This is 2023's last and probably longest episode. To lead off, Megan Stifel takes us through a batch of stories about ways that AI, and especially AI "alignment" efforts, manage to look remarkably fallible. Anthropic has released a paper showing that race, gender, and age discrimination by AI models is real but could be dramatically reduced simply by instructing the model to "really, really, really" avoid such discrimination. (The Techcrunch headline writers had fun snarking on the idea that "racist" AI could be cured by asking nicely, but in fact the discrimination identified by Anthropic was severe bias against older white men, and so was the residual bias that asking nicely didn't eliminate.) The bottom line from Anthropic seems to be, "Our technology is a really cool toy, but it can't be used for for anything that matters.") In keeping with that theme, Google's highly touted OpenAI competitor Gemini was released to mixed reviews; the model couldn't correctly identify recent Oscar winners or a French word with six letters (it offered "amour"). There was good news for people who hate AI's ham-handed political correctness; it turns out you can ask another AI model how to jailbreak your model, a request that can make the task go 25 times faster.
This could be the week that determines the fate of FISA section 702, David Kris reported. When we recorded, it looked as though two bills would go to the House floor, and only one would survive. (Since then, that plan has been dropped, at least for now, in favor of a short-term extension of 702 into April.) The two bills reflect a split between the two committees overseeing the program. Judiciary's bill is a grudging renewal of 702 for a mere three years, full of procedures designed to cripple the program. The intelligence committee's bill also harries the FBI for its past failures but preserves the core of 702.
Gus Hurwitz looks at the FTC's last-ditch appeal to stop the Microsoft-Activision merger. The best case for the Commission, he suspects, is that the appeal will be rejected without actually repudiating the pet theories of the FTC's hipster antitrust lawyers.
Megan and I examine the latest HHS proposal to impose new cybersecurity requirements on hospitals. David, meanwhile, looks for possible motivations behind the FBI's procedures for companies who want help in delaying SEC cyber incident disclosures. Then Megan and I consider the tough new UK rules for establishing the age of online porn consumers. I argue that, if successful, they'll undermine Pornhub's litigation campaign against American states trying to regulate children's access to porn sites.
The race to 5G is over, Gus notes, and it looks like even the winners lost. Faced with the threat of Chinese 5G domination and an industry sure that 5G was the key to the future, many companies and countries devoted massive investments to the technology, but it's now widely deployed and no one sees much benefit. There is more than one lesson here for industrial policy and the unpredictable way technologies disseminate.
23andme gets some time in the barrel, with Megan and I both dissing its "lawyerly" response to a history of data breaches – namely changing its terms of service to make it harder for customers to sue over data breaches.
Gus reminds us that the Biden FCC, which only gained a working majority in the last month or two, is apparently determined to catch up with the FTC in advancing foolish and doomed regulatory initiatives. This week's example, remarkably, isn't net neutrality. It's worse. The Commission is building a sweeping regulatory structure on an obscure and laconic section of the 2021 infrastructure act that calls for the FCC to "facilitate equal access to broadband internet access service…": If you think we're hyperventilating, read Commissioner Brendan Carr's eloquent takedown of the whole initiative.
Senator Ron Wyden (D-OR) has a bee in his bonnet over government access to smartphone notifications. Megan and I do our best to understand his concern and how seriously to take it.
Wrapping up, Gus offers a quick take on Meta's broadening attack on the constitutionality of the FTC's current structure. David takes satisfaction from the Justice Department's patient and successful pursuit of Russian Hacker Vladimir Dunaev for his role in creating TrickBot. Gus notes that South Korea's law imposing internet costs on content providers is no match for the law of supply and demand.
And finally, in quick hits we cover:
- The guilty plea of the founder of a cryptocurrency exchange accused of money laundering
- Rumors that the ALPHV ransomware site has been taken down by law enforcement
- IBM's long-term quantum computing research milestones
- The UK's antitrust throat-clearing about the OpenAI-Microsoft tie-up
- And Europe's low-on-details announcement of a deal on the world's first comprehensive AI rules
You can subscribe to The Cyberlaw Podcast using iTunes, Google Play, Spotify, Pocket Casts, or our RSS feed. As always, The Cyberlaw Podcast is open to feedback. Be sure to engage with @stewartbaker on Twitter. Send your questions, comments, and suggestions for topics or interviewees to CyberlawPodcast@gmail.com. Remember: If your suggested guest appears on the show, we will send you a highly coveted Cyberlaw Podcast mug! The views expressed in this podcast are those of the speakers and do not reflect the opinions of their institutions, clients, friends, families, or pets.
The post Adventures in AI alignment appeared first on Reason.com.
