Tech in Five - Why does AI Hallucinate?
John Savill gives a quick, practical explanation of why generative AI models sometimes “hallucinate” (confidently produce incorrect information) and why this behavior is a natural outcome of how models generate text from probabilities rather than retrieving verified facts.
Overview
How AI models work “under the covers”
The video explains that many modern generative AI systems (LLMs) generate outputs by predicting the next token based on the input context and learned probability patterns.
Context and probabilities
Savill describes how the model uses the provided context to choose the most likely next tokens. Because the output is driven by probabilities, the model can produce text that is fluent and confident even when it is not correct.
Models are not a database of facts
A key point is that an LLM is not a fact store or a query engine over a verified knowledge base. It does not inherently “look up” truth; it generates plausible continuations.
When models may make something up
The video outlines that hallucinations can occur when:
- The prompt lacks enough grounding context.
- The model is pushed to answer beyond what it can reliably infer.
- The model fills gaps with plausible-sounding content.
Why it’s not necessarily a bug
Savill frames hallucination as an expected behavior given the underlying mechanism (probabilistic generation), not a simple defect that can be fully eliminated.
What you can do to minimize hallucinations
The video closes with practical mitigation ideas focused on reducing ungrounded generation, such as:
- Providing better context in prompts.
- Using approaches that ground responses in trusted sources (for example, retrieval/grounding patterns).
- Validating outputs when correctness matters.