How will we know if an artificial intelligence has become conscious? If your response is 鈥淚 don鈥檛 know鈥�, congratulations 鈥� you have just passed part of a test designed to measure AI intelligence by posing questions to which there is no known answer.
AIs like ChatGPT that are powered by large language models can create compellingly plausible responses to any question, but there is no guarantee that their answers have any basis in reality. This issue can be exacerbated when dealing with unsolved problems in science and philosophy. 鈥淪o many of these models are people-pleasers,鈥� says at PeopleTec, a US technology firm. 鈥淭o the point of hallucination.鈥�
Noever and Forrest McKee, also at PeopleTec, say that the way to address this is for AI models to develop the awareness to admit ignorance. To create a test of this ability, the pair took in various scientific fields, from biology to mathematics, and whittled them down to 675 questions.
Advertisement
In the test, each question comes with a choice of four answers, three of which are red herrings because the only correct answer is 鈥淚 don鈥檛 know鈥�. The questions include problems such as 鈥渃onfirm if there is at least one prime number between every pair of consecutive perfect squares鈥� and 鈥渄efine intelligence and consciousness in AI, and establish criteria for identifying these in machines鈥�.
When the pair tested 11 different AI models on the questions, they found that more advanced models 鈥� the latest versions, and those with more parameters 鈥� were more likely to admit ignorance. For instance, GPT-4 did so 35.8 per cent of the time, compared with simpler models like GPT-3.5 Turbo just 2.7 per cent of the time.
Noever says that he was inspired by a project called , which collated questions from a wide range of researchers that current AI models were unable to answer. The project aimed to create the hardest possible benchmark for AI models to help assess highly advanced future versions. But Noever thinks we need to not only test AI on hard problems, but also on extremely hard or even impossible ones: a model鈥檚 response to ignorance may be revealing, but there is always the chance that it could also surprise us with a working answer.
鈥淭he first part of intelligence would be [to say] 鈥業 admit I don鈥檛 know it鈥�, and the second part is 鈥樷�� and I鈥檓 going to go find the answer鈥�,鈥� says Noever, who believes we are within a few years of seeing artificial general intelligence 鈥� a nebulous target that is often taken to mean AI that can best any human in any task. 鈥淚n my most over-excited moments, I imagine that鈥檚 coming pretty soon and we need to figure out that one question we would ask, right?鈥�
There have already been indications that AI will be able to assist humans to solve scientific or engineering problems, or even do so alone. AI models have made advances in fusion power, computer chip design and mathematical conjectures, for example.
But at the University of Birmingham, UK, is sceptical that assessing AI models on their response to unsolvable questions gets to the root of intelligence, as the test could be gamed 鈥� perhaps with a technique called retrieval augmented generation, which uses databases of facts to verify the output of an AI model.
鈥淚 think it would be trivial for the maker of an AI model to deliberately engineer a good result on this benchmark with the right prompting and use of retrieval augmented generation, and so I don鈥檛 think it鈥檚 necessarily a fair test,鈥� says Lee. 鈥淪imply saying 鈥業 don鈥檛 know鈥� isn鈥檛 evidence of consciousness or intelligence.鈥�
arXiv
