I think my main objection is a variant of #1: Voters don't care about AI safety and focusing on a single issue will reduce the party's ability to gain support
In fact, some and perhaps many care about AI safety. However, almost none care ONLY about AI safety. A political party (or governing coalition in parliamentary systems) needs to have positions (and, if successful, actually pursue and enact them) on every single issue. Even for voters for whom AI is their top issue, it may not even be 20% of their policy-preference set.
I further think that if there's enough support to make an AI Safety party, with no other significant topical positions, there's MORE than enough for members of existing parties to adopt that as part of their platform.
Thanks for stating your objection to my argument. I agree with you and that is why I argue for the formation of a centrist party with AI Safety as an important issue, rather than a party which literally only speaks about AI, as the latter party would come across as unserious (even the Green Party does not do this with climate change). You can see the "What about other issues?" section about some general positions I think a party should take.
I also think an effective political party could grow support for AI Safety by leveraging media coverage in order to bring to light the dangers of AI and the necessity of regulation, so even if it starts as a fringe prospect, it will gain momentum over time.
To your second point about existing parties being able to adopt this as part of the platform, along with the potential bad optics of EAs/AI doomers "invading" existing parties, I think it is simply harder to succeed in political party primaries than as a third party candidate due to political party primaries having an unlevel playing field which favors catering to the idiosyncrasies of the party's base and major party players.
To back this up, I do not see Chase Oliver getting anywhere as a Republican nominee, but he did gain some notoriety as a third party candidate. While it is in theory possible for pro-AI Safety rationalists to gain power through traditional parties, I feel like this would take a very long time and is less likely to succeed.
Ultimately, however, even if you think this has a slim chance of succeeding, I see some electoral effort as the only plausible way of achieving necessary government support for AI Safety, so I feel like something like this is still worth it if there is not a clear alternative.
Artificial Intelligence is advancing rapidly, raising significant concerns about its safe development and deployment. Despite widespread public concern about AI, there is a notable absence of a sustained political movement dedicated to addressing these issues. While certain organizations and individuals are engaged in shaping AI policy, these efforts have generally avoided the domain of electoral politics.
This post argues for the establishment of a centrist political party that is firmly committed to promoting AI safety. The argument rests on three main claims:
Governments are needed to promote AI safety because the current dynamics of free-market AI development make voluntary caution difficult, and because AI carries unprecedented risk and transformative potential. Furthermore, the US government can make a huge difference for a relatively insignificant slice of its budget.
While self-regulation should be an important component of AI Safety, there are many reasons to believe it might not be sufficient for ensuring AI is developed safely.
Firstly, there’s potentially a massive first-mover advantage in AI. The first group to develop an AI system which dramatically increases the productivity of ML research could theoretically secure overwhelming economic power by utilizing said AI to kick off a chain of recursive self improvement. This could start with human AI researchers gaining dramatic productivity boosts by using AI, which then allows for the development of more advanced systems which boost productivity even more, and finally results in the vast majority of AI improvement being driven by AI systems alone, with the only effective constraints on improvement being compute and energy.
As a result of such reinforcing feedback loops, the first company to create AI capable of greatly accelerating AI research could establish an insurmountable advantage, first by improving the capabilities of their AI systems far beyond the competition, then by using such systems to gain economic resources and power.
Even without recursive improvement, however, being a first mover in transformational AI could still have dramatic benefits. By establishing brand loyalty among customers attracted to transformational AI and increasing access to investment capital by demonstrably being the leading AI company, first movers could gain a significant advantage over competitors.
Thus, incentives are distorted accordingly. Major labs are pressured to move fast and cut corners—or risk being outpaced. Slowing down for safety often feels like unilateral disarmament. Even well-intentioned actors are trapped in a race-to-the-bottom dynamic, as all your efforts to ensuring your model is safe is not that relevant if an AI system developed by another, less scrupulous company becomes more advanced than your safer models. Anthropic puts it best when they write: "Our hypothesis is that being at the frontier of AI development is the most effective way to steer its trajectory towards positive societal outcomes."
The actions of other top AI companies also reflect this dynamic, with Meta having spent hundreds of millions if not billions of dollars to poach talented individuals from other firms, and many AI firms barely meeting basic safety standards.
This is exactly the kind of environment where governance is most essential. Beyond my own analysis, here is what notable advocates of AI safety have said on the necessity of government action and the insufficiency of corporate self-regulation:
“‘My worry is that the invisible hand is not going to keep us safe. So just leaving it to the profit motive of large companies is not going to be sufficient to make sure they develop it safely,’ he said. ‘The only thing that can force those big companies to do more research on safety is government regulation.’”
“I don't think we've done what it takes yet in terms of mitigating risk. There's been a lot of global conversation, a lot of legislative proposals, the UN is starting to think about international treaties — but we need to go much further. {...} There's a conflict of interest between those who are building these machines, expecting to make tons of money and competing against each other with the public. We need to manage that conflict, just like we've done for tobacco, like we haven't managed to do with fossil fuels. We can't just let the forces of the market be the only force driving forward how we develop AI.”
“Many researchers working on these systems think that we’re plunging toward a catastrophe, with more of them daring to say it in private than in public; but they think that they can’t unilaterally stop the forward plunge, that others will go on even if they personally quit their jobs. And so they all think they might as well keep going. This is a stupid state of affairs, and an undignified way for Earth to die, and the rest of humanity ought to step in at this point and help the industry solve its collective action problem."
Beyond the argument from competition, there is also the question about who gets to make key decisions about what type of risks should be taken in the development of AI. If AI has the power to permanently transform society or even destroy it, it makes sense to leave critical decisions about safety to pluralistic institutions rather than unaccountable tech tycoons. Without transparency, accountability, and clear safety guidelines, the risk for AI catastrophe seems much higher.
To illustrate this point, imagine if a family member of a leader of a major AI company (or the leader themselves) gets late stage cancer or another serious medical condition that is difficult to treat with current technology. It is conceivable that the leader would attempt to develop AI faster in order to increase their or their family member's personal chance of survival, whereas it would be in the best interest of the society to delay development for safety reasons. While it is possible that workers in these AI companies would speak out against the leader’s decisions, it is unclear what could be done if the leader in this example decided against their employees' advice.
This scenarios is not the most likely but there are many similar scenarios and I think it illustrates that the risk appetites, character, and other unique attributes of the leaders and decision makers of these AI companies can materially affect the levels of AI safety that are applied in AI development. While government is not completely insulated from this phenomenon, especially in short timeline scenarios, ideally an AI safety party would be able to facilitate the creation of institutions which would utilize the viewpoints of many diverse AI researchers, business leaders, and community stakeholders in order to create an AI-governance framework which will not give any one (potentially biased) individual the power to unilaterally make decisions on issues of great importance regarding AI safety (such as when and how to deploy or develop highly advanced AI systems).
Finally, I think the massive resources of government is an independent reason to support government action on AI safety. Even if you think corporations can somewhat effectively self-regulate on AI and you are opposed to a general pause on AI development, there is no reason the US government shouldn't and can't spend 100 billion dollars a year on AI safety research. This number would be over 20 times greater than 3.7 billion, which was Open AI's total estimated revenue in 2024, but <15% of US defense spending. Ultimately, the US government has more flexibility to support AI safety than corporations, owing simply to its massive size.
Despite many compelling reasons existing for the US government to act on AI safety, the US government has never taken significant action on AI safety, and the current administration has actually gone backwards in many respects. Despite claims to the contrary, the recent AI action plan is a profound step away from AI safety, and I would encourage anyone to read it. The first "pillar" of the plan is literally "Accelerate AI Innovation" and the first prong of that first pillar is to "Remove Red Tape and Onerous Regulation", citing the Biden Administration's executive action on AI (referred to as the "Biden administration's dangerous actions") as an example, despite the fact the executive order did not actually do much and was mainly trying to lay the ground-work for future regulations on AI.
The AI Action plan also proposes government investment to advance AI capabilities, suggesting to "Prioritize investment into theoretical computational and experimental research to preserve America's leadership in discovering new and transformative paradigms that advance the capabilities of AI", and while the AI Action plan does acknowledge the importance of "interpretability, control, and robustness breakthroughs", it receives only about two paragraphs in a 28 page report (25 if you remove pages with fewer than 50 words).
However, as disappointing the current administration's stance on AI Safety may be, the previous administration was not an ideal model. According to this post NSF spending on AI safety was only 20 million dollars between 2023 and 2024, and this was ostensibly the main source of direct government support for AI safety. To put that number into perspective, the US Department of Defense spent an estimated 820.3 billion US dollars in FY 2023, and meaning the collective amount spent by represented only approximately 0.00244% of the US Department of Defense spending in FY 2023.
Many people seem to believe that governments will inevitably pivot to promoting an AI safety agenda at some point, but we shouldn't just stand around waiting for that to happen while lobbyists funded by big AI companies are actively trying to shape the government's AI agenda.
In contrast to lobbying, which we have good reason to suggest will fail, there are several good reasons to assume that a political party centered around AI safety could work. While it is tough to make firm conclusions about the effectiveness of certain political strategies, there are many third parties from the US and around the world to learn from.
Green Parties:
Green parties have obtained many electoral successes world-wide, with a recent example being the Green Party in Germany, which a few years ago entered into German government as part of a coalition. Even the US Green Party, which has note been the most electorally successful, has had a significant influence on climate policy in the US. In 2012, Jill Stein proposed a "Green New Deal", and I think its not a stretch to argue that this had some influence on the later Green New Deal proposed by AOC in 2019.
Additional Reading on Green parties' impacts on Politics:
2.) In the United States specifically, Green Party candidates running in competitive elections has been associated with a pro-environmental platform shift in the Democratic party.
Ross Perot/The Reform Party:
Ross Perot's campaigning has influenced American politics in many ways, especially by increasing the prominence of its core issues in political discussions. While by no means the first organization to talk about these issues, Ross Perot's campaigning played a key role in moving the issues of deficit reduction, protectionism, and anti-interventionism into the mainstream of American politics. While it is difficult to ascertain cause and effect, it should be noted that Donald Trump first ran for president as a Reform Party candidate, and many echoes of the Reform Party's ideology can be found in the thinking and policies of the current presidential administration.
Additional Reading on the Impact of Perot/The Reform Party:
In the book Three's a Crowd: The Dynamics of Third Parties, Ross Perot, and Republican Resurgence, Authors Ronald B. Rappaport and Walter J. Stone conclude Ross Perot's Reform Party has had a lasting effect on the two-party system, caused by strategic platform shifts in the Republican party and the migration of former Perot voters and activists into the Republican party. The authors of that book also wrote a shorter piece on the same subject.
RFK JR:
RFK JR has dramatically shifted US health policy. Under his leadership, DHHS has canceled hundreds of millions of dollars of funding for mRNA vaccines, stopped US funding of "Gavi, the vaccine alliance", an international vaccination organization. While the impact of RFK Jr's decisions are arguably very negative, what is inarguable is that his campaign demonstrates the potential power of savvy third-party candidacies.
Despite only averaging around 5% in the polls and not having a substantial electoral track record, RFK JR was able to cut a deal with Donald Trump which yielded him control of the DHHS in exchange for him dropping out and endorsing Trump. One could argue that Trump would have done similar things in a world where Kennedy didn't run, but while Donald Trump had a mixed effort on health before his second term, there is little to suggest that, without Kennedy, a hypothetical DHS would have made such dramatic changes from pre-existing policy.
While future major party political candidates might be less willing to offer a similar deal to future third parties as the one Kennedy got from Trump, this is a separate avenue a hypothetical AI safety party would have for achieving some of its goals.
Along with prior examples of third party success at influencing politics, the available evidence suggests the American public could support the agenda of an AI Safety party.
According to polling by the Pew Research Center (which I linked to at the beginning of this post) 58% of Americans polled were concerned that the government "will not got far enough in its regulation of AI".
According to a Reuters/Ipsos Poll which concluded last Monday, 47% polled agreed with the statement that "AI is bad for humanity", opposed to just 31% who agreed that "AI is good for humanity, and 58% agreed with the statement that "AI could risk the future of the human kind". Additionally, "71% were concerned that too many people will lose jobs" when asked about their "concerns about artificial intelligence",
According to a poll of 1,481 voters by the AI Policy Institute, 80% of those surveyed supported a policy approach to AI regulation which emphasized government oversight into the release of new AI models to ensure that the new models were safe, as opposed to 24% who supported an approach with basically no AI regulation, 46% who supported a ban on frontier AI development, and 50% who supported regulation based primarily on self-reporting by AI companies.
While it could be argued that Americans do not feel very confident in their beliefs on AI and these polling results do not really mean much for that reason, it is remarkable that, no matter what polling you examine, the general American public seems somewhat wary of AI. A 2025 report by Brookings concluded that "Overall, the U.S. and U.K. publics tend to be more concerned than optimistic about AI’s impacts, though many hold mixed or even inconsistent views".
Thus, an AI Safety political party would not have to start from square one in convincing the public of the risks of AI and to support government action on AI Safety, because many Americans already feel anxious about AI to varying degrees.
All in all, while the evidence is mixed and draws from differing political systems, there seems to be good reason to think that a party committed to AI safety and related causes could be effective at influencing American government and society to take AI safety more seriously.
If it is clear that governments must act to ensure that AI is developed safely, and many well-resourced, intelligent people, clearly think that future AI systems have a significant chance of killing every human currently alive, then I find it odd that there isn’t already a political party built around at least partially this issue.
While it may be unlikely that a candidate from an AI safety party wins a major national election anytime soon or cuts a deal with a major party candidate, that doesn’t mean such a party wouldn’t matter. Electoral success isn’t the only—or even primary—way political movements shape the future. A focused party could still have significant impact by any of the following:
Shifting the Overton window on AI regulation and safety.
Educating elite institutions and policymakers on the importance of AI safety.
Pressuring mainstream parties to adopt AI safety platforms .
Supporting and legitimizing pro-safety viewpoints.
Making AI safety a higher salience issue
These effects don’t require sweeping victories and aiming for them is far more promising than passively waiting for a political establishment—much of which still doesn’t grasp the existential risks posed by AI—to change on its own. And it’s likely more scalable than relying solely on individual advocates and underfunded safety orgs operating without any electoral leverage.
I concede a party which literally only talks about AI risk would not be effective, but the formation of a reasonable centrist party that has AI safety as one of its core goals and has leadership and members that care deeply about AI safety could greatly increase the likelihood that AI is safely developed.
But just because a pro AI safety political party could, in theory, lead to a positive change in governmental attitudes towards AI safety, does not mean it is the best cause area. How much time and money might it take to start a successful pro AI safety party?
Estimating the likelihood of electoral success is inherently difficult, as campaigns are shaped by a wide array of factors, including candidate quality, strategic execution, and message resonance. Nevertheless, by examining the costs and outcomes of previous political campaigns, several conclusions can be drawn. Many campaigns led by candidates with limited personal appeal or unremarkable platforms have nonetheless managed to achieve electoral visibility—and, in some cases, success—with relatively modest expenditures.
Not only does this exercise help us estimate costs, but I argue that this reinforces the notion that a political party organized around a timely and substantive issue such as AI safety, supported by a clear and distinctive message, could plausibly exert substantial influence on government, as arguably many candidates without a very compelling message, developed organizational structure, or personal charisma did quite well at running efficient campaigns.
Using data provided by Open Secrets and the FEC, I compiled the spending of all presidential campaigns tracked by Open Secrets between 2004 and 2024. There are many omissions of minor candidates, but I think this can at least serve as a starting point for analysis.
| Presidential Candidate | Election Year | Dollars Spent by campaign | Dollars Spent by Affiliated Outside Groups[1] | Total Spent by Campaign and Affiliated Outside Groups* | Votes cast for candidate | Percent of total vote cast for candidate | Approximate Total Dollars Spent/ vote gained | Approximate Total Dollars Spent/ Percentage of Vote Share Won |
| Chase Oliver | 2024 | 463,348$ | 0$/N/A | 463,348$ | 650,126 | 0.418792264 | 0.71$ | 1,106,391.02$ |
| Jill Stein | 2024 | 2,233,635$ | 0$/N/A | 2,233,635$ | 862,049 | 0.555306898 | 2.59$ | 4,022,343.33$ |
| Cornell West | 2024 | 1,275,968$ | 0$/N/A | 1,275,968$ | 82,644 | 0.053236862 | 15.44$ | 23,967,753.95$ |
| Kamala Harris | 2024 | 1,154,978,762 | 839,559,258 | 1,994,538,020 | 75,017,613 | 48.32416487 | 26.59$ | 41,274,133.25$ |
| Donald Trump | 2024 | 448,966,052 | 1,021,577,998 | 1,470,544,050 | 77,302,580 | 49.79607417 | 19.02$ | 29,531,325.00$ |
| Joe Biden | 2020 | 1,042,748,801 | 572,094,939 | 1,614,843,740 | 81,268,924 | 51.31151526 | 19.87$ | 31,471,371.13$ |
| Jo Jorgensen | 2020 | 2,757,703 | N/A | 2,757,703 | 1,865,724 | 1.177979488 | 1.48$ | 2,341,045.01$ |
| Donald Trump | 2020 | 778379130 | 312,254,786 | 1,090,633,916 | 74,216,154 | 46.85854237 | 14.70$ | 23,275,028.65$ |
| Hillary Clinton | 2016 | 563,433,611 | 205,144,296 | 768,577,907 | 65853514 | 48.18457808 | 11.67$ | 15,950,703.27$ |
| Donald Trump | 2016 | 325,515,461 | 97,105,012 | 422,620,473 | 62,984,828 | 46.08557961 | 6.71$ | 9,170,340.84$ |
| Gary Johnson | 2016 | 11,956,242 | 1314095 | 13,270,337 | 4,489,341 | 3.284820942 | 2.96$ | 4,039,896.61$ |
| Jill Stein | 2016 | 3,587,105 | N/A | 3,587,105 | 1,457,218 | 1.066236716 | 2.46$ | 3,364,267.00$ |
| Evan McMullin | 2016 | 1632885 | N/A | 1,632,885 | 731,991 | 0.535592945 | 2.23$ | 3,048,742.55$ |
| Virgil Goode | 2012 | 93,794 | N/A | 93,794 | 122,389 | 0.094812419 | 0.77$ | 989,258.59$ |
| Gary Johnson | 2012 | 2,507,763 | 551,386 | 3,059,149 | 1,275,971 | 0.988470347 | 2.40$ | 3,094,831.33$ |
| Barack Obama | 2012 | 721,397,677 | 75,145,374 | 796,543,051 | 65,915,795 | 51.06370658 | 12.08$ | 15,599,005.72$ |
| Mitt Romney | 2012 | 449,507,659 | 145,132,478 | 594,640,137 | 60,933,504 | 47.20402097 | 9.76$ | 12,597,234.83$ |
| Jill Stein | 2012 | 893,636 | N/A | 893,636 | 469,627 | 0.363811061 | 1.90$ | 2,456,318.94$ |
| Randall Terry | 2012 | 270,405 | N/A | 270,405 | 13,107 | 0.010153742 | 20.63$ | 26,631,067.59$ |
| Chuck Baldwin | 2008 | 208,229 | N/A | 208,229 | 199,750 | 0.15211651 | 1.04$ | 1,368,878.37$ |
| Bob Barr | 2008 | 1,393,262 | N/A | 1,393,262 | 523,715 | 0.398827024 | 2.66$ | 3,493,399.19$ |
| John McCain | 2008 | 333375676 | N/A | 333375676 | 59,948,323 | 45.65271424 | 5.56$ | 7,302,428.38$ |
| Cynthia McKinney | 2008 | 145,020 | N/A | 145,020 | 161,797 | 0.123213992 | 0.90$ | 1,176,976.72$ |
| Ralph Nader | 2008 | 3,996,305 | N/A | 3,996,305 | 739,034 | 0.562799864 | 5.41$ | 7,100,756.87$ |
| Barack Obama | 2008 | 729,519,581 | N/A | 729,519,581 | 69,498,516 | 52.92551538 | 10.50$ | 13,783,891.87$ |
| George W. Bush | 2004 | 345,259,170 | N/A | 345,259,170 | 62,040,610 | 50.7301484 | 5.57$ | 6,805,798.54$ |
| John Kerry | 2004 | 309708100 | N/A | 309708100 | 59,028,444 | 48.26712251 | 5.25$ | 6,416,543.68$ |
| Ralph Nader | 2004 | 4549143 | N/A | 4549143 | 465,650 | 0.380758564 | 9.77$ | 11,947,578.92$ |
| Michael Badnarik | 2004 | 1073945 | N/A | 1073945 | 397,265 | 0.324840655 | 2.70$ | 3,306,067.091$ |
| Michael Peroutka | 2004 | 708227 | N/A | 708227 | 143,630 | 0.11744519 | 4.93$ | 6,030,276.77$ |
| David Cobb | 2004 | 385,712 | N/A | 385,712 | 119,859 | 0.098007819 | 3.22$ | 3,935,522.75$ |
To begin the analysis, let's start with raw averages of how many dollars these campaigns spent per percentage of vote share.[2] The average is 10,535,457.35$, unadjusted for inflation, and $18,017,249.48, using a conservative estimate of inflation[3]. However, most winning campaigns spend far more than this per percent of vote share. Why?
One potential reason is simple: diminishing returns. As previously theorized, money in politics has its limits. After a certain amount of time and money spent campaigning, people will already know about your political party and form somewhat firm opinions about whether they will support you or not, however, many of the smaller party candidates are not known by a majority of Americans, and so see greater benefits from campaigning. The R^2 value for the total money spent by campaigns and the amount of dollars they spend per percentage of the vote share is 0.6875.
Another possible reason is strategic. A few thousand voters in swing states for a Democrat or Republican presidential candidate are far more influential than an additional million voters in a state like California. For this reason, large political parties concentrate their resources into the swing states, naturally reducing efficiency, as more resources are spent on a smaller number of voters.
Whatever the reason, the data (and my verdict) is clear: third parties (which, it is important to note, are almost by definition smaller than their major party counter parts) are more efficient at converting dollars into votes than larger ones. But how much more efficient? To start, the average amount of dollars spent per percentage of vote share won for all third parties is 5,969,545.93$ unadjusted for inflation, and $10,412,796.29 adjusted for inflation using a conservative estimate.
Even if we exclude the long-established Green and Libertarian parties—who benefit from name recognition and a longstanding party infrastructure—the efficiency gap remains. The dollars spent to percent vote share won ratio for the remaining third-party candidates is still only 9,885,294.69$ unadjusted for inflation ($17,243,113.82 adjusted for inflation) and this average is skewed by several outliers.
For instance, some of the candidates did not even get on the ballot in a majority of states. Cornell West, who had a dollars spent to percent vote share ratio of $23,275,028.65 (unadjusted for inflation) in 2024, he did not appear directly on the ballot in the majority of states, including very populous states such as California and New Jersey. While there is little available information on Randall Terry's campaign, it appears he also ran in the Democrat Primary, skewing his expenses while not directly contributing to his general election vote total (Randall Terry netted around 20,000 votes in the Oklahoma primary, more than his total general election vote total), and did not appear directly on the ballot in the majority of states.
While there are many factors which affect campaign finance, an AI third-party, assuming it is somewhat competently run and gets ballot access in a majority of states, should expect to spend somewhere between $1,158,877.12 (the lowest inflation adjusted 2025 dollars spent per vote share by any candidate, which was achieved by Chase Oliver's 2024 campaign), and $18,017,249.48 (the inflation adjusted average of dollars spent per vote share of all campaigns), with something on the lower end of that being far more likely, as third-party candidates typically spend far less per percent of vote share than larger parties.
So what does this imply for a hypothetical AI safety political party? Even in the worst-case scenario—the party is far less efficient than the average third-party campaign and about as efficient as the average campaign—it would cost about $90 million to secure 5% of the vote in 2028. That is a substantial amount, but far from impossible, and the potential payoff is enormous. Even a 10% chance of brokering policy concessions with a major-party nominee could be worth billions in redirected funding for AI safety.
And again, $90 million represents something close to the worst-case. A campaign operating at the same efficiency of Jill Stein’s 2024 campaign ($4,213,159.35 per percent of vote share) would need only about $20 million to reach 5%. If an AI safety political party achieved a similar efficiency as Chase Oliver's campaign–$12 million could buy not 5% but 10% of the national vote—a remarkably high return on investment.
Finally, an AI safety party wouldn’t need to run at full scale immediately. Early milestones, like gathering enough signatures for ballot access, could serve as low-cost tests of viability. If the project falters at that stage, it can be re-evaluated before major spending begins.
Beneath are my responses to potential objections to my argument. This is not a complete list, and I would like to hear your objections to my arguments in the comments.
This is probably the best objection to an AI safety political party, and after thinking about it I conclude I agree to an extent. An AI safety political party which mainly focuses on AI issues would probably not get anywhere.
However, I think a centrist party that spends 10-25% of its time talking about AI issues would be able to be just as effective (if not more effective) at promoting AI Safety, while also being far more likely to succeed electorally.
As already mentioned, winning isn't required for impact. A party could elevate AI risk in national debates, attract additional media attention, and legitimize pro-safety viewpoints in existing parties.
This may be true. But even if it is more likely than not its not a useful reason to do nothing. AI timelines are still uncertain and will be potentially bottlenecked by hardware, funding, energy, and/or algorithmic progress. Even small shifts to the amount of time it takes for AGI or ASI to be developed or the amount of or quality of safety measures in place when AGI or ASI will be developed could matter a lot.
Additionally, even if we can't convince governments to actively regulate AI development or increase funding for AI safety research, we can at least try to ensure governments aren’t actively making it worse (e.g., by subsidizing unsafe training runs or defunding AI safety).
I did mostly write this with the US in mind. After all, the most likely scenario in my view is that a company centered in the US develops the first AGI. However, the majority of people on this forum live in the Anglosphere, and almost all of these countries can influence the trajectory of AI. Ideas from Canada, Australia, the UK, or even New Zealand can easily spill over into U.S. discourse. More broadly, there’s also no reason only the U.S. can fund AI safety research. In an interconnected research ecosystem, investments in AI safety anywhere can help globally. By advancing AI safety research, creating public awareness, or just slowing down capabilities races via international coordination. Even if your country isn’t training frontier models, the actions of its government can still matter a lot. While the cost benefit is definitely more tricky, it is at least worth pondering.
Along with AI Safety, a Pro-AI-Safety political party would need to take stances on various other issues. While maintaining AI as a primary focus (and making that clear to the general public), it should also endorse a common sense platform generally based around evidence-based solutions, classical liberal ideals (such as respect for personal liberty and democracy), and political moderation (Not having a clear skew towards left or right, especially on polarizing culture war issues). This will help the party appeal to the largest number possible and to not explicitly paint AI -Safety as a "Left-wing" or "Right wing issue", as this will broaden the appeal of the party.
It should also emphasize that AI has the potential to lead to mass unemployment, and there should be a method of ensuring unemployed workers are able to survive (maybe an Automation dividend funded by taxes on AI companies and robotics company, maybe a UBI, maybe a federal job guarantee. Whichever is most popular should do) to tie the issue of AI safety to broader, more relatable anxieties about how AI will transform the economy.
In first-past-the-post systems, where there is only one round of voting and then the candidate with the most votes wins, there is a very prominent spoiler effect. In the US, where we have a first-past-the-post electoral system, I think the traditional pro-business, anti-regulation slant of the GOP will present obstacles to an ambitious AI-safety agenda from being supported by Republicans. Thus, I understand if some would be nervous that a potential AI third-party would strengthen the GOP by siphoning away voters from the Democratic party.
However, it remains to be seen what the ultimate approach of the Democratic party will be towards AI regulation, and it makes sense to try to pressure Democrats into adopting more comprehensive AI-safety political positions until the issue is more clearly politicized. Many Democrats in Congress do not understand the massive threat that artificial intelligence poses to humanity.
I will concede that in a hypothetical 2028 election between a Democrat who is somewhat friendly to AI safety policies and a Republican who is hostile to them, it might make sense for a hypothetical AI party to endorse the Democrat in swing states or even drop out of the race entirely and endorse the Democrat. However, this does not mean a third-party would be pointless, as it can still extract concessions from the larger party.
Thus, if an AI safety political party is constructed in a thoughtful manner, its leadership would be able to leverage spoiler effects to its advantage, turning a potential weakness into a strength without inadvertently promoting anti-safety politicians.
This topic deserves a post of its own, but to put it succinctly: there are no winners in an AI race. If a misaligned, uncontrollable superintelligence is created, it won’t matter whether it originates in the U.S. or China—the likely outcome could be human extinction or some other catastrophic loss.
While you may not share the values of the Chinese government, those values are at least recognizably human—shaped by the same kinds of goals, drives, and emotional states we all share. The same cannot be said for future AI systems, which may form entirely alien objectives and surpass human intelligence to the point where our survival becomes irrelevant to them.
In this light, a “safe” AI superintelligence aligned with Chinese interests would almost certainly be better for Americans—and for the world—than a misaligned AI superintelligence developed in the U.S. Moreover, the gap between a safe AI aligned with China and a safe AI aligned with the U.S. is tiny compared to the chasm between either of those and an unaligned AI.
This reasoning naturally leads to an argument for unilateral disarmament:
But there’s a better alternative: cooperation. Instead of racing or unilaterally slowing down, the U.S. and China could jointly commit to rules governing how AI should be developed, used, and safeguarded.
This cooperative approach is plausible for several reasons:
Firstly, the U.S. currently leads in both AI hardware and software. China has much to gain and little to lose from agreeing to a binding safety framework. Secondly, the Chinese government has consistently signaled that it cares about AI safety[4]—arguably more so than the U.S government ever has. Finally, International precedents exist. Despite immense challenges and distrust between superpower, the world has successfully negotiated enforceable treaties on biological weapons and nuclear arms.
Cooperation is therefore the best path forward: it reduces existential risk, preserves geopolitical stability, and avoids the trap of an all-out race. With credible U.S. commitment, there is little reason to think China would reject such an agreement, and history shows that global safety agreements are possible.
One major way in which people who care to influence governmental support for AI safety policies have sought to influence government has been through lobbying organizations and other forms of activism. However, there is reason to doubt they will be able to cause lasting change. First of all, there is significant evidence that lobbying has a status quo bias lobbying has a status quo bias. Lobbying is most effective when it relates to preventing changes, and when there are two groups of lobbyists on an issue, the lobbying to prevent change win out, all else being equal. In fact, according to a study by Dr. Amy McKay, "it takes 3.5 lobbyists working for a new proposal to counteract just one lobbyist working against it".
Even if this effect did not exist, however, it is very unlikely AI safety groups will be able to compete with anti-AI-safety lobbyists. Naturally, the rise of large, transnational organizations built to profit around AI has also lead to a powerful pro-AI lobbyist operation. This indicates we can't simply rely on current strategies of simply funding AI-safety advocacy organizations, as they will be eclipsed by better funded pro-AI-Business voices.
I feel like a political party would help raise the profile of activism and visa versa, and I feel like political parties are more effective than activism, but I don't really know enough about activism to say much here, so I am happy to simply facilitate discussion on this issue. Do you feel like an effective activist campaign would be more able to sway government decision-makers and the major parties than a political campaign? If so, I would like to hear why in the comments.
This could potentially be viable but I have a couple of concerns:
A third party would likely face less direct hostility from party elites/interest groups and would face less pressure to directly compete against major party candidates. I can imagine many being hostile to an external political force trying to "take over" a political party (NYT: Rationalists attempt takeover of Democratic Party!) if there was a serious attempt to try to gain influence in the major parties from within.
Furthermore, there is a greater risk of "politicizing" AI safety if successful, as compared with a third party run (although I personally do not think that is that much of a problem[5]). Right now I would say both parties do not really have established positions, but once one major political party starts arguing for substantial AI safety measures, tribalism might kick in and stifle bipartisan support for AI safety.
If you care about the risks AI pose to society, you should consider that electoral politics, while slow and flawed, may still be one of the highest-leverage tools available to bolster AI safety in the long run.
Lobbying, advocacy, and research are all important, but they operate in a political environment where governments still hold the decisive levers of power: regulation, funding, enforcement, and international negotiation.
An AI safety political party doesn’t need to win the presidency or even significant congressional representation to matter. Its possible it can reshape political agendas, shift the Overton window, and pressure larger parties to adopt their priorities. By doing so, it can achieve outsized influence compared to its vote share or funding base.
Even modest success could bring enormous benefits: greater public awareness of AI risks, expanded government research budgets, and the creation of institutions capable of managing transformative technologies safely. And while the costs of launching such a party are real, they pale in comparison to the potential benefit of ensuring significant US government commitment to AI safety.
Acknowledgements:
Written with assistance from Chat GPT and Claude. Also shout out to the Transformer Substack, which has been a good general source of news info on AI.
Open Secrets indicates that many candidates did not have outside groups who raised and spent money on their behalf. I suspect that some of these candidates did have some marginal support from outside groups, but as these weren't significant enough to be reported by Open Secrets, for the sake of my calculations when Open Secrets reports that campaigns affiliated outside group spent 0$, or, before 2012, if Open Secrets omits this entirely, I treat this as if they received none.
While the dollars per vote paints a similar story, I believe average dollar spent per percent vote share is a better metric as the data does not get skewed by voter turnout. A candidate who got 10,000,000 votes in an election where 100,000,000 people voted did better from an efficiency perspective than a candidate who got 19,000,000 votes in an election where 200,000,000 people voted, assuming their campaign spending was equal.
I did not want to have to manually adjust every value for inflation individually, so I just took the average and set the starting date on the "US Inflation Calculator" to January 2004, which is an overestimate as most candidates analyzed ran in latter election years. It probably wouldn't be to hard to create code to do this, but I would need to find the inflation rates from each year and the final result would still be a bit inaccurate. Similarly, for the inflation-adjusted estimated of Chase Oliver and Jill Stein's dollars spent per percent of vote share won, I just used January 2024 as the baseline.
Additional Sources about China's positive view of AI safety:
https://www.mfa.gov.cn/eng/xw/zyxw/202405/t20240530_11332823.html?utm_source=chatgpt.com
While I would rather have both major parties agree on AI Safety than one agree and one disagree, the latter is still better than neither major party having a clear position on the issue.