Asking for forgiveness, rather than permission, is Silicon Valley’s favorite business model—from Uber’s early days entering cities without seeking approval from local officials to the social networking companies’ loose treatment of user data.
With the AI market booming, the forgiveness cycle is kicking into high gear once again.
Consider Google’s latest AI imbroglio. On Thursday, the company published a lengthy blog post explaining why its new AI search—a feature that it automatically activated for all U.S. users this month, without any ability to opt out—was telling people to put glue on their pizzas and to eat rocks.
It turns out AI search isn’t smart enough to recognize satirical and troll-y content that exists on the web (especially in online discussion forums like Reddit), Google acknowledged. As a result, the company is now limiting the amount of such content it includes in its AI search results.
The incidents “highlighted some specific areas we needed to improve,” Google VP Liz Reid wrote.
Also this week we finally heard from Helen Toner, the former OpenAI board member who was ousted in the fallout of the Sam Altman crisis last year. (The board, as you’ll recall, had briefly fired Altman for not being “consistently candid” in his role as CEO.)
According to Toner, one of the reasons the board lost trust in Altman stemmed from the launch of OpenAI’s most popular product, ChatGPT, in November 2022. The board was never informed beforehand of the launch, Toner claims, and found out about it after the fact, as people were discussing it on Twitter.
None of these incidents are catastrophic—hopefully, no one was daft enough to add glue to their pepperoni pizza—but they underscore an entrenched behavior in Silicon Valley that shouldn’t be glossed over at a time when we’re trying to determine how much regulation to impose on the AI industry and how much to allow the industry to regulate itself.
There are signs that tech companies are acting more responsibly. In recent weeks, OpenAI has signed a string of multimillion-dollar deals with publishers such as Vox Media, The Atlantic, and News Corp. The deals allow OpenAI to train its large language models on the content of these publishers, rather than just scrapping it all off the web for free.
Of course, OpenAI is currently being sued by the New York Times for allegedly doing exactly that. Would any of these content deals be happening if OpenAI hadn’t already been challenged for its behavior?
Alexei Oreskovic
Want to send thoughts or suggestions to Data Sheet? Drop a line here.
Today’s edition of Data Sheet was curated by David Meyer.
NEWSWORTHY
TikTok’s U.S. copy. TikTok is creating a U.S.-specific clone of its recommendation algorithm that could operate independently of the China-owned company’s main code, Reuters reports. Although parent ByteDance apparently ordered this effort before the U.S. passed a law giving it the choice of selling TikTok or seeing it banned in the U.S., the report suggests that the code split could make such a sale easier. The Chinese government says it opposes the sale, and it would be able to block the export of TikTok’s recommendation algorithm as it stands.
OpenAI reports on disinformation. OpenAI has for the first time detailed how influence operations have been using its AI. In a blog post, OpenAI said it had terminated the accounts of five such operations, which were the work of groups in Russia, China, Iran, and Israel. The firm said they had used its tech to generate texts, including long articles and fake replies on social media—which apparently failed to attract much real engagement. As NPR notes, OpenAI’s move comes just after Meta gave its own report on similar issues.
Siri to tap apps. Bloomberg reports that Apple’s soon-to-be-overhauled Siri will be able to control functions within apps on behalf of users. Apparently, Apple will start with its own apps. The report also says that, although Apple “has been forging a deal with OpenAI” to bake ChatGPT into iOS—another report earlier this week said that was a done deal—Apple is also still talking to Alphabet about using its Gemini AI in the future.
Drop your market share. Tencent’s WeChat is China’s biggest mobile payment player, followed by Alipay. According to Nikkei, Chinese regulators have asked Tencent to reduce WeChat’s market share in its home country, though “it is not clear whether Tencent was given any precise numerical target to meet.”
IN OUR FEED
“People have been claiming [Y Combinator] fired Sam Altman. That’s not true.”
—Y Combinator cofounder Paul Graham hits back at claims that the startup accelerator sacked Altman as president for putting his own interests first. Having withheld comment when the Washington Post first published that claim half a year ago, Graham now says YC merely gave Altman the choice between running YC and running OpenAI, once the latter company launched its for-profit subsidiary with Altman as CEO. Some people have recently been citing Altman’s YC departure as an example of him being called out by an employer for untrustworthiness.
IN CASE YOU MISSED IT
Trump’s media company stock spirals in the wake of fraud verdict, by the Associated Press
Tesla claps back at adviser criticizing Elon Musk’s stock options—giving the Tesla CEO $45 billion is the ‘ethical’ thing, by Marco Quiroz-Gutierrez
Some TikTok creators surprisingly support a TikTok ban to protect national security and help people ‘live in the moment again’, by Alexandra Sternlicht
AI companies are competing for talent with more than pay packages and GPU access. Their position on AI safety may be starting to matter too, by Sage Lazzaro
Meta AI can answer search queries and recommend a cocktail bar—but you can’t really turn it off even if you want to, by the Associated Press
BEFORE YOU GO
Spotify’s Car Thing. Enraged buyers of Spotify’s Car Thing streaming device—which the company will deactivate later this year—this week filed a class-action lawsuit against Spotify over the move. And now TechCrunch reports that Spotify has agreed to refund those who bought the automotive gadget during its brief window of availability in 2022, as long as they have proof of purchase.
