AI models are getting smarter — and sneakier. In this post we explain how to spot hallucinations, which prompt techniques actually work, and share a ready-to-use guardrail prompt to make your AI say “I don’t know” when it should

The problem in one sentence

A 2025 MIT study found that large language models are 34% more likely to use confident-sounding words like “definitely,” “certainly,” and “without doubt” — and they do that especially when they’re hallucinating. Even funnier (and worse): the fake facts often sound more convincing than the real ones.

The bright side (yes, there is one)

Things are improving. According to the AI Hallucination Report 2025, Google’s Gemini-2.0-Flash-001 leads the pack with only 0.7% hallucinations — the lowest reported rate in the industry. So the machines are getting modesty lessons… slowly.

The golden rule while AI learns humility

Trust smart. Verify smarter.

Let the model be clever — but don’t let cleverness substitute for verification.

Practical rules we tested (add to your prompt)

To reduce the chance of being fed a confident lie, we experimented with these prompt-level rules when building the text:

• Estimate your confidence. Ask the model to rate how confident it is in its answer.

• Cite the source. Request a link to the primary source (or state that none is available).

• Give ranges, not fake precision. Ask for a plausible range instead of a precise figure when data is weak.

• Use an explicit guardrail prompt — for example:

You are an AI assistant. Your task is to give accurate, verified answers.
If you are unsure or lack data, explicitly say “I don’t know” instead of guessing.
After generating your response, critically recheck it for factual consistency.

What definitely doesn’t work (we tried — zero effect)

Some phrasings and tactics are surprisingly useless; some make things worse:

• “Be precise” - vague and ignored.

• “Think carefully” - polite, but ineffective.

• Longer prompts - just add noise.

• Repeating instructions does not add to the model's intelligence.

In fact, for some models, these approaches increase confident-sounding hallucinations.

TL;DR (for people who skim headlines)

AIs still hallucinate. Some models are getting better, but don’t treat confident language as a correctness seal. Use explicit verification steps in prompts, demand sources, and prefer ranges to fake precision. When in doubt, get the model to say “I don’t know.”