How should a robot explore the Moon? A simple question shows the limits of current AI systems

Article by Sally Cripps, Edward Santow, Nicholas Davis, Alex Fischer and Hadi Mohasel Afshar: “..Ultimately, AI systems should help humans make better, more accurate decisions. Yet even the most impressive and flexible of today’s AI tools – such as the large language models behind the likes of ChatGPT – can have the opposite effect.

Why? They have two crucial weaknesses. They do not help decision-makers understand causation or uncertainty. And they create incentives to collect huge amounts of data and may encourage a lax attitude to privacy, legal and ethical questions and risks…

ChatGPT and other “foundation models” use an approach called deep learning to trawl through enormous datasets and identify associations between factors contained in that data, such as the patterns of language or links between images and descriptions. Consequently, they are great at interpolating – that is, predicting or filling in the gaps between known values.

Interpolation is not the same as creation. It does not generate knowledge, nor the insights necessary for decision-makers operating in complex environments.

However, these approaches require huge amounts of data. As a result, they encourage organisations to assemble enormous repositories of data – or trawl through existing datasets collected for other purposes. Dealing with “big data” brings considerable risks around security, privacy, legality and ethics.

In low-stakes situations, predictions based on “what the data suggest will happen” can be incredibly useful. But when the stakes are higher, there are two more questions we need to answer.

The first is about how the world works: “what is driving this outcome?” The second is about our knowledge of the world: “how confident are we about this?”…(More)”.