In this issue:
AI voting advice - Chatbots provide welcome election guidance to voters who know what to ask
What the Founding Fathers can tell us about AI - An essay finds parallels between the Founders’ challenge and our own
Dispatches from Mitch
AI voting advice
Chatbots provide welcome election guidance to voters who know what to ask
America’s 250th birthday has a lot of people pondering, with some anxiety, the durability of our democratic system. Our polarized, hyperfragmented social media landscape seems to discourage people from seeking out impartial information, which might inconveniently violate the tribal narratives.
Indeed, the very idea of “impartiality” has been under attack for some time. To a news junkie like me, the word has come to feel quaint, bordering on naive. I despair at this. Such cynicism feels like the natural enemy of civic mindedness.
This might be why I was strangely heartened by Jennifer Medina’s piece in the New York Times today about earnest voters seeking advice from equally earnest chatbots.
Claude, or at least the persona we call Claude — the character Anthropic’s chatbot plays when it’s working as intended — is nothing if not earnest. But as Medina explains, Claude, like other leading chatbots, is “trained to avoid answering political questions that could expose biases.”
So when a user profiled in the piece wanted to inform herself, she had to phrase her prompt not in terms of election or voting advice, but as a series of research requests about choices that would be “strategic” for groups with her values.
A different user, working with ChatGPT, routed around a refusal to tell him which candidate was “the most libertarian” by asking instead for analysis of the candidates’ voting history.
User after user is described as pleased with the assistance. One said, “I felt so refreshed. That’s the most informed voting I have ever done. It felt like some political expert that knew all of the research and we just sat down over coffee and chatted and they took notes.”
Viewed through one lens, this article is about how to circumvent chatbot guardrails for benign and even patriotic reasons. But Medina doesn’t lean into this, instead working to balance the glowing testimonials with concerns about how AI election advice could go wrong. According to experts she consults: The bots could be overly persuasive, as they are superhumanly good at coming up with facts and explaining them clearly. The facts they give could be wrong, informed by flawed internet sources — but those same experts concede that this could be true for users doing independent research as well.
Medina reports that campaigns are looking for ways to influence how chatbots talk about their candidates, such as by playing to bots’ known love of bullet points.
Interestingly, the possibility of chatbots hallucinating facts doesn’t come up in this piece, which may be an indication that this error mode is in decline, as many insiders are claiming and as I am finding in my own use.
Given all of the horrible ways AI could undermine democracy — deepfakes, social media bots, concentration of power, killing all the voters (and everyone else) in pursuit of goals nobody programmed in — I don’t think “Claude giving bad election advice” even makes my top ten list of election concerns.
I’d even go so far as to say that I think current AIs are probably a boon for democracy when used in this role. The default alternative is often voting based on appearance and vibes. I actually wonder if it would be good for the chatbots to proactively offer to help voters do their election homework. I doubt that will ever happen — it would expose the companies to too many accusations of partiality for too little profit — but anyone trying to envision a future where democracy coexists with powerful AI should probably start thinking about how to augment the electorate’s ability to make informed decisions.
What the Founding Fathers can tell us about AI
An essay finds parallels between the Founders’ challenge and our own
The 4th of July weekend has brought an unusual calm to the AI news cycle. I was thus in a patriotically contemplative mood when I encountered a new essay from Peter Wildeford, of the AI Policy Network.
Called “The Alignment Problem of 1776”, it draws parallels between the problems the Founding Fathers were trying to solve and the problems our generation must solve with respect to AI. It’s a great essay. A+.
To oversimplify, the Founders were in a pickle, and they knew it. To become independent and stay that way would require the creation of a government strong enough to oppress them all over again.
[H]ow do you safely live alongside a powerful force you can never fully trust, fully predict, or fully control? 250 years ago, the question was the King … and then the new American government itself. Today, the same question applies to superintelligence: can it be similarly controlled, and can it be similarly made accountable to the people?
In Wildeford’s telling, the Republic has held together as long as it has thanks in no small part to the individual choices of early leaders like General George Washington, who established a precedent of military subordination to civilian command, and as President established a precedent of stepping down after two terms.
Wildeford finds no such virtue among those racing to superintelligence. In a section called “Consent of the governed,” the leaders of AI companies are called out for dragging us all down the path to AI superintelligence against our wishes, corrupted by the pursuit of an unstoppable power that will not be theirs to control.
Wildeford assigns additional credit for the longevity of the American experiment to the system of checks and balances established in the Constitution — the result, in part, of careful research by James Madison, who had “systematically studied every ancient and modern confederacy he could find sources on, cataloguing their failure modes the way [an AI] safety team catalogues jailbreaks.”
But checks and balances, Wildeford explains, only work if the bodies pitted against each other are roughly matched in power. Human institutions can sometimes keep other human institutions in check. They can sometimes keep today’s AIs in check. They cannot expect to hold superintelligence in check. Therefore, he says:
[Madison’s] machinery must be applied while it still can be. Checks and balances bind humans — humans can be voted out, subpoenaed, outcompeted, shamed, fired. For at least a few more years, the Ring is held entirely by humans — executives, boards, engineers, the officials who could govern them — and everything about them remains within reach of the oldest tools of the republic. The question is not whether the Constitution can restrain a superintelligence. It is whether the American people will assert their authority over the people building one, during the window in which those people can still be obliged to answer.
In a footnote, Wildeford, one of those “superforecasters” my colleague Robert talked about yesterday, says he is “50% sure we will have such ‘superintelligent’ AI systems before the end of 2036 absent some major war or regulation disrupting current technological progress.”
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.