AI makes things up. Yes. Here's how to spot its lies
Here are 5 techniques to detect AI lies:
🔍 1. Always ask for sources
The simplest reflex: "What sources are you basing this on?" If the AI can't cite a verifiable source or invents references, it's a red flag. An AI that says "according to a Harvard study" without giving a year, author, or title is probably inventing. The RAG approach (Retrieval-Augmented Generation) is the solution: the AI draws answers from a verified database rather than internal memory .
⚖️ 2. Use specialised AI, not generalists
Large generalist models (ChatGPT, Gemini) excel at everything but master nothing. For legal questions, use AI trained on Kenyan laws. For tax questions, use AI that has integrated KRA guidelines. Tools like Knowlix AI (29 African countries) are trained on specific data to reduce hallucinations . The more specialised the AI, the less it invents.
👁️ 3. Read between the lines: AI loves unnecessary details
Hallucinations often hide in superfluous details. An AI giving names, dates, overly precise figures without sources is suspicious. A model that lacks nuance and presents everything as absolute truth is also suspect. If the answer is too good to be true, it probably is.
🔄 4. Cross-check with other sources
AI is not a source. It's a tool. Ask the same question to two different tools (ChatGPT and Claude). If their answers diverge, dig deeper. If information isn't confirmed by an official source (KRA, KNBS, CBK), don't take it at face value.
❓ 5. Ask trap questions
When AI gives you advice, ask: "What are the risks of this approach?" Inventing AIs often provide one-sided answers. A reliable AI will alert you to risks. Also ask: "In what cases would this recommendation not work?" Models that hallucinate can't see their own limits.
Tip: The European AI Act, adopted in 2024, imposes transparency requirements on AI systems, including the right to clear explanations and to challenge automated decisions . Even in Kenya, use tools that integrate RAG for answers grounded in reliable data.
Do it now: Take a recent response from ChatGPT or another AI. Verify every piece of numerical data. If you can't find a reliable source, don't use it.
Sources: TechCabal – Knowlix AI Launch (2026); Meta – AI Hallucinations Report (2026); HBR – AI Transparency Framework (2025); AI Act – European Union (2024); RAG – Retrieval-Augmented Generation Research (2025).
👍 If this article helped you, thumbs up — if you have questions, reply to this message.