WhatsApp and artificial intelligence
Hallucination
A hallucination, in artificial intelligence, is an answer a model gives with confidence but that rests on no data at all: a timetable that does not exist, a made-up rule, a service the residence does not offer. It sounds just as convincing as a correct answer.
Also known asAI hallucination · made-up answer · confabulation · alucinación
Why it happens and how it is contained
A language model generates text word by word according to what seems likely, and each word builds on the ones before, so an early error can drag the rest along with it. If someone asks what time the gym closes and that fact is not in what it has been given, the model may fill the gap with a time that sounds reasonable. The phenomenon is not fully understood and research into reducing it is ongoing; what is known is how to contain it:
- Give it the source. Answering from the residence's knowledge base reduces hallucinations but does not eliminate them: the model can also go wrong when reading that source.
- Let it not know. "I don't know, I'll pass it to the team" should count as a correct answer, not as a failure.
- Take decisions away from the model. An exception to the house rules or a promise on someone's behalf goes to a person: that is the handoff to a human.
Worth knowing
What people take for granted: that hallucination is a rare glitch that goes away with a better model. The technical sources themselves treat it as an open problem. The useful question for a residence is not whether the assistant can make things up, but what it does when it does not know something: whether it says so and passes it on, or fills the gap.
Sources
Related terms
- Language modelA language model is an artificial intelligence model trained on vast amounts of text that generates language by predicting the next word; a very large one is called an LLM. It is what lets an assistant understand a message written in someone's own words and draft the reply.
- Knowledge baseA knowledge base is the organised body of information an AI assistant draws its answers from: in a residence, the house rules, the timetables, the services and the frequently asked questions. When the model looks it up before answering, the technique is called retrieval-augmented generation (RAG).