examples have the model extract exact quotes first and base its analysis only on those quotes, stating plainly when no relevant quote exists[1]. The library guide's Trace step sends you to the primary record for the same reason[2]. Both are the same idea at different scales: keep the evidence and the interpretation visibly separate.
The practical benefit is review speed. When every claim sits next to a quote you supplied, checking the answer means spot-checking the pairings, not re-researching the topic.
Can RAG remove hallucinations?
Neither source discusses retrieval-augmented generation by name, so this section is my reading of what they imply. Giving the model relevant documents gives it real text to work from, which is the idea behind Anthropic's quote-first technique. It does not follow that the model cannot still misread a document, combine two sources incorrectly, or fill a gap the documents leave open.
Even with documents supplied, Anthropic's recommended pattern is quotes first, analysis second, and an explicit "No relevant quotes found" when the material does not answer the question[1], and the page closes by saying its techniques do not eliminate hallucinations entirely[1].
So treat retrieval as a way to reduce risk, not remove it. The checking routine still applies, and it gets easier, because every claim should trace to a document you can open.
When agreement is not verification
Running the same prompt several times and comparing outputs is a documented technique: Anthropic calls it best-of-N verification and says inconsistencies across outputs could indicate hallucinations[1]. Read that rule in one direction only. Inconsistency is a warning, but consistency is not proof. A model can repeat the same plausible mistake across runs, and two different assistants can rely on the same flawed source.
The same caution applies to research tools that search scholarly literature. The library guidance says research-focused tools such as Elicit and Consensus draw on scholarly literature, and that you should still confirm their results in the library databases[2].
The tie-breaker is always a record that exists independently of any model: the paper, the dataset, the official page, the primary document. One verified record outweighs any number of confident paragraphs.
If the claim matters and the record conflicts with what the model said, the record wins and the model's answer gets logged as a miss. Keeping a small note of the misses you catch is underrated: after a few weeks you will know which kinds of questions your tools handle well and which ones always need the manual route.
The high-stakes rule
Both sources land on the same bottom line. These techniques significantly reduce hallucinations but do not eliminate them, so critical information always needs validation, especially for high-stakes decisions[1]. The library version is shorter: generative AI is a starting point, not a source[2].
Match the checking effort to the cost of being wrong. A brainstorm for internal naming ideas needs a glance. A number going into a client report, a legal deadline, a medical detail or a citation in published work needs the full routine, every time. The model does not know which conversation matters. You do.
FAQ
Can AI hallucinations be eliminated?
No. Anthropic's documentation says its techniques significantly reduce hallucinations but do not eliminate them, and to always validate critical information. Plan a checking step for anything important rather than waiting for a perfect model.
Why does AI make up fake citations?
The BU library says AI tools can produce entirely made-up information, including fake citations. As general background, a citation has a recognizable shape, authors, year and title, and text with that shape can be produced without any record behind it.
Does asking for sources fix hallucinations?
Not by itself. The model can invent sources as fluently as it invents claims, and even real links may not support the sentence they are attached to. Open every source and find the claim in it.
Can RAG remove hallucinations?
Neither source claims it can. Anthropic says its techniques do not eliminate hallucinations entirely, and the sources do not discuss RAG by name. Keep the verification routine, and require exact quotes for document-based work.
Is a claim safe if two AI tools agree?
No. Agreement is not independent verification - both tools can share the same flawed source or repeat the same plausible error. Confirm important claims against a primary record.
What is the fastest way to check one citation?
Search the exact title in a scholarly search tool. If it does not appear, search by author and keywords, because AI often gets titles or years slightly wrong. If nothing turns up, treat the source as nonexistent.
Sources
Elsewhere on this site
The AI assistant comparison applies the same evidence-first thinking to choosing a tool. The selected work catalogue covers the products behind these engineering notes.
