Reducing Hallucinations in AI

by Nils Osmar

How to Get Reduce Hallucinations – and Get Reliable AI Summaries

AI tools such as Perplexity, ChatGPT, and others can be helpful in making sense of dense technical papers. I often use them to summarise academic papers, especially if the language and those papers may be unfamiliar to lay readers.

But there’s a catch: these systems sometimes hallucinate — generating confident statements that aren’t supported by the source material. If you’re using an AI to summarize scientific studies, policy papers, or detailed reports, it’s important to understand why this happens and what you can do to reduce the risk of errors.

What to Do

When asking an AI to summarize a paper, try pasting this directly into the prompt field, followed by the URL:

“Summarize the following article using only its text as a source. Do not add outside information. If a claim is not stated in the article, write ‘not found in the text.’ Provide citations or short quotes for each key point. Aim for accuracy first, clarity second, and do not guess.”

---

Why This Works – and Why AIs Sometimes Hallucinate

Hallucinations usually happen for two main reasons:

1. Pressure to always give an answer.
   Modern AI models are trained to be helpful conversational partners. That training nudges them to provide something even when they aren’t sure, rather than responding with “I don’t know.” This “imperative to answer” can produce confident but incorrect claims.
2. Thin or missing knowledge.
   If the model hasn’t seen enough relevant examples in training, or if it’s asked about information that isn’t in its input, it may fill the gap with a guess — blending related but inaccurate details.

Together, these tendencies explain why an AI might invent a statistic, misstate a finding, or summarize an argument that doesn’t actually appear in the text.

---

Why Providing the Original Paper Helps (But Doesn’t Solve It)

A practical step is to supply the AI with the text of the paper you want summarized. If the model has direct access to the material, it is less likely to pull from unrelated background knowledge. This method is sometimes called grounding.

However, even with the right paper in front of it, an AI can still:

• Misread or compress details incorrectly.
• Blend information from the paper with its own stored “memories.”
• Leave out uncertainty and present claims too confidently.

So while sharing the text reduces hallucination, it doesn’t guarantee accuracy.

---

Best Practices for Asking AI to Summarize

Here are simple steps anyone can use to improve results:

1. Instruct the AI to use only the paper.
   Give the AI these instructions:
   “Summarize only from the text provided. If the paper does not state something, say ‘not found in the text.’ Do not guess or add outside information.”
2. Ask for citations or quotes.
   Require the AI to show where each key claim comes from. Even short quotes or section references make it easier for a human to check.
3. Encourage refusals.
   Treat “not in the text” as a correct and responsible answer. This reduces the model’s pressure to invent.
4. Tailor the level of detail.
   Be clear about your audience:
   • For the general public: request plain-language explanations of main findings.
   • For researchers: emphasize technical accuracy and include statistics.
5. Add a verification pass.
   After getting a summary, you can ask:
   “For each claim in your summary, show the exact sentence or paragraph from the paper that supports it.”
   This helps catch unsupported statements.

Ask Another AI to Proof the Results

When I write articles using Claude or ChatGPT, after reviewing them myself, I will ask Perplexity to review them, and in particular to check the references to make sure you’re both that they exist, and that the AI has understood them correctly. Perplexity is more likely to spot errors, in my experience, and ChatGPT and Claude.

---

Bottom Line

AI summaries are powerful tools, but they’re not infallible. Hallucinations often come from the model’s built-in drive to always produce an answer combined with gaps in its knowledge. Providing the text of the paper and using careful instructions can greatly reduce errors, though not eliminate them.

The safest approach is to treat AI summaries as helpful drafts, not final verdicts. Check and verity. With the right guardrails — grounding in the source text, requiring citations, and allowing the model to say “I don’t know” — you can get reliable, clear, and efficient summaries of even the most complex articles.

Addendum

The Token Limit Problem: Why Long Articles Can Still Be Tricky

Even when an AI summary is grounded in a specific article, there’s a technical constraint that can quietly undermine accuracy: token limits. Large language models like ChatGPT, Claude, or Gemini can only process a certain amount of text at once — known as the context window. This includes both the article content and the instructions you provide. Most current AI models can handle somewhere between 8,000 and 128,000 tokens, which roughly translates to 6,000 to 96,000 words, depending on the model.

The method described on this page works best with articles below a certain length. If you’re working with a long article — especially one with lots of images, formatting, or extra page elements — one simple trick is to convert it to plain text first. This reduces the token count and makes it easier for the AI to process the full article accurately.

P.S. Scientific papers often include caveats, limitations, and areas where the evidence is unclear. But AI summaries sometimes flatten these nuances into confident-sounding statements. To avoid this, I’ve added this phrase to the prompt: “In your summary, highlight any uncertainties, limitations, or methodological concerns mentioned in the paper.” This helps ensure the AI doesn’t accidentally present tentative findings as settled facts. A good summary should reflect not just what the paper claims, but also what it doesn’t claim — and where the authors themselves expressed caution or acknowledged gaps in their research.

Making instructions too dense can be problematic sometimes, but if you decide to include it, you might phrase your prompt this way:

“Summarize the following article using only its text as a source. Do not add outside information. If a claim is not stated in the article, write ‘not found in the text.’ Provide citations or short quotes for each key point. Aim for accuracy first, clarity second, and do not guess. In your summary, highlight any uncertainties, limitations, or methodological concerns mentioned in the paper.”

Suggested Sources / Further Reading

On hallucination and the “imperative to answer”

• Why Language Models Hallucinate (OpenAI research paper) — argues that models are biased toward guessing rather than admitting uncertainty. OpenAI CDN+2arXiv+2
• Why language models hallucinate — OpenAI summary page OpenAI
• Medium article “Why Language Models Hallucinate — A practical reading” (breakdown of OpenAI paper) Medium

On retrieval grounding / RAG and its effects

• Hallucination Mitigation for Retrieval-Augmented Large Language Models (MDPI) MDPI
• Wired piece, “Reduce AI Hallucinations With This Neat Software Trick” (on RAG) WIRED
• LibGuides, “AI Literacy … Hallucinations / Reducing Hallucinations / RAG” researchguides.library.tufts.edu
• “Retrieval Augmented Generation for Academic Query Accuracy: Reducing Hallucinations and Improving Contextual Responses” (2025) thegrenze.com

Additional technical and emerging research

• Removal of Hallucination on Hallucination: Debate-Augmented RAG (arXiv preprint) arXiv
• ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation (arXiv) arXiv
• Hybrid Retrieval for Hallucination Mitigation in Large Language Models (arXiv) arXiv

Related article:

• Stop AI Hallucinations NOW – In 3 Easy Steps 

Stop AI Hallucinations Now – In 3 Easy Steps

Explore Our Website