Artificial intelligence? Maybe we need artificial stupidity instead. In order to avoid an apocalyptic scenario where machines take over the world, some are suggesting we should limit AI to human-level intelligence.
At least then we’ll stand a fighting chance, says at the University of Louisville, USA.
Although artificial general intelligences – AIs that can perform any task that a human brain can – do not yet exist, Yampolskiy thinks planning how to limit their computational abilities and make them behave in ways that are familiar to humans may help avert disaster in the future.
Advertisement
Yampolskiy and his colleague Michaël Trazzi at Sorbonne University in Paris suggest constraining AI by introducing “artificial stupidity”. Using rough estimates of the storage space, bandwidth and processing power of the human brain, they propose, capping the number of operations per second or restricting the capacity of an AI’s working memory to comparable levels for the human brain.
However, putting such a cap on a machine brain may not work as intended. Humans might not use their processing power optimally, meaning an AI could outmatch us even though it could carry out far fewer operations per second.
Humanlike AI
So two further measures are suggested: ensure the AI is not able to self-improve to make itself more efficient and give it humanlike cognitive biases – mistakes in reasoning, often made because of deeply held beliefs, that mean we deviate from rational judgement. A machine mind with such foibles might in theory act more like us, allowing us to anticipate its moves better.
Conservatism is one such bias that could be introduced: “the [AI] will keep the same initial values”. Courtesy bias is another: “the [AI] will try to not offend anyone, avoiding aggressive behaviours”. Humans attempting to constrain such an AI could use their political or social skills to ensure a happy symbiosis with it.
Some of the suggestions make a lot of sense, says at the University of Oxford’s Future of Humanity Institute. “Things like processing limitations could be used to make the AI’s thinking interpretable – not enough to cripple it, but to make sure it can’t get away with things we aren’t able to follow,” he says.
However, Armstrong questions whether the overall approach that Yampolskiy suggests will be effective in the end. Take the attempt to prevent a highly intelligent AI from improving itself. What is to stop a machine from making a copy of itself that doesn’t have such a constraint? Then the jig would be up.
And while cognitive biases may be helpful, perhaps we don’t understand enough about how they exist in human minds yet to know how they could be transported reliably to machines.
Journal reference:
