The Fine Print When Wrong Sounds Right

Dead and Gone… The Fine Print When Wrong Sounds Right By Gary Payne, MBA Founder of FuneralCostOntario.ca The Fine Print When Wrong Sounds Right For a while, one of the easiest ways to make artificial intelligence look ridiculous was to ask it about a strawberry. Specifically, how many times does the letter R appear in the word strawberry? The answer is three, something most of us can establish without advanced mathematics or a graduate degree. ChatGPT, however, became famous for sometimes answering two and then, in the especially entertaining versions shared online, explaining why two was correct. This went on long enough that on April 28 of this year, the official ChatGPT account on X posted "at long last" alongside a demonstration showing that ChatGPT could finally get the strawberry question right. The internet responded exactly as you would expect. Somebody changed the fruit. Users tried cranberry, and screenshots appeared showing ChatGPT confidently getting that wrong instead. A TechRadar writer tried it independently and also received the wrong answer. I find this funny. There is something wonderfully humbling about technology capable of writing computer code being defeated by the produce section. The problem is that most of the questions we are beginning to ask these machines aren't about berries. Earlier this year, two Tennessee lawyers learned that distinction rather expensively. In Whiting v. City of Athens, a U.S. appeals court found more than two dozen fake citations and factual misrepresentations in briefs submitted to the court. Some cases didn't exist. Others existed but didn't say what the lawyers claimed they said. Each lawyer was ordered to pay a $15,000 punitive sanction, along with legal fees and double costs. The court did not determine that artificial intelligence had created the bad citations. In fact, its larger point was more useful than that. It didn't really matter whether the information came from AI, another lawyer or somewhere else. The lawyers were responsible for checking it before putting it in front of the court. That is a long way from miscounting the R's in cranberry. The same problem becomes easier to understand when it leaves the courtroom and lands somewhere much more familiar: trying to get a straight answer from an airline. In November 2022, Jake Moffatt was travelling after the death of their grandmother and went to Air Canada's website looking for information about bereavement fares. The airline's chatbot said a ticket could be purchased first and the reduced bereavement rate requested afterward, as long as the application was made within 90 days of the ticket being issued. Moffatt relied on that information, travelled and later asked Air Canada for the refund. Air Canada said no. Its actual policy said the opposite. Bereavement fares could not be applied retroactively after the trip. Moffatt eventually took the dispute to British Columbia's Civil Resolution Tribunal, where Air Canada argued, in effect, that it should not be responsible for inaccurate information provided by its chatbot. Tribunal member Christopher Rivers called that a "remarkable submission." The chatbot was part of Air Canada's website. The company was responsible for the information it provided there. Moffatt was awarded $650.88 in damages, plus interest and tribunal fees. That case gets closer to what bothers me about all of this. The machine doesn't have to sound confused when it is confused. There is no hesitation, no scratching of the head and no "I might be wrong about this." A completely invented answer can arrive in the same calm, polished language as one that is absolutely correct. Calling that a lie is technically unfair. A lie requires somebody to know the truth and deliberately say something else. ChatGPT isn't sitting somewhere thinking, "I'm going to have some fun with Gary today." But from the other side of the screen, the distinction can become surprisingly academic. If I believe the answer and act on it, the damage doesn't particularly care whether the machine intended to deceive me. None of this has made me stop using AI. Quite the opposite. I use it because it can be extraordinarily useful, and I suspect most of us are going to use much more of it in the years ahead. What it has changed is the weight I give the answer itself. For most of my life, confidence has been one of the little signals we use to decide whether somebody knows what they're talking about. We listen for hesitation, notice uncertainty and become suspicious when somebody starts qualifying every second sentence. AI changes that equation because the confidence can be manufactured along with the answer. Sometimes it will be right and sometimes it will be wrong, but I no longer confuse how confidently something is written with how likely it is to be true.