🧠 Turn AI Hallucinations Into Reliable Evidence

Kia ora, Namaskaram 🙏🏾

Do you trust your favourite AI chatbot?

If you're using ChatGPT, Claude, Copilot—or any other LLMs for research—you're bound to encounter “hallucinations.”

These are answers that may sound right—but are often made up, taken out of context, or just confident-sounding waffle.

📚 Evidence on Hallucinations

To reduce hallucinations, it helps to know which type you're dealing with—so you can prompt effectively and get more reliable evidence.

Type I – Factual Inaccuracies (Sounds true, but wrong facts)

👉🏾 Use RAG i.e. Retrieval-Augmented Generation
Cross-check outputs with trusted sources—like giving your AI assistant access to a fixed but reliable research library.

📈 Effectiveness: Very High
📝 Why: The evidence highlights RAG as one of the most potent strategies. It grounds LLM outputs in external, trustworthy information (e.g., knowledge graphs, databases, APIs). This significantly reduces reliance on memorised or fabricated content.

Type II – Semantic Distortions (Misunderstands the context)

👉🏾 Use Prompt Engineering
Be clear and specific—like you're briefing a junior colleague. The more context you give, the better the AI understands what and how you're asking it to research.

📈 Effectiveness: Medium to High
📝 Why: The evidence notes that well-structured, context-rich prompts (including techniques like in-context learning and instruction tuning) can improve the relevance and precision of outputs. It’s especially helpful for improving how LLMs interpret meaning—but it depends heavily on user skill.

Type III – Fluency Discrepancies (Overconfident output)

👉🏾 Use Chain-of-Thought (CoT) Prompting
Ask AI to think and write down step by step—it makes the logic visible and easier to understand how it arrived at the final output.

📈 Effectiveness: High
📝 Why: Chain-of-Thought prompting encourages step-by-step reasoning, which the evidence shows improves factual consistency and logical coherence. It reduces the risk of “fluent nonsense” by making the AI reveal its internal logic.

Citation: Jaya Nanwani, Rakesh Kadu K. Advances in Reducing AI-Generated Hallucinations: Techniques and Open Challenges. Authorea. March 10, 2025.

Hallucinations have reduced—but they haven’t disappeared.

With the best prompts and latest model updates, they’re likely here to stay.

Because that’s just how the models works.

The real question is …
When should we use LLMs for more reliable research?

Want my full ChatGPT Playbook for Behaviour Change—taught live? Join my upcoming 2-week course.

P.S. Newsletter readers can use the code EVIDENCE to get $50 off the course (Offer expires in 48 hours)