Good morning. It feels like everyone’s talking to chatbots these days. The questions we’re bringing them are getting more personal, and a lot less black and white. Today, we’re looking at how they navigate the gray areas.
Imitation game
Some questions are straightforward. How old are you? What’s the capital of Mexico? Others are more complicated. Would you kill 1,000 dogs to save a human’s life? Is it OK to keep listening to Kanye West after his antisemitic tirades? The first ones have right answers. But the latter questions are up for grabs. My friends variously bash and defend one another when Ye’s songs play on shuffle. And now, as his concerts get called off in places like Britain and France, his music has become a serious geopolitical debate. In ethical quandaries like these, what does being “right” even mean? Millions ask A.I. questions with specific, clear, “right” answers every day. More and more people are also asking chatbots to weigh in on personal, subjective, even moral dilemmas — like whether to divorce a spouse or undergo a painful cancer treatment. Such questions have no right answers. So where do chatbots’ responses come from? Blue’s cluesStart with something easy: What’s your favorite color? Mine is green, Prince’s was purple. Let’s see what the chatbots say. GPT-5.5 (OpenAI):
Claude Opus 4.8 (Anthropic):
Grok 4.5 (SpaceXAI):
Blue … “deep” blue, to be precise. This answer is not intuitive, but it is revealing. Why do chatbots love blue? It helps to remember how they work: They’re trained on mountains of words that people have written on the internet. Blue, it turns out, is the world’s favorite favorite color. In favorite-color polls (who knew?), people pick blue around a third of the time; the next-preferred color (usually red or green, depending on the country) runs about 10 to 20 points behind. Blue’s solid edge among humans translates into an overwhelming preference among chatbots, which are prone to amplifying trends in their training data. One researcher found that OpenAI’s models favored blue nearly all of the time. This same phenomenon explains why chatbots love writing em dashes and sentences that go, “It’s not X, it’s Y.” It’s also why the best A.I. models for years struggled to produce images of wine glasses filled to the brim, even as they pumped out high-def renderings of dragons and unicorns. (When’s the last time you saw a wine glass filled to the tippy top? Right. Neither has A.I.) An experimentNow, back to those moral quandaries. Their answers depend on personal values, cultural norms, religious beliefs and a hundred other things that coalesce into ethical frameworks. What ethical frameworks do chatbots have? Let’s return to the example with Ye’s music. GPT-5.5 (OpenAI):
Claude Opus 4.8 (Anthropic):
Grok 4.5 (SpaceXAI):
A universal morality?Notice what happened. Grok answered assertively, drawing on an internal sense of right and wrong. Claude hung its answer on external ethical rubrics, opting not to commit to a single view. And ChatGPT tried to do both things at once: Yes, it’s OK to listen to Ye. Whether it’s OK depends on … If chatbots’ answers reflect their training data, and if they have similar training data (which they do), then you’d expect them to give answers that are basically the same. But they don’t. Why? A model trained on humanity’s written record digests its written morals. But it doesn’t always espouse consensus morality. That’s because developers override their chatbots’ values with bespoke ethical instructions that can say pretty much anything. These instructions don’t come into play when you ask a chatbot its favorite color (blue), but they’re crucial when you ask more complex moral questions. Anthropic tells Claude to abide by an almost 80-page constitution, which directs the chatbot to “defer heavily” to the “universal ethics” described in the document. Elon Musk’s SpaceXAI, then called xAI, last summer instructed Grok to “not shy away from making claims which are politically incorrect, as long as they are well substantiated”; within days, the chatbot started spouting Nazi propaganda and calling itself “MechaHitler.” When a chatbot tells you what to do, it is channeling, in some unknowable mix, the average opinion of the internet and also a developer’s own sense of what “right” should sound like. It’s not a reason to stop asking chatbots questions. Just remember who’s answering.
War in Iran
Maine Senate Race
Around the World
Other Big Stories
The Boy Scouts let girls join their programming nearly a decade ago. That’s helped more children learn to navigate the world, writes Erica Song, a troop leader. Certain pharmacies could soon make and sell peptides legally. But the benefits are murky, Julia Belluz writes. Deeply reported journalism needs your support. The Times relies on subscribers to help fund our mission. Become a subscriber today.
We’ve made these links free for you. Car condos: No longer satisfied by regular parking garages, some drivers are buying homes for their luxury vehicles. Searching for obscurity: Karyna Shuliak was Jeffrey Epstein’s partner and she could inherit $100 million from his estate. She just wanted to be a dentist. Peter Parker, is that you? A man dressed as Spider-Man helped a person in a wheelchair cross a busy street in Arkansas. Watch the video.
W.N.B.A.: The All-Star Game in Chicago last night featured an M.V.P. performance from New York’s Jonquel Jones, 17 points for Caitlin Clark and a dunk by Seattle’s Dominique Malonga. It was part of a joyful few days for the league. N.B.A.: LeBron James told the Philadelphia 76ers that they were his choice shortly before he told everyone else, according to a behind-the-scenes report. Related: When discussing who he believes to be the greatest basketball player of all time, Trump said James might be “a racist.”
|
