AI hallucination refers to when an artificial intelligence system generates false or incorrect information, presenting it as if it is true.
For example, an AI may falsely identify objects in images or video that are not actually present, hear words that were not spoken in audio, or make incorrect predictions that have no factual basis, going so far as to provide “evidence”, though its resources are not real.
Why do we get AI hallucinations?
As with all things AI, it comes down to the data. If there’s an issue with the data being used to train the AI, then its outputs are going to be problematic. Datasets that contain errors, anomalies, biases, or are insufficiently diverse, may teach the AI to recognise things that are not actually there. Garbage in, garbage out.
AI models must also be carefully tuned to avoid things like overfitting, where the model is memorising data rather than learning patterns because its sample size is too small, or the training contains large amounts of irrelevant, “noisy data”.
How do we prevent AI hallucinations?
Let’s go back to data. By carefully training models on relevant data with specific sources, we can limit the possibility of it hallucinating. We can create templates for it to follow, with clear guidance to help it recognise the right patterns and data it needs. The great thing about AI is that you can tell it what you want and what you don’t, so that it can learn what it should be looking for.
