~~ recommended by newestbeginning ~~
 |
In the debate about what threats artificial intelligence poses, everyone is trying to sell you something. If you follow the corporate press, artificial intelligence is a looming mass extinction event waiting to happen. If you listen to Trump, it’s part of the next Cold War with China, meaning we can’t afford to slow down.
For the appistocracy, there’s profit in all of the attention — literally. We have a new generation of technologists who generate publicity warning of everything from bioterrorism to the machines taking over. Thus AI is turned into a “national security” issue; and as such, it is elevated to the same status as nuclear weapons, deemed too important for interference from a meddlesome public.
The reality though? Listen to the spy agencies. Though they’re far from immune to threat inflation, these agencies echo none of the hysterics dominating the news. As the U.S. intelligence community sees it, AI simply magnifies threats that are already there. And that’s basically right. (More on this in a moment.)
We’ve all seen the sensational headlines. And it’s not just the old school media. Social media influencers credulously amplify this coverage to their millions of followers, going viral with talk of human extinction. Thus dire warnings about a looming AI apocalypse (called things like catastrophic risk, existential risk or X-risk) dominate Americans’ collective imagination.
 |
Then there are the anti-doomers, who run a racket of their own. The most prominent are Yann LeCun, until the end of last year Meta’s chief AI scientist, and Andrew Ng, who co-founded Google Brain and now runs the venture studio AI Fund.
In October 2023, LeCun accused Sam Altman, Demis Hassabis and Dario Amodei — the heads of OpenAI (ChatGPT’s developer), Google DeepMind (Gemini) and Anthropic (Claude) — of “massive corporate lobbying” aimed at regulatory capture, warning that if their fearmongering succeeded, a small number of companies would end up controlling AI. Ng, in an interview the same week, called the idea that AI could drive humans to extinction a “massively, colossally dumb idea,” and said large tech companies were manufacturing extinction fear because they would rather not compete with open source.
They may well be right about the motive. But this is a fight between two business models, not between an industry and its critics.
LeCun and Ng are the two loudest champions of open-weight AI — models a company gives away, so that anyone can download the finished product, run it on their own machines and modify it, free, with no permission required. Anthropic does the opposite: it keeps its models on its own servers and rents access. So does OpenAI, with everything at the frontier. A law requiring a government license to release a powerful model would fall hardest on the side that gives its models away for free.
So Anthropic and OpenAI’s loss is people like LeCun and Ng’s gain. They’re defending a different kind of racket: a regulatory regime with no frontier-model licensing requirements, in which the value accrues to the startups rather than the labs.
What every faction has in common —from the proponents of some AI apocalypse to the anti-regulation types — is their disinterest in responding to the realistic risks AI poses.
The U.S. intelligence community provides the most reasonable account of the actual dangers. The Annual Threat Assessment is the U.S. intelligence community’s yearly inventory of everything that might plausibly threaten national security: China’s military, Russian sabotage, fentanyl, lone-wolf shooters, drug cartels, collapsing states. It is the closest thing the government produces to an official list of what it’s afraid of.
In this year’s report, artificial intelligence is described as a threat in three different ways:
“[I]t is essential to make sure that humans maintain control of the machines and how AI is used.”
“These applications, however, also carry risks that require careful human engineering to appropriately mitigate risk of AI autonomy before they are broadly deployed.”
“[E]merging technologies such as AI and quantum computing are expected to have a significant impact on national security.”
That’s the sum total of what the eighteen agencies of the U.S. intelligence community have to say about the dangers of AI. No human extinction, no superintelligence, no machine uprising.
As the most recent Department of Homeland Security Threat Assessment puts it AI is merely “adding layers of complexity to the threats we face.” It identifies the following five problems:
Disinformation, deepfakes, and election interference,
Cyber operations and financial crime,
Violent extremist exploitation and radicalization,
Proliferation of chemical and biological knowledge, and
Breaking AI's safety controls, and poisoning the data it trains on.
Virtually all of these threats existed before AI. For instance, AI-fueled disinformation is simply more of the same with better production values. Cyber operations and financial crimes are cyber operations and financial crime, just conducted faster or with greater stealth. “Extremists” were radicalizing people on message boards for two decades before anyone typed a prompt. The chemical and biological concern goes back to at least the Clinton administration, biotechnology being the last apocalyptic threat. And the fifth is a threat to AI systems, not from them: poison the training data and the expensive machine learns what you want instead of what its owners intended.
None of this is nothing. It’s just not the end of the world.
When Congress passed the Global Catastrophic Risk Management Act in 2022, it forced the executive branch to take a formal look at everything that might end civilization over the next thirty years. RAND’s Global Catastrophic Risk Assessment, produced for the Department of Homeland Security, categorizes AI primarily as an “entropy” engine.
“One can think of AI as adding entropy and chaos to already-wicked problems that humans face,” the RAND report says. “Chaos does not require the development of superintelligent or supercapable AI and is possible with existing and near-future AI capabilities.”
 |
RAND is not a dovish institution. The Pentagon-funded think tank was built to think about thermonuclear war.
Its companion study, On the Extinction Risk from Artificial Intelligence, explored similar questions specific to AI. Rather than speculate about machine consciousness, the authors went looking for the physical mechanism. If AI is going to kill every human being, how, specifically? They examine three plausible routes: nuclear fallout, engineered pathogens, and deliberate geoengineering.
Here’s the part that never makes the Sunday shows.
The report concludes that there are major constraints preventing an AI-driven extinction event. They are physical and operational barriers — such as processing, weaponizing, and physically deploying pathogens, maintaining infrastructure without human maintainers, and controlling physical industrial or nuclear system.
RAND also ran a red-team exercise to see whether large language models help people plan a large-scale biological attack. The authors found no statistically significant difference in the viability of plans produced with the models versus without them.
The National Academies of Sciences, in another congressionally chartered review of AI in the life sciences, echoed RAND’s conclusion, adding more detail. It says that the bottleneck to bioterrorism has never been information. It’s DNA synthesis screening, tacit laboratory skill, culturing, formulation, aerosolization — the physical, expensive, failure-prone middle of the process.
What all of these studies point to is that the boring risks are the real ones. Yet the doomers have made the risk sound like a movie, which causes the actual risks read as anticlimax by comparison. A poisoned training set, a spoofed sensor, an early-warning system that sees a missile that isn’t there, a protein design tool that outruns the screening software. These are solvable problems with identifiable owners and unremarkable price tags, and they are being crowded out of public attention by a debate over whether a chatbot is going to become God.
The government’s own analysts, who have no consumer product to move and no valuation to defend, have been telling us the answer for years in language so measured that nobody has bothered to read it. AI merely increases the magnitude and speed of threats we already have, ones that can be solved and will exist regardless of AI.
Is hyperbole the only register we’re able to communicate in?
It currently is, and none of this is going to change on the supply side. Businesses are going to cast their products in the most favorable light. Media is going to go with the headlines that draw attention.
I know, because I sometimes do it. When I published a piece last week headlined “Fuck 9/11,” I also knew the headline would generate pearl-clutching outrage, and that the outrage would carry the story a lot further than the argument would have on its own.
In a debate centered on whether a chatbot is going to become God, every major faction is cashing in — at the expense of us all.
No comments:
Post a Comment