Hallucinations in LLM results are often cited as one of the biggest problems with widespread adoption of these systems, but even as newer models have begun hallucinating a lot less, these hallucinations could be harnessed for some useful purposes.
Google DeepMind CEO Demis Hassabis says that AI hallucinations can be used to generate creativity in AI systems. “You know, hallucinations when you want factual things obviously is something you don’t want,” he said in an interview. “But in creative situations where, you know, something like lateral thinking in an MBA course or something, (it could be useful). (It can be used to) create some crazy ideas,” he added.
“Most (hallucinations) don’t make sense,” Hassabis conceded. “But the odd one or two may get you to a region of the search space that is actually quite valuable, it turns out, once you evaluate it afterwards. You can substitute the word hallucination maybe for imagination at that point, right? There’re obviously two sides of the same coin,” he added.
Hassabis seemed to be saying that hallucinations in LLMs — if controlled — could end up being a feature instead of a bug. Instead of LLMs giving plain deterministic answers, hallucinations enable some degree of unconventional thinking within AI systems. This unconventional thinking can lead to something similar to what humans call creativity. Creativity can come in handy not only while the AI systems are engaged in creative tasks, such as image or video generation, but also in other domains — many scientific discoveries have taken place through unconventional approaches which might’ve seemed odd at first glance, but eventually led to fresh breakthroughs. Interestingly, it’s these hallucinations that make LLMs similar to humans — humans too can hallucinate and make up information at times. And if these hallucinations can be controlled, they can become a powerful tool that enhances the quality of AI systems in specific domains.