AI, with a little help from their friends

By now, everyone's clued in on artificial intelligence's advantages over its human variety. But recently, Microsoft Copilot leader Vik Singh aired the counterintuitive view that a way be found by which AI chatbots know when they should seek (human) help. GenAI has a problem with fact. Bots trained on LLMs spew out factual errors almost as many times as they get answers right. Researchers still don't know why machines 'hallucinate' beyond a vague notion that neural networks can imagine data they have not been fed. Humans have a better understanding of how to handle imagination. Paul Cezanne crafted 'The Card Players' from figures of farmhands to deliver a composition of intense detail and focus. Asked to render the image of a 19th-century railway bridge, AI is fairly accurate - except it paints in an extra track.

The goal of AI is to mimic human creativity. It needs to learn how humans imagine without getting answers wrong as often as machines. There are several ways to go about this. Additional context helps. So, throw more data at the bots. Improve how bots learn - therefore, tweak their inference architecture. Use a time-tested method and have bots debate each other to arrive at the truth. Or, simply get them to seek help from humans. Improvements to data and learning models are delaying AI's commercial rollout. It may make sense to hand bots over to users who can then train them according to their needs to distinguish fact from fiction. This is admittedly a suboptimal solution. But it works. Humans could train bots to handle simple tasks and build in the complexity over time. Humans learn to eliminate hallucination this way.

Instead of waiting for AI to reach a cognitive threshold, less sophisticated versions could be commercially exploited. AI is years away from 'solving' chess. Yet, the game has transformed to a point that a grandmaster with a well-trained machine is unbeatable. This model of augmented intelligence offers immediate benefits. Human and machine intelligence can co-evolve profitably.