"What makes the goal of accuracy so vexing for chatbots is that they operate probabilistically when choosing the next word in a sentence; they aren’t trying to find the light of truth in a murky world. “These models are built to generate text that sounds like what a person would say — that’s the key thing,” Jesse Dodge says. “So they’re definitely not built to be truthful.” I asked Margaret Mitchell, a computer scientist who studied the ethics of A.I. at Google, whether factuality should have been a more fundamental priority for A.I. Mitchell, who has said she was fired from the company for criticizing how it treated colleagues working on bias in A.I. (Google says she was fired for violating the company’s security policies), said that most would find that logical. “This common-sense thing — ‘Shouldn’t we work on making it factual if we’re putting it forward for fact-based applications?’ — well, I think for most people who are not in tech, it’s like, ‘Why is this even a question?’” But, Mitchell said, the priorities at the big companies, now in frenzied competition with one another, are concerned with introducing A.I. products rather than reliability. The road ahead will almost certainly lead to improvements. Mitchell, who now works as the chief ethics scientist at the A.I. company Hugging Face, told me that she foresees A.I. companies’ making gains in accuracy and reducing biased answers by using better data. “The state of the art until now has just been a laissez-faire data approach,” she said. “You just throw everything in, and you’re operating with a mind-set where the more data you have, the more accurate your system will be, as opposed to the higher quality of data you have, the more accurate your system will be.” Jesse Dodge, for his part, points to an idea known as “retrieval,” whereby a chatbot will essentially consult a high-quality source on the web to fact-check an answer in real time. It would even cite precise links, as some A.I.-powered search engines now do. “Without that retrieval element,” Dodge says, “I don’t think there’s a way to solve the hallucination problem.” Otherwise, he says, he doubts that a chatbot answer can gain factual parity with Wikipedia or the Encyclopaedia Britannica."