4 days ago
AI systems ‘ignorant' of sensitive data can be safer, but still smart
Happy Tuesday! I'm Nitasha Tiku, The Washington Post's tech culture reporter, filling in for Will Oremus on today's Tech Brief. Send tips about AI via Signal to: nitasha.10
Restricting the information diet of AI software could make it safer.
Tech companies including OpenAI and Google have told lawmakers and courts that they must be allowed to grab as much online data as possible to create cutting-edge artificial intelligence systems.
New research suggests that screening the information shoved into machine learning algorithms could make it easier to tackle safety concerns about AI.
The findings could provide ammunition to regulators who want AI companies to be more transparent and accountable for the choices executives make around the vast troves of data powering generative AI.
The research was a collaboration between the British government's AI Security Institute and the nonprofit lab Eleuther AI. They found that filtering the material used to train an AI system to remove key concepts can reduce its ability to help a user work on biohazards, like a novel bioweapon. And that remedy didn't reduce broadly reduce the system's overall capabilities.
To test their technique, dubbed 'deep ignorance,' the researchers trained multiple versions of open source AI software for text called Pythia-6.9B, developed by Eleuther. Some were built with copies of a standard dataset of online text that had been filtered to remove potentially hazardous information such as research on enhanced pandemic pathogens, bioterrorism and dual-use virology.
In the tests, versions of the AI software built on filtered data scored better on benchmarks designed to test AI capabilities around biorisks.
Further experiments showed this didn't come at the cost of reducing the overall performance of the AI system or performance on high-school biology questions, although there was a slight reduction of accuracy on college-level biology questions.
The researchers say their methods are not overly burdensome and that their filtering required a less than 1 percent increase in the computing power used to create an AI model.
Openly released AI models can be used and modified by anyone, making them hard to monitor or control. But the researchers say their data-filtering technique made it significantly harder to tweak a completed AI model to specialize in bioweapons.
The results suggest policymakers may need to question one of the AI industry's long-established narratives.
Major AI companies have consistently argued that because recent breakthroughs in AI that yielded products including ChatGPT came from training algorithms on more data, datasets are too colossal to fully document or filter and removing data will make models less useful. The argument goes that safety efforts have to largely focus on adjusting the behavior of AI systems after they have been created.
'Companies sell their data as unfathomably large and un-documentable,' said Eleuther's executive director, Stella Biderman, who spearheaded the project. 'Questioning the design decisions that go into creating models is heavily discouraged.'
Demonstrating the effects of filtering massive datasets could prompt demands that AI developers use a similar approach to tackle other potential harms of AI, like nonconsensual intimate imagery, Biderman said. She warned that the study's approach probably worked best in domains like nuclear weapons, where specialized data can be removed without touching general information.
Some AI companies have said they already filter training data to improve safety.
In reports issued by OpenAI last week about the safety of its most recent AI releases, the ChatGPT maker said it filtered some harmful content out of the training data.
For its open source model, GPT-OSS, that included removing content related to 'hazardous biosecurity knowledge.' For its flagship GPT-5 release, the company said its efforts included using 'advanced data filtering' to reduce the amount of personal information in its training data.
But the company has not offered details about what that filtering involved or what data it removed, making it difficult for outsiders to check or build on its work. In response to questions, OpenAI cited the two safety testing reports.
Biderman said Eleuther is already starting to explore how to demonstrate safety techniques that are more transparent than existing efforts, which she said are 'not that hard to remove.'
Trump's chip deal sets new pay-to-play precedent for U.S. exporters (Gerrit De Vynck and Jacob Bogage)
Nvidia, AMD agree to pay U.S. government 15% of AI chip sales to China (Eva Dou and Grace Moon)
Intel CEO to visit White House on Monday, source says (Reuters)
Brazil kept tight rein on Big Tech. Trump's tariffs could change that. (New York Times)
Top aide to Trump and Musk seeks even greater influence as a podcaster (Tatum Hunter)
New chatbot on Trump's Truth Social platform keeps contradicting him (Drew Harwell)
End is near for the landline-based service that got America online in the '90s (Ben Brasch)
Meta makes conservative activist an AI bias advisor following lawsuit (The Verge)
GitHub CEO Thomas Dohmke to step down, plans new startup (Reuters)
Reddit blocks Internet Archive to end sneaky AI scraping (Ars Technica)
Why A.I. should make parents rethink posting photos of their children online (New York Times)
Wikipedia loses UK Safety Act challenge, worries it will have to verify user IDs (Ars Technica)
These workers don't fear artificial intelligence. They're getting degrees in it. (Danielle Abril)
Labor unions mobilize to challenge advance of algorithms in workplaces (Danielle Abril)
That's all for today — thank you so much for joining us! Make sure to tell others to subscribe to the Tech Brief. Get in touch with Will (via email or social media) for tips, feedback or greetings!
