In this issue:
Congress needs your call before the next AI incident - A one-minute voicemail can do more than you think
Nothing to see here - There are two kinds of robot reality check. Which do you need?
Guest post by Haven Harms
Congress needs your call before the next AI incident
A one-minute voicemail can do more than you think
Picking up my phone
I dialed the number and waited expectantly as the phone rang.
“Thank you for calling the district office of Congressman Kevin Kiley,” a recorded voice began. “Our staff is currently unavailable. Please leave your name, contact information, the city you reside in, and your reason for calling, and we’ll get back to you as soon as we can.”
In stereotypical millennial fashion, I don’t make many phone calls these days. When I do, it’s for a good reason.
I heard the beep giving the cue that it was my turn to speak. “I’m calling about OpenAI’s disclosure this week that its AI models autonomously escaped containment and hacked another company, Hugging Face.” After naming AI policy safeguards I wanted Congressman Kiley to support, I finished, “We really need more federal oversight on AI development.”
In the past, I typically called my representatives to ask them to support or oppose a specific bill. After hearing about OpenAI’s models escaping a testing environment and hacking another company, however, I felt a sense of urgency to act. We don’t actually know how bad the situation is, as new information continues to emerge. For example, we’ve recently learned that other services beyond Hugging Face were targeted by OpenAI’s models, and that Anthropic’s AI models also hacked into other companies over the past few months. To me, this situation appears to be a major fork in the road, and whether our leaders act or not could mean the difference between a future where humans remain empowered, and one where humans lose control to AI. Future AI incidents could be much worse.
Our elected officials need to know what happened, and they need to take action. If you have concerns about the way AI development is headed, one of the best ways to communicate about those concerns is to call your representatives.
“But wait, I’m just one person,” you may object. “What could my phone call possibly change?”
Do politicians actually listen?
In 2017, Senate Republican leaders needed nearly unanimous party support to repeal key parts of the Affordable Care Act. Senators Susan Collins and Lisa Murkowski faced intense pressure from their fellow Republicans to support the measure. However, their offices also received calls, emails, letters, and personal stories from constituents urging them to protect the ACA. Ultimately, Collins and Murkowski ended up joining Senator John McCain and the Senate’s Democrats to defeat the measure 51 to 49.
When asked whether constituents’ stories affected the senators’ votes, Murkowski replied, “How could they not?”
When you call an elected official’s office, a staff member logs the issue, your position, and the action you want. Then, your message is tallied and routed to the staffer who covers that policy area. Elected officials will not personally hear every voicemail, but staff brief them on what constituents are saying. Calls signal which issues constituents care about and what they want their representatives to do. They prompt staff research and shape legislative priorities. As calls accumulate, this can change what an office pays attention to, what is included in a bill, and how representatives vote.
Politicians have listened on AI before
When SB 1047 was introduced in 2024, members of the AI safety community joined with allies such as SAG-AFTRA, the union representing actors and other media professionals, to mobilize in support. The California bill, introduced by Senator Scott Wiener, sought to reduce catastrophic risks from the most powerful AI systems through safety testing, security protocols, incident reporting, and other safeguards. It also faced fierce opposition, particularly from AI labs, investors, and parts of the open-source and startup communities.
As people reached out with concerns and proposed changes, Wiener’s office revised the bill. Some changes appeared within weeks of the feedback, including a provision considering how much money developers spent training a model when determining which AI systems the law would cover.
“It’s not about just winning the fight. I want to get this right,” Wiener said, crediting the people who worked to provide constructive feedback.
Supporters also called and wrote lawmakers as the bill moved through the legislature. SB 1047 ultimately passed both chambers but was vetoed by Governor Gavin Newsom. The process later helped give rise to SB 53, which enacted some of the AI transparency protections advocates had sought.
Calling for the first time
For Max Harms, an AI alignment researcher (and my husband), SB 1047 was the nudge that pushed him to call an elected official for the first time. Reflecting back, Max explained why he had avoided making calls like this before:
“I had this notion when I was younger that, even though people sort of wanted me to call my elected officials, I didn’t have a lot of power to actually influence things. That was the excuse that I often told myself when I was too lazy to actually contact people — sort of this sense of powerlessness and smallness.”
SB 1047 gave him a reason to act. The bill faced organized opposition, and its prospects were uncertain.
“And as a result, I was like, oh yeah, well, I’ll do my part as a California citizen. I’m sure there are a bunch of other people who care about this issue, but the fact that I live in California and the fact that this is a state-level issue means that I have a particular responsibility to step forward.”
Max expected the call itself to be more involved.
“I was sort of expecting to talk to a staffer, to have a real human being on the line who wanted to know things like where I live and what my phone number is, and would need to have me explain the situation to them. But I just got an answering machine.”
His message was brief.
“I said my name, where I lived, and what I cared about. I was like, ‘please vote yes on SB 1047.’ And that was it. It took, like, at most five minutes, but probably more like one.”
The hardest part was getting himself to make the phone call. It was something Max had never done before, and he worried that he might find himself in an awkward situation or say something dumb. The call itself was easy.
Immediately afterward, he felt “relieved and proud. And a little bit embarrassed, even — embarrassed that I had gone so many years with this impression that contacting my government representatives was an ordeal, like something I had to brace myself for, instead of something that would only take a couple minutes and not even require talking to a human being.”
The experience changed how he approaches political outreach.
“It’s not like I look forward to contacting people, but now when somebody’s like, hey, there’s an AI thing that has happened or some other thing that matters — I consider it a pretty easy thing to do to just put it on my to-do list, you know, during lunch or something. Just make a few calls.”
When asked whether the calls had been worth it, even though SB 1047 was vetoed, Max’s answer was immediate: absolutely. “Actions should always be judged on the information available to the action taker at the time that they were made. Anything else is outcome bias. If SB 1047 had not gotten vetoed, then my actions would have been just as right or wrong to have taken. And I think there’s a decent shot that it could have become law. It’s only through everybody speaking up that politicians can feel where the tide of public opinion is turning.” These days, his outreach efforts have turned to calling members of Congress, because AI is a global issue, and action can’t stop at the state level.
The void answers back
From the outside, calling a politician’s office can feel like shouting into the void, especially when all you reach is voicemail. However, as I was writing this article, my phone rang.
It was Congressman Kiley’s office.
The staffer summarized the message I had left: I was concerned about the OpenAI incident and believed stronger federal oversight of AI development was needed. Then he asked whether I had anything else to add. After I encouraged Congressman Kiley to support meaningful safeguards, the staffer said he would pass my comments along to the congressman.
It turns out I wasn’t shouting into the void at all.
Call your senators and your representative. Finding the right phone number takes about thirty seconds: house.gov has a lookup for your representative, and senate.gov lists contact information for both of your senators. While I’ve been writing from the U.S. perspective, the principle travels. In democracies everywhere, elected officials need to hear from their constituents about AI. Calling ahead of a vote matters, but you don’t have to wait. It’s likely that the OpenAI and Anthropic hacking incidents won’t be the last of their kind, and the safeguards we get may depend on whether officials hear from constituents before the next one.
Make the call, and find out how easy it is for yourself. Then tell your friends and family about the anticlimactic voicemail you left, and invite them to contact their representatives too.
Haven Harms is the founder and principal of Harms Research & Consulting, supporting organizations working on AI safety, governance, and advocacy.
Dispatch from Mitch
Nothing to see here
There are two kinds of robot reality check. Which do you need?
(With apologies to podcast listeners; you really need to see the videos at these links!)
Plenty of people, after you show them something like the robot martial arts demo from the Spring Festival Gala in February, will give you the kind of robot reality check that goes something like, “These robots are just running a pre-recorded routine. It’s not like they could actually fight you.” While this child and this engineer might beg to differ, the critique is probably true, for those robots.
Indeed, today’s multi-purpose robots have many shortcomings. They have to work as a team to make a bed. They’re slow and imprecise when folding laundry. They make a mess in the kitchen. The refrigerator they bring you might arrive with scratches and dents. They need multiple battery changes to run a marathon faster than the human record-holder. Sometimes, they turn out to be remotely operated, falling on their backs when the operator takes off their headset.
But I don’t show those videos to say, “Robots are ready to take your job and nunchuck you in a back alley.” The robot reality check I’m offering is that, just a year earlier, humanoids could barely do any of those household chores, the robot marathon took more than twice as long, and the Spring Festival Gala robots looked like this.
The progress is unlikely to stop. The AIs driving the robot bodies are getting dramatically smarter, and are training in accelerated simulations that allow them to acquire centuries of practice in days or weeks. At the same time, robot bodies are becoming more dexterous and reliable, thanks to soaring investment.
The sparks don’t fly until these trends combine to yield robots that approach the utility of paid humans in different roles. But don’t assume robots need to be faster, better, and cheaper than their replacements to start earning their keep. Excelling at any one of those traits may be enough to make inroads.
For many tasks, a dramatically lower cost can forgive limitations in speed or precision. Security guards and traffic cops probably fall into this category.
For tasks where speed or precision is everything, more specialized robots may win out even if they cost more to employ than the best available humans. I think some kinds of surgery and field welding probably fall into this category.
And then there are the tasks humans can’t do at all, like vaulting down a rocky slope, or spinning in the air so quickly that they become effectively transparent. (I can see military applications for both kinds of bots, though more conventional drones would probably do a lot more for the money.)
And humans, while very talented at picking things up with their articulated limbs, struggle to hover above the ground while doing so, unlike this multi-jointed aerial robot — though sometimes what you really want in that situation are aerial tentacles.
I’m on record recently as being deeply uncertain about whether humanoid robots will actually matter, because I think AI capable of piloting those robots well enough to replace a lot of jobs might not be far from leapfrogging to more exotic technologies. I suggested nanotechnology as one possibility, but mere microtechnology might do for a lot of roles — think swarms of insect-like bots buzzing around and teaming up to get things done. When I try to picture such swarms doing the same household chores the humanoids were bad at in the above videos, I get something like that scene from Snow White, only even more horrifying.
But nothing like that is here yet! So for now, enjoy your nunchuck supremacy, laugh at the latest robo-foibles, and whistle while you work.
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.
 | A guest post by
|