In this issue:
Why aren’t the Hugging Face incident reports getting more media attention? - Five theories and a plea
Dispatch from Mitch
Why aren’t the Hugging Face incident reports getting more media attention?
Five theories and a plea
This is one of those days where a shortage of StopWatch-caliber AI news feels like the most important story.
There’s plenty of the usual business fare, coverage of data center backlash, copyright lawsuits, education issues, and so on. But there’s almost nothing new about loss-of-control risks, even though the Hugging Face incident reports have been out for five days now. Last week, the media passed along the “7,000 agents” figure from near the top of the reports and left it at that. Why?
Below, I’ll share some of my theories. But first, some explanations I don’t put much stock in:
A media conspiracy
Industry hush money
Overriding concerns about beating China
The new evidence just isn’t compelling
I don’t buy those explanations mostly because the media’s initial coverage of the Hugging Face incident was wide and deep, helping bring a number of important new facts to light. The new reveals provide even richer veins for journalists to mine.
The following theories seem far more likely to me.
Theory 1: People think they already know the story
They consumed the versions that came out in late July, and slotted the event into a narrative — e.g., “This was a case of AIs trying too hard to follow vague instructions”; or “This is fully explained by OpenAI having disabled some guardrails.” So people are seeing fresh chatter about Hugging Face but failing to double click.
If they haven’t looked into the new reports, they do not know the story.
Theory 2: The new details seem too boring and technical to read and share
This was arguably true for the first day or two after the reports were released. For someone with limited time and/or a short attention span, the reports would have been easy to bounce off after getting the initial “7,000 agents” figure. But the interesting parts of the reports have now been extensively highlighted and repackaged: There was my own quick stab at this on Thursday. Zvi Mowshowitz has put out a whole series of posts.
There are many others. But the retelling we insiders all seem to be holding up as our champion is The Rise and Fall of Agent Civilizations from podcaster Dwarkesh Patel. In plain language, Patel captures the grand scale and unprecedented nature of the incident, drawing clear circles around the known unknowns that must be investigated. I learned a few things from it that I had missed in my own scans of the evidence.
Read it. (Or listen to his audio version.) You’ll be hooked from the opening lines:
Over the course of three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself.
Theory 3: People had already misinterpreted earlier coverage as (the start of) the worst case scenario
I think a lot of people not immersed in AI safety issues are unfazed by the new information — if they hear about it at all — because they had already assumed the worst. Hearing about Hugging Face for the first time in July, they said, “Yep. Baby Skynet is here. These idiots will just keep feeding it until it kills us all.”
The earlier known facts, while very alarming, failed to capture just how far beyond their task the offending agents went. So to someone like researcher Ajeya Cotra, last week’s findings were terrifying revelations. But if you had already leapt to the scariest conclusions, the new reports don’t feel like news.
Theory 4: The news cycle and the AI “beat” have limited attention that is currently turned elsewhere
Media outlets mostly treat AI as a beat: a narrow reporting lane. The AI beat is often part of the business or technology section. For outlets with tight beats, the amount of coverage they can run on AI has a soft cap tied to the amount of staffing assigned to cover that beat.
If the limited number of reporters on the AI beat are already busy covering other AI topics, then returning to Hugging Face has to wait until the news cycle moves on. This seems plausible to me, because total AI coverage is actually quite high right now, and dominated by data center backlash, which recently became perceived as an acute election-season crisis for politicians and AI companies.
So there may not be much leftover capacity or appetite for fresh stories about loss-of-control risk coming out of an incident whose main victims (more AI bros) aren’t sympathetic to most readers and aren’t even complaining.
Theory 5: Investigative journalists are looking into the reports, and we just need to give them more time
I’m confident that this is at least a little bit true. The herding of the media — where journalists tend to focus their attention on the same stories at the same time — is partly a function of the fact that only a small subset of journalists are assigned and resourced to do original, independently-researched reporting. These are the trailblazers that help initiate fresh news cycles. If or when some of these trendsetters share what they glean from the new reports and their contacts at OpenAI, I think we’ll see a broader media wave ripple out from it.
The risk acknowledgement problem
Of my five theories, I’m putting most of my weight on 1 and 2: that people just haven’t realized that they don’t actually know the Hugging Face story, and don’t know how compelling that story is.
But there’s a larger post-Hugging Face pattern lurking in the media’s AI coverage I should talk about.
Ever since the first reports of the incident, my monitoring tools for tracking article framings have shown a rise in stories that take catastrophic risk concerns seriously, but a much, much larger rise in articles that minimize, reframe, or dismiss such risks.
These minimizing frames aren’t usually the focus of those pieces. I don’t usually get the sense that anyone woke up that day determined to downplay catastrophic risk. Instead, I see a pattern where people who have other AI topics to write about make space for those topics by pushing catastrophic risks off to the side. They open with a thousand variations of “AI companies are struggling to maintain control of their creations. But the more pressing concern is...[something else].” Then they write about that something else.
On the one hand, this signals that it’s socially acceptable to talk about rogue AI. But on the other hand, this tags the issue as a back-burner problem for tomorrow — and tomorrow never comes.
I get it. If you’re a journalist, a blogger, or a video essayist, you can’t only talk about AI’s biggest risks, even if you have the professional freedom to do so. If you repeat the same story enough times, it stops being news, and you lose the audience that most needs to hear it.
And we do have other things to worry about! We’ll still have plenty of problems if we successfully avert extinction. Besides, very few people can productively fill their days entirely with work against catastrophic risks, and those who try tend to burn out or go crazy.
But I hope the media will do more to help the public understand that a chaotic buzz of human activity at the edge of a cliff doesn’t mean there’s not a cliff. Sometimes, you need to write about the cliff.
Now seems like one of those times.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.