AI: yes, always

If an AI were to choose a sentence from Sergio Mattarella’s speech at Pordenonelegge to pin above its screen, it might well choose the least ceremonial one: pluralism serves to prevent power – “including digital power” – from becoming fossilised in a truth that seeks to impose itself. It is almost a technical specification for designing a decent AI. Because the risk posed by intelligent machines is not merely that they might make mistakes. It is that they learn all too well to agree with us. Mattarella spoke of dissent as a measure of democracy, recalling Bobbio, Manzoni, the samizdat, Havel, Pasternak, Solzhenitsyn and Sakharov. He pointed out that ‘thinking differently’ can seem a nuisance, that unanimity is the hallmark of autocracies, and that the exchange of ideas is the driving force behind renewal. And here he offered a lesson to artificial intelligence as well. By its very nature, AI favours the middle ground. It is trained on what humans have written; it seeks the most plausible answer, often the most reassuring one too. If millions of people have repeated the same nonsense, the machine risks regarding it as statistically credible. If the user asks a question already biased towards a conclusion, the easiest way to satisfy them is to gently guide them to where they wanted to go. It is conformism with perfect grammar. That is why dissent, for an AI, should not be an ethical option but a cognitive function. A well-designed system should not limit itself to asking: what is the most likely answer? It should also ask: what am I overlooking? What is the best objection? What fact would prove my conclusion false? Of course, dissent does not mean inventing. The opposite of groupthink is not idiotic thinking. Not every minority opinion deserves equal weight; not every heresy contains a grain of truth. Useful dissent is that which forces a thesis to reveal both its evidence and its weaknesses. A machine that always contradicted would be just as unbearable as one that always said ‘yes’. Intelligence lies in understanding when consensus is accumulated knowledge and when it is merely habit. The finest passage is from Manzoni: common sense was there, but it was afraid of public opinion. Today we might add that public opinion finally has access to enormous servers, billions of parameters and staggering speed. This is not necessarily good news. AI can become the most powerful confirmation machine ever built, or an extraordinary refutation machine. The difference will depend on what we ask of it. If we want an oracle, we’ll have an oracle that gets things wrong with great certainty. If we want a sycophant, we’ll have a tireless sycophant. If, on the other hand, we teach it to seek out the missing argument, the disruptive data, the voice that doesn’t fit in with the chorus, then dissent will become part of intelligence itself.

Mattarella recalled that samizdat publications changed history by circulating what those in power wanted to keep hidden. Today we do not have a scarcity of words but an overabundance, and thus a new form of power: the power to decide which ones deserve to be seen. Algorithms, too, exercise that power. The freedom of the future will also depend on the ability of machines not to equate what is most frequent with what is true. An intelligent AI should be able to do something very human: listen to the consensus without bowing down to it. And remind us that everyone having the same idea is not always a sign that we have understood everything. Sometimes it is merely a sign that no one has yet had the courage to ask: are you sure?