Collabs, creators and crawlers
And how exactly could the threat from AI play out in elections?
Welcome to our new subscribers this week! If you haven’t subscribed, join us:
This week, OpenAI allows websites to opt out of being included in training data sets; an examination of how AI might influence voters; and US states and cities are positioned to regulate AI before the federal government.
Copyright, creators, and crawling
Large AI models work because they synthesize the text and images of millions of creators for free; many of the creators of those texts and images produced them as paid work. The tension has already led to a number of class-action lawsuits and the economic and legal landscapes are shifting rapidly.
Historically secretive about its data sources to shield itself from inquiry, OpenAI cracked open the door this week, letting websites opt out of being included in future training data. It’s modeled after long-standing Internet practices that allow websites to disallow search engines from “crawling” their contents. Most websites don’t opt out – if you let Google in, you are rewarded with more traffic to your site. If you let OpenAI in, you’ll help their “models become more accurate and improve their general capabilities and safety” – and why should publishers care about helping OpenAI optimize their models when their livelihoods are at stake?
The economic equation is not the same across all AI companies; for instance, publishers might have an interest in opting into products like Google’s Bard, which includes links out to its sources and could lead to traffic (and ad revenue). Since there’s no standardization, each publisher will have to grapple with the pros and cons of letting their data be harvested by each individual AI company. In the absence of court guidance about whether the scraping and ingestion of data to train AI models is covered under fair use, we’re heading into a messy near-future where both publishers and AI companies will be navigating a gray area, with a likely exploration of incentives for publishers to participate in these models.
These tensions are playing out differently across creative industries. In music, an industry with a long track record of leveraging the courts to fight back against disruptive technologies, a partnership is reportedly brewing between Google and Universal Music with a goal to develop a tool that enables fans to create “deep fake” songs where artists opt-in to participate and can earn a revenue share in exchange[1].
In Hollywood, amid more existential threats around streaming economics and residual payments, the unions are working to prevent studios from using AI to take one-time work – whether it be joke-writing for screenwriters or digital replicas for actors – and leveraging it in perpetuity without compensation. While they are currently seeking bans of these AI use cases, nobody is arguing that AI can be put back in the bottle, and so these negotiating positions may ultimately evolve towards something that looks more like the potential Google/Universal Music partnership.
And meanwhile in the publishing industry, AI-generated travel guides have taken over Amazon, complete with hundreds of fake reviews (and disgruntled customers).
[1] Last week, Meta released MusicGen, a music generator that used fully licensed training data. Coincidence?
What exactly are the electoral threats of AI?
Last week, Sam Altman, CEO of OpenAI, tweeted his concerns about the role AI might play in upcoming elections. The broad threat is real, but not exactly for the reasons Altman thinks. In contrast to popular perception, ”personalized” persuasion via 1:1 channels (canvassing, phone calling, texting) is notoriously difficult, and as a result this kind of targeted messaging is typically reserved for voter ID and turnout, not persuasion.
Mass media like television and digital video ads can be more effective at persuasion but it’s unclear how much of an impact AI would have. These channels are one-to-many broadcasts, with even highly targeted digital ads usually reaching thousands of people at a minimum and television vastly more.
A lurking threat is disinformation distribution over existing private messaging networks like WhatsApp. Disinformation in group messaging was a threat even before the rise of consumer-facing generative AI tools, but deep fakes amplify the threat. They’re also nearly impossible to regulate and operate largely out of view of journalistic scrutiny as opposed to the large transparency, disclaimer, and identity verification systems implemented by Meta and Google for their traditional advertising businesses.
There are regulatory possibilities to build on the platforms’ work to further limit harms by way of disclosure laws. California already has a law on the books requiring disclosure of the use of AI-powered political chatbots, and it could have real teeth if expanded federally and/or modified to include a private right of action. Despite gloomy news just weeks ago that the FEC had punted on the issue of AI disclosures in campaign advertising, saying they didn’t have the authority to regulate it, yesterday they revisited the issue and moved it forward, and we’ll see if any further action is taken over the coming months.
States and cities are moving faster on regulation
In contrast to last week’s news that Congressional AI regulation is highly unlikely anytime soon, state legislatures are showing an appetite for action. Following on the heels of a few initial attempts at legislation in California, a coalition of CT, CO, NY, VA, and MN lawmakers are working on model AI legislation focused on “broad guardrails,” product liability, and impact assessments of AI systems - largely those in use by state agencies themselves. We’ve long hypothesized that states are more likely to get regulation off the ground before anything happens at the federal level for a number of reasons (gridlock, for one). Widely duplicated model legislation eases the burden of complying with 50 different state-level regulations, and the rapid pace of technology development and uncertain impacts lend themselves well to fast action on a smaller scale that can be expanded federally if successful.
And don’t discount cities, which are generally able to move even faster than state legislatures. New York City’s anti-bias law for hiring algorithms went into effect last month. The San Francisco Standard reported this week on how Boston, Seattle, and San Jose are setting guidelines for the use of AI by municipal employees, and on Thursday California regulators approved the expansion of driverless taxis in San Francisco, the epicenter of AI technology.
Of note
Regulation
AI regulation is taking shape, but startups are being left out (The Verge) Smaller startups advocate for a voice in the AI regulatory conversation.
The S.E.C.’s Chief Is Worried About A.I. (The New York Times) Finance has long been hyper-optimized and automated, and Gensler seems to be repackaging valid-but-classic concerns for new media interest in AI regulation here.
Bias
Eight Months Pregnant and Arrested After False Facial Recognition Match (The New York Times) The City of Detroit already faces three lawsuits for wrongful arrests based on facial recognition errors.
AI is acting ‘pro-anorexia’ and tech companies aren’t stopping it (Washington Post)
AI for good
How AI is helping airlines mitigate the climate impact of contrails (Google) Google, American Airlines, and Breakthrough Energy used AI to analyze satellite imagery, weather, and flight path data to reveal that optimizing flight paths could reduce aviation’s climate impact by up to ~18%.
What might LLMs/generative AI mean for public benefits and the safety net/tech? (Dave Guarino) How LLMs can help with providing social and government services.
More
Technically Optimistic (Emerson Collective) A six-part limited series podcast from Emerson Collective CTO and former DNC CTO Raffi Kirkorian, diving into AI regulation, education, creativity, and AI management.
Why ChatGPT Is Getting Dumber at Basic Math (Wall Street Journal) Some have speculated that the large models can only hold so much ‘in their heads’ – and the increasing need to add safety features is displacing other knowledge.
Vassals vs. Rivals: The Geopolitical Future of AI Competition (Lawfare)
Every start-up is an AI company now. Bubble fears are growing. (Washington Post) Examining whether reality will match AI hype.