Keeping AI away from the bad guys
Illustration: Sarah Grillo/Axios
In the U.S. and Europe, Big Tech is under fire — hit with big fines and the threat of stiff regulation — for failing to thwart the profound consequences of its inventions, including distorted elections, divided societies, invaded privacy, and sometimes deadly violence.
Driving the news: Now, artificial intelligence researchers, facing potentially adverse consequences from their own technology, are seeking to avoid being ensnared by the same "techlash."
- AI researchers are working to limit dangerous byproducts of their work, like race- or gender-biased systems and supercharged fake news.
- But the effort has partly backfired into a controversy of its own.
What's going on: As we reported, OpenAI, a prominent research organization, unveiled a computer program last week that can generate prose that sounds human-written.
- It described the feat and allowed reporters to test it out (as we did — see this), but OpenAI said it would withhold the computer code.
- It said it was attempting to establish a new norm around potentially dangerous inventions in which, for the sake of preventing their possible misuse, researchers would continue their work but keep some advances under wraps in the laboratory.
- In the case of its own new invention, OpenAI said it feared that somebody could use it to effectively develop a weapon for mass-producing fake news.
- This was the first time a major research outfit is known to have used the rationale of safety to keep AI work secret.
But the move met massive blowback: AI researchers accused the group of pulling off a media stunt, stirring up fear and hype, and unnecessarily holding back an important research advance.
Why it matters: Against the backdrop of the techlash, we're seeing a messy debate play out around an urgent question: what to do with increasingly powerful "dual-use" technologies — AI that can be used for good or for ill.
- The outcome will determine how technology that could cause widespread harm will — or won't — be released into the world.
- "None of us have any consensus on what we're doing when it comes to responsible disclosure, dual use, or how to interact with the media," Stephen Merity, a prominent AI researcher, tweeted. "This should be concerning for us all, in and out of the field."
Details: OpenAI says its partial disclosure was an experiment. In a conversation with two top AI researchers from Facebook, OpenAI's Dario Amodei held up social media companies as a cautionary tale:
"The people designing Twitter, Facebook, and other seemingly innocuous platforms didn't consider that they might be changing the nature of discourse and information in a democracy … and now we're paying the price for that with changes to the world order."
- Several researchers praised OpenAI's decision to withhold code as a vital step toward rethinking norms. "I think it's amazingly responsible," said Kristian Hammond, a Northwestern professor and CEO of AI company Narrative Science.
But other academic researchers came down hard.
- While the new program is often impressive, its researchers admit that they simply used a scaled-up version of previous work. It's very likely therefore that someone could replicate the feat at relatively minimal cost. OpenAI says that's why it sounded the alarm.
- But Sam Bowman, a professor at New York University, said the move "feels like a worst-of-both-worlds compromise that slows down the research community without actually having a real long-term safety impact."
- Several experts said OpenAI's warnings of potential societal impacts are exaggerated. "We're still very far away from the risks," says Anima Anandkumar, a Caltech professor and Nvidia's machine learning research director. She said it's early to be withholding any research at all.
What's next: Computer science is lurching toward the same robust discussion that biologists and nuclear scientists had before them — when to circumscribe openness in the name of safety and ethics.
- Notably, Google recently said it will consider potential harms of its AI research before deciding to publish it.
- "I'm not sure what alternative there was," says Jeremy Howard, a founder of AI company Fast.ai. "I think OpenAI did the right thing here, even if they communicated it sub-optimally."