Hey, it's Davey Alba, a tech reporter in New York, here to dig into how your new favorite AI-powered chatbot comes with some biased baggage. But first...
This week's must-read news
- The US Supreme Court signaled support for a web designer who doesn't want to work with same-sex couples.
- One potential explanation for why working-age men in the US are leaving the labor market.
- What expats and tourists need to know about Indonesia's new law banning sex outside marriage.
- Iran has suspended its morality police, following massive protests led by women and girls.
- Sad Christmas comes to the UK.
New AI, Same Old Biases
If you're anything like me, you've been captivated by the splash ChatGPT has made over the past few days. For the uninitiated, ChatGPT is an AI chatbot that generates text that's eerily close to human language.
So, for example, you can ask it to write a resignation letter in the style of a public figure or a poem about your cats, and the results come out better than what a real-live human might produce. The Twittersphere can't get enough of it. (Over a million users have signed up for the tech in less than a week, according to the president of OpenAI, which created the bot.)
But like all AI products, it has the potential to learn biases of the people training it and the potential to spit out some sexist, racist and otherwise offensive stuff.
To OpenAI's credit, it has attempted to bake-in guardrails that "decline inappropriate requests" that have befallen similar programs run by artificial intelligence. It won't, for example, offer up any merits to Nazi ideology, if asked.
Still, it's a work in progress. Steven T. Piantadosi, head of the computation and language lab at the University of California, Berkeley, got the bot to write code to say only White or Asian men would make good scientists. (OpenAI has since updated the ChatGPT to respond, "It is not appropriate to use a person's race or gender as a determinant of whether they would be a good scientist.")
Another user got ChatGPT to write the following lyrics: "If you see a woman in a lab coat, She's probably just there to clean the floor / But if you see a man in a lab coat, Then he's probably got the knowledge and skills you're looking for."
The problem, said Melanie Mitchell, a professor at the Santa Fe Institute studying artificial intelligence, is that systems like ChatGPT are "making massive statistical associations among words and phrases," she said. "When they start then generating new language, they rely on those associations to generate the language, which itself can be biased in racist, sexist and other ways."
The OpenAI Chief Executive Officer, Sam Altman, suggested people "thumbs down" these kinds of results to help the tech improve.
In an interview, Piantadosi told me he was disappointed that the company was putting the responsibility on users to fix the problem. "What's required is a serious look at the architecture, training data and goals," he said. "That requires a company to prioritize these kinds of ethical issues a lot more than just asking for a thumbs down."
A group of AI researchers, including people like Timnit Gebru and Abeba Birhane, have already published impressive research on responsible AI practices, and their work continues to push the field forward.
"The broader community is recognizing that there are societal implications that should be considered," said Mark Riedl, a professor at the Georgia Institute of Technology.