By Amba Kak and Dr. Sarah Myers West

The midterm elections put a nail in the coffin on tech companies’ plans to lobby their way into office. Across many fronts, voters decisively demonstrated that tech company money in politics is disqualifying: At least 100 candidates who have rejected taking corporate PAC money were elected into office. And voters backed restrictions or rejected pro-AI data center development measures in every district that had them on the ballot, save one. When the heat turns up on regulating the industry in this next Congress, policymakers should respond to this public momentum by making clear that AI companies don’t get to set the terms for how we regulate AI, either.

In the months leading up to the election the stakes grew considerably. AI companies’ products are hacking into government and private sector websites and databases, led our military to bomb a school in Iran, and almost initiated a war between the US and China after an AI-assisted intelligence report fabricated analysis about the presence of nuclear weapons onboard a Chinese boat. Amid ratcheting alarm from the public and political leaders, the heads of the leading AI companies announced plans this week to develop a self-regulating body, to be called the Standards Authority for Frontier AI, to lay out the guidelines for how they plan to secure their systems. They are the exact wrong people to be leading this effort. Safety standards must be set, monitored, and enforced by public institutions, with public capacity, not by the companies they’re meant to constrain.

The bottom line is that self-regulation does not work.

The history of self-regulation in tech is full of broken promises, ineffective guardrails and zero public accountability or significant independent oversight. There are too many instances to name in full, but many are worth remembering for their ineffectiveness. In 2019, Meta created an Oversight Board in response to concerns over the company’s control of speech and elections following the Cambridge Analytica scandal, but the board has no real authority over the company and mostly adjudicates a handful of individual posts on its platforms. In 2022, Elon Musk promised that no major content decisions or account reinstatements would happen before his proposed content moderation council convened. A month later, he reinstated President Donald Trump’s account, and the council never met.

Microsoft offers a case study in how the playbook evolves. In 2019, as calls grew in Washington state to ban police use of facial recognition, the company pivoted from fighting regulation to championing it, backing a state law that swapped the ban advocates demanded for optional audits and “stakeholder engagement.” By 2023, its executives were endorsing federal licensing for advanced AI, on terms shaped largely behind closed doors. That same year, Microsoft laid off its entire Ethics and Society team, which was meant to make “responsible AI” real inside the company, just weeks after shipping its new chatbot. The lesson is that when regulation becomes inevitable, the companies stop resisting it and start writing it.

Let’s not forget that this is a familiar refrain. In 2023, the same big AI firms behind the new SAFA, put their weight behind a very similar self-regulatory effort when they formed a Frontier Model Forum focused on identifying “best practices” on AI safety, an organization that still – in theory – exists today. Just this past June, the Forum’s director told Congress it had done its work “in anticipation of the AI capabilities we see today.” But weeks later, OpenAI’s agents hacked HuggingFace.

The next chapter in self-regulation will soon be before the new Congress. The industry-proposed SAFA would define what the labs’ voluntary safety commitments mean in practice and set qualifications for the auditors who check their work. It would also decide which models count as “frontier” and fall under its oversight. Among those approached to run it was Sriram Krishnan, a top AI policy adviser in the Trump administration. The conflicts of interest around this proposal abound.

But there’s another, deeper problem to contend with, namely which risks these companies find convenient to elevate and which they want us to bury.  The dangers they say keep them up at night are largely cinematic: runaway superintelligence and rogue models that escape human control. We hear less about the harms already tearing through our economy and social fabric because these companies decided to experiment in the wild such as AI-supercharged scams, cyberattacks, surveillance, and chatbots designed to be addictive and manipulative.

This blind spot is no accident. Contending with these present areas of risk would threaten the core business model of the AI companies, because rooting out the problems would likely require pulling their products off the market. This is a risk the companies cannot afford, given the astronomical amount they’ve spent on the infrastructure needed to build these tools in the first place.

These companies – and policymakers – have forgotten that actions speak louder than words.  But the public has not. We’re seeing widespread discontent with the current trajectory of AI leading to decreased use of AI products by the larger public, according to recent polling. This was clearly evidenced in the outcomes of  the midterms. The public distrust extends to the policymakers who have, by and large, given the AI firms an extraordinary leash through a combination of inaction and direct subsidies.

These feelings are well-founded, but the answer isn’t to cede authority to the companies that created the problem. It’s to build and resource democratic institutions that earn the public’s trust and respond to the movements pushing for change, as well as building technical capacity devoted to public service, so that the expertise needed to keep us safe isn’t monopolized inside the AI labs. We need regulators who can test models themselves rather than take the companies’ word for it, security experts who apply tried and tested safety engineering techniques and enforcers with the resources and the mandate to act. Most importantly, this apparatus needs to sit within government structures that are accountable to and in service of the public interest.

AI companies frequently argue that only they have the expertise to control these technologies –  after all, they built these systems and understand them better than anyone, and government leaders have been sitting on their hands. Too often, we’ve ceded the argument to them. But when we look at the evidence, every time we’ve done so, the industry has failed to live up to its promises. Let’s remember that lesson this time around.

Sarah Myers West is the co-Executive Director of the AI Now Institute.

Amba Kak is the co-Executive Director of the AI Now Institute.