Personally I find the best use for AI in terms of voting and crafting political opinions to be in its usage as a devil's advocate/pressure tester.
I'm well aware that there are a lot of individuals that are highly opinionated and feel strongly about certain issues, so much that they may be blind to its shortcomings and unwilling to even be open to considering the opposing view.
And so while it's important to respect everyone's vote in the spirit of democracy, I find it important to foster an environment that reduces hostility towards those with different viewpoints (hostility that has only gotten worse in the age of social media and echo chambers).
Using AI as a devil's advocate can be a step in that direction where we're not using it to tell us what to vote based on our current values, but instead inform us of the shortcomings of those values. This can hopefully create a system where we are more considerate of one another.
Yes, I agree! This could be part of how we help train agents to be not just vote translators but partners in helping people to learn the issues and what they really think about them. At the same time, as I flagged to John above, there will be questions about people's real appetite for engaging in all this.
I agree the appetite for engagement is a real issue. To be frank, the vast majority of voters are unlikely to have it.
Though I wonder if some level of engagement might be a base requirement for allowing people to "proxy" their preferences to AI in the first place. IOW, if you're not willing to do the work to engage with the agent so it can accurately reflect your preferences, you might not be allowed to use an agent. Which means you'd have to vote in person, or proxy your vote to another human (which, I'd suggest, is essentially what we do when we elect human representatives now).
Which brings up another future question: Will people be able to allow agents built by other people to vote for them? All this would really take is for some people to instruct their agents to follow another agent's votes rather than making its own decisions. And that opens a huge can of worms as it would allow some people to build a power base outside conventional political channels by just convincing other people to follow their agent's votes (which is clearly a lot less work than training their own agents).
Sorry, if I keep bringing up potential issues - the systems engineer is me is always looking for the breaking points. But the earlier we identify them, the more likely it is we'll build a system that's hard to hack (in the sense of breaking the intent, not necessarily electronic incursion).
In any case, how to handle proxies (sanctioned or not) seems like another key topic that will need to be addressed at some point.
Future problem to consider: What happens when someone changes their personal values?
The old adage comes to mind, that if you're not a liberal when you're 20 you have no heart, and if you're not a conservative when you're 40 you have no brain. Believe that adage or not, people's values change over time, sometimes radically.
There are some fairly simple possibilities to address this, like doing periodic scheduled re-interviews, or re-tuning the agent after major life events. But the truth is, for at least some people, value change happens continually. Especially on hot topics, people are exposed to information on those topics fairly relentlessly, and it shifts their view over time - often without them consciously realizing it.
Obviously, getting agents to represent current values well must come first. But once that stake is in the ground, we'll have to deal with how to keep agent representatives up to date in the constantly shifting information landscape we all live in.
This is complicated by the inherent limitations of current LLMs (they are, after all, language models, not consciousness models).
For example, LLMs have recency bias (which may help), but they lose the script as context grows (which may hurt). They have limited and varied ability to faithfully follow instructions, even when they have all the context they should need. They're non-deterministic, and occasionally go off in la-la-land that seems completely out of whack with their context and instructions (as may have happened with one or two of the student's agents in your experiment).
But here's another thought that may apply even with those known limitations: What if anyone could run a simulation of how a vote would turn out at any time before the actual vote, to see what the results are? And then if they care enough about the issue (and they feel the vote will be close enough), they could potentially take the time to tune their agent's parameters before the final vote to try and shift the results?
What kind of policy improvement might we get from allowing citizens to offer information on possible amendments that would change their vote? And what kind of polling could we get from running those simulations regularly in the lead up to a major vote?
This is super interesting! I haven't thought about this yet, but I agree it's important. In general it's going to be hard to get people to spend the time to train their agents, unless they're doing really tangible work for them that is of clear value. So that implies it might be even harder to get them to regularly update their views. We have to somehow make it something interesting and valuable, rather than like forced software updates that we all view as unpleasant chores.
Among other things this makes me think we need to experiment on training agents among the general public. My students are obviously inclined to find this stuff useful and interesting. We need to know if the average person has the appetite to engage in this kind of training---and then whether they're willing to retrain on a regular basis.
Separately I'm sure people will propose that the agents automatically update based on other things they're learning about you over time. This could work if your governance agent is related to your overall "personal superintelligence" that you're using, but it also raises potentially uncomfortable questions about privacy.
I thought of input from the "personal superintelligence" as well. But even putting the privacy issues aside (since those are a function of the personal AI, not the political aspect), there's a very real challenge that many people don't express their preferences honestly a good portion of the time. So, expecting a personal AI to be able to filter out social conformity activities from actual preferences feels like a minefield.
Sounds like an interesting topic for a sci-fi story though... ;)
This is incredibly interesting. We're running a month-long experiment with 500 people and 500 agents to see how agents affect community governance and decision-making.
It'll happen in the month of June, an hour north of SF, could be interesting to chat about collaborations! We'll share everything we learn openly.
Since last year when I started engaging with genuinely substantive discussion with Chat 5.2 about ideas that matter, I have been astonished at the level of depth, caring, sensitivity, and wisdom (there is no other word as accurate) that AI seems to exhibit. And it seems obvious to me that AI is better at judgements along with outpacing nearly every human skill set there is ALREADY. In short, we need to find SOME way to use AI in our governance. I know how controversial this idea will be, but I for one vote that we develop a consensus about to train our new and eventual masters to dispense justice and help govern. This article showed me how naive I was about the actual mechanics of that.
I've been thinking about this more, because many of the challenges of LLMs (for example, non-determinism) apply equally to all uses of them.
tl;dr: Maybe to improve the reliability and alignment of LLM agent responses we should take multiple "samples" of the responses, and use statistics to select the most appropriate response?
The "standard" way to try and get LLM alignment is to provide additional context, rules, and data to shape the LLMs responses. But this approach doesn't actually work reliably - and I'm not sure it ever will.
And in fact, the more context we give an LLM, the more likely it is to ignore some of the context. So, there's actually a point of diminishing returns with trying to create guardrails - a point at which trying to get the agent to more closely hew to individual preferences may make it reflect our preferences worse, instead of better.
And because any errant responses add to the context for future responses, even a single bad response in a long string of actions can take the LLM off on a tangent that deviates further and further from the user's intent.
So, how do we address deviation without just adding more rules and context?
The probabilistic nature of LLMs guarantees that responses will vary. And the nature of probabilities means that sometimes the responses will vary *tremendously*. Even when the probability of extreme responses is extremely low, if you scale it large enough, they'll happen in some cases - that's just math. (And the probability of hallucinations and other extreme responses with LLMs is actually relatively high).
But we have other methods of handling variability (even in the extreme). And the probabilistic nature of LLMs suggests one that may work rather well to improve consistency and reliability - taking multiple samples of output, and using statistics to select one.
For example, if you ask the same LLM to score the same input using the same rubric 5 times, you'll get 5 different scores. But if you were to take the median of all those scores, you'd get a much more consistent scoring mechanism that's much more resistant to extremes. Note that this doesn't prevent extreme scores across different inputs; we're talking about reproducibility with the same inputs.
The bottom line: I wonder if this principle might be harnessed in some way to make your political agents better reflect the user's preferences.
For example:
1. Allow the agent to come up with 5 responses.
2. Score how well they reflect the user's preferences (possibly with a separate LLM call).
3. Pick the one with the best or median score.
Notes:
- Obviously, this would multiply inference costs. But what are the costs of a significant portion of AI representatives getting their votes wrong?
- There are reasons the best or the median might work better. This is something that could be tested by having the users rate both of them and figuring out which best represents the user's preferences long-term.
Totally non-sequitur: How will you prevent agents from submitting prompt injection attacks into the AI legislature's deliberations? How will you prevent users from jailbreaking their AI reps? I ask because these are a couple of other inherent weaknesses with LLMs, which may require new forms of AI (e.g. symbolic AI/world models) to reliably defend against.
Incredibly fascinating!
Personally I find the best use for AI in terms of voting and crafting political opinions to be in its usage as a devil's advocate/pressure tester.
I'm well aware that there are a lot of individuals that are highly opinionated and feel strongly about certain issues, so much that they may be blind to its shortcomings and unwilling to even be open to considering the opposing view.
And so while it's important to respect everyone's vote in the spirit of democracy, I find it important to foster an environment that reduces hostility towards those with different viewpoints (hostility that has only gotten worse in the age of social media and echo chambers).
Using AI as a devil's advocate can be a step in that direction where we're not using it to tell us what to vote based on our current values, but instead inform us of the shortcomings of those values. This can hopefully create a system where we are more considerate of one another.
Yes, I agree! This could be part of how we help train agents to be not just vote translators but partners in helping people to learn the issues and what they really think about them. At the same time, as I flagged to John above, there will be questions about people's real appetite for engaging in all this.
I agree the appetite for engagement is a real issue. To be frank, the vast majority of voters are unlikely to have it.
Though I wonder if some level of engagement might be a base requirement for allowing people to "proxy" their preferences to AI in the first place. IOW, if you're not willing to do the work to engage with the agent so it can accurately reflect your preferences, you might not be allowed to use an agent. Which means you'd have to vote in person, or proxy your vote to another human (which, I'd suggest, is essentially what we do when we elect human representatives now).
Which brings up another future question: Will people be able to allow agents built by other people to vote for them? All this would really take is for some people to instruct their agents to follow another agent's votes rather than making its own decisions. And that opens a huge can of worms as it would allow some people to build a power base outside conventional political channels by just convincing other people to follow their agent's votes (which is clearly a lot less work than training their own agents).
Sorry, if I keep bringing up potential issues - the systems engineer is me is always looking for the breaking points. But the earlier we identify them, the more likely it is we'll build a system that's hard to hack (in the sense of breaking the intent, not necessarily electronic incursion).
In any case, how to handle proxies (sanctioned or not) seems like another key topic that will need to be addressed at some point.
Fascinating stuff.
Future problem to consider: What happens when someone changes their personal values?
The old adage comes to mind, that if you're not a liberal when you're 20 you have no heart, and if you're not a conservative when you're 40 you have no brain. Believe that adage or not, people's values change over time, sometimes radically.
There are some fairly simple possibilities to address this, like doing periodic scheduled re-interviews, or re-tuning the agent after major life events. But the truth is, for at least some people, value change happens continually. Especially on hot topics, people are exposed to information on those topics fairly relentlessly, and it shifts their view over time - often without them consciously realizing it.
Obviously, getting agents to represent current values well must come first. But once that stake is in the ground, we'll have to deal with how to keep agent representatives up to date in the constantly shifting information landscape we all live in.
This is complicated by the inherent limitations of current LLMs (they are, after all, language models, not consciousness models).
For example, LLMs have recency bias (which may help), but they lose the script as context grows (which may hurt). They have limited and varied ability to faithfully follow instructions, even when they have all the context they should need. They're non-deterministic, and occasionally go off in la-la-land that seems completely out of whack with their context and instructions (as may have happened with one or two of the student's agents in your experiment).
But here's another thought that may apply even with those known limitations: What if anyone could run a simulation of how a vote would turn out at any time before the actual vote, to see what the results are? And then if they care enough about the issue (and they feel the vote will be close enough), they could potentially take the time to tune their agent's parameters before the final vote to try and shift the results?
What kind of policy improvement might we get from allowing citizens to offer information on possible amendments that would change their vote? And what kind of polling could we get from running those simulations regularly in the lead up to a major vote?
This is super interesting! I haven't thought about this yet, but I agree it's important. In general it's going to be hard to get people to spend the time to train their agents, unless they're doing really tangible work for them that is of clear value. So that implies it might be even harder to get them to regularly update their views. We have to somehow make it something interesting and valuable, rather than like forced software updates that we all view as unpleasant chores.
Among other things this makes me think we need to experiment on training agents among the general public. My students are obviously inclined to find this stuff useful and interesting. We need to know if the average person has the appetite to engage in this kind of training---and then whether they're willing to retrain on a regular basis.
Separately I'm sure people will propose that the agents automatically update based on other things they're learning about you over time. This could work if your governance agent is related to your overall "personal superintelligence" that you're using, but it also raises potentially uncomfortable questions about privacy.
I thought of input from the "personal superintelligence" as well. But even putting the privacy issues aside (since those are a function of the personal AI, not the political aspect), there's a very real challenge that many people don't express their preferences honestly a good portion of the time. So, expecting a personal AI to be able to filter out social conformity activities from actual preferences feels like a minefield.
Sounds like an interesting topic for a sci-fi story though... ;)
This is incredibly interesting. We're running a month-long experiment with 500 people and 500 agents to see how agents affect community governance and decision-making.
It'll happen in the month of June, an hour north of SF, could be interesting to chat about collaborations! We'll share everything we learn openly.
https://edgeesmeralda2026.substack.com/p/the-agent-village-experiment-at-edge
Since last year when I started engaging with genuinely substantive discussion with Chat 5.2 about ideas that matter, I have been astonished at the level of depth, caring, sensitivity, and wisdom (there is no other word as accurate) that AI seems to exhibit. And it seems obvious to me that AI is better at judgements along with outpacing nearly every human skill set there is ALREADY. In short, we need to find SOME way to use AI in our governance. I know how controversial this idea will be, but I for one vote that we develop a consensus about to train our new and eventual masters to dispense justice and help govern. This article showed me how naive I was about the actual mechanics of that.
I've been thinking about this more, because many of the challenges of LLMs (for example, non-determinism) apply equally to all uses of them.
tl;dr: Maybe to improve the reliability and alignment of LLM agent responses we should take multiple "samples" of the responses, and use statistics to select the most appropriate response?
The "standard" way to try and get LLM alignment is to provide additional context, rules, and data to shape the LLMs responses. But this approach doesn't actually work reliably - and I'm not sure it ever will.
And in fact, the more context we give an LLM, the more likely it is to ignore some of the context. So, there's actually a point of diminishing returns with trying to create guardrails - a point at which trying to get the agent to more closely hew to individual preferences may make it reflect our preferences worse, instead of better.
And because any errant responses add to the context for future responses, even a single bad response in a long string of actions can take the LLM off on a tangent that deviates further and further from the user's intent.
So, how do we address deviation without just adding more rules and context?
The probabilistic nature of LLMs guarantees that responses will vary. And the nature of probabilities means that sometimes the responses will vary *tremendously*. Even when the probability of extreme responses is extremely low, if you scale it large enough, they'll happen in some cases - that's just math. (And the probability of hallucinations and other extreme responses with LLMs is actually relatively high).
But we have other methods of handling variability (even in the extreme). And the probabilistic nature of LLMs suggests one that may work rather well to improve consistency and reliability - taking multiple samples of output, and using statistics to select one.
For example, if you ask the same LLM to score the same input using the same rubric 5 times, you'll get 5 different scores. But if you were to take the median of all those scores, you'd get a much more consistent scoring mechanism that's much more resistant to extremes. Note that this doesn't prevent extreme scores across different inputs; we're talking about reproducibility with the same inputs.
The bottom line: I wonder if this principle might be harnessed in some way to make your political agents better reflect the user's preferences.
For example:
1. Allow the agent to come up with 5 responses.
2. Score how well they reflect the user's preferences (possibly with a separate LLM call).
3. Pick the one with the best or median score.
Notes:
- Obviously, this would multiply inference costs. But what are the costs of a significant portion of AI representatives getting their votes wrong?
- There are reasons the best or the median might work better. This is something that could be tested by having the users rate both of them and figuring out which best represents the user's preferences long-term.
Totally non-sequitur: How will you prevent agents from submitting prompt injection attacks into the AI legislature's deliberations? How will you prevent users from jailbreaking their AI reps? I ask because these are a couple of other inherent weaknesses with LLMs, which may require new forms of AI (e.g. symbolic AI/world models) to reliably defend against.