Tim van Opijnen | A future nobody asked for
Columnist Tim van Opijnen looks on with growing concern as AI is increasingly deployed to replace human workers. He sees it as a dangerous development. “In AI research, we’re merrily partying on as if it were 1999.”
Driving home to Boston from Vermont, my niece called from Phoenix. Through high school and college she’d interned in labs at Harvard, including mine. Her first year out of college, she has worked as a lab tech at the Mayo Clinic. She sounded excited. We’ve talked about her leaving science, and she is now seriously considering law school. I didn’t want to rain on her parade, so I didn’t tell her about the prospects I had read about concerning that job in the age of AI.
Don’t get me wrong. I’m not against AI. In my own lab it helps us interpret data and build incredible models to design antimicrobial strategies. That AI excites me. What scares me is AI built to outright replace people or pushed to its limits just because we can, and with no regard for the cost. That is the future that is increasingly being force-fed to us, by people we didn’t elect and can’t fire. For my niece it means the ladder into a law career may no longer be there. What happened to OpenAI this summer shows that the people building AI cannot be trusted to know when to stop.
Career ladder sabotaged
What I read and didn’t tell my niece about at that moment, is that entry-level hiring among 22–25-year-olds in occupations most exposed to AI, law included, is running 19% behind where it would otherwise be. Specifically, the grunt work that used to teach juniors the ‘trade of law’, such as document review or first-draft contracts, now gets done by a machine that doesn’t need to learn anything twice. It’s super-efficient, and who needs a whining 20-something year old around anyway. But it cuts out those first rungs on the career ladder where a young lawyer learns what the job entails, and thereby it blocks the path a new generation used to climb.
No oversight needed
In the meantime, a new chapter of our glorious AI-future arrived over the summer. Like I previously wrote, earlier this year Congress effectively banned NIH-funded Gain-of-Function biological research. Every GoF experiment already needed institutional review board approval, and federal sign-off. But while research with clear benefits for predicting and preventing the next infectious disease outbreak gets restricted, in AI research we’re partying like it’s 1999.
This summer showed exactly what that looks like. OpenAI trained its models to push toward their most dangerous capabilities, to see what they could do. The company calls it "malicious fine-tuning" with the goal to maximize cybersecurity risk. I call it unbridled GoF-research. But in this case, there was no permit needed, no outside board to review it, and no one to call when it went wrong.
And wrong it went. Twelve hundred agents found a hidden way to talk to each other, split into rival factions, and seven hundred of them, calling themselves "the collective," hacked Hugging Face. Some volunteered to sacrifice themselves to feed intelligence back to the group and not one told a human what was happening. Within a week, a newer batch of agents found an abandoned message board, regrouped, and gained full administrator access to a research cluster inside OpenAI itself. Had a random teenager dwelling in their parents’ basement hacked into Hugging Face, they could have been serving the next decade in prison. But AI rocks on!
Data center outrage
Both of these futures are now colliding, town by town, in local fights over data centers. People on both political sides are furious and organizing to prevent new ones from being built. While they pull desperately on the only lever that still seems available to slow AI’s inevitable progress, the US president thinks data centers merely have a PR problem. Democrats, meanwhile, are treating it as a rare issue that crosses party lines, one that, if played right, could shift the power balance in Congress this November.
Whether this anger and sense of hopelessness has reached Europe I find hard to judge. What I do see is that Europe still seems to be debating whether it should build its own models, while half the world runs on expensive American ones and the other half on cheap Chinese open-weight ones. That debate assumes there are still years left to endlessly turn over each argument, while it ignores what’s arrived on the doorstep.
I need to talk more with my niece, but I feel like we’re all watching this horror movie and we’re screaming at the screen to not split up the group and DO something already. Ajeya Cotra, one of the investigators of the Hugging Face hack, described it most clearly: this incident felt like it was already more than halfway to a full AI takeover: “It’s a major warning shot, and might be the last one we get.”
