AI Campaign Infrastructure (Voter Vectors, Voter Simulations, Persuasion Reward Models)
This doc is only a rough outline and needs more work, but IMO something in this space will become very important over the next 5-10 years, even if we're not sure exactly what.
Background facts
I've listed some relevant facts below for context, to help frame what good campaign infrastructure ideas might look like. This is intended as a review of what I think is (both) true and important to keep in mind, not as a "case" to persuade someone who strongly disagreed. However, I think there's pretty solid evidence for the facts below, and am happy to make arguments/provide more detail as needed.
Political facts
It's common for swing voters to be on the opposite end of a spectrum from almost everyone who's really into politics, on either left or right; not just different but actively opposite. This includes staffers, campaign volunteers, journalists, election data nerds, activists and Twitter hobbyists ("politicos").
Most politicos are highly educated, and at least went to college. Most swing voters are working class.
Politicos are much younger than average, swing voters are older.
Politicos are Very Online, swing voters still mainly get their info from TV news.
Politicos usually have a strong ideology that they try to make consistent. This can be "Democratic" or "Republican", but it also includes smaller groups like woke progressives, libertarians, true religious fundamentalists, etc. Swing voters are often inconsistent and idiosyncratic, and they mostly don't punish you for "flip-flopping".
Politicos write a lot, either professionally or as a hobby. Swing voters largely don't write, don't have strong writing skills, and when they do it's for a small audience of family and personal friends.
Politicos use Twitter and Reddit a lot. Swing voters use Facebook, to the extent they have social media at all.
Swing voters vastly outnumber politicos, but politicos tend to be much more visible as a group, for all the above reasons.
Politicos live in a handful of coastal cities (New York, SF, DC), swing voters are concentrated in suburbs in swing states.
Many more politicos lean left than right, but the right-leaning ones are much more strongly conservative than swing voters. Politicos tend to be ideological extremists and to bite bullets philosophically, eg. leftists want no restrictions on abortions, while rightists want a total ban, zero exceptions.
Campaign ads which politicos especially like are less likely to persuade swing voters, and vice-versa.
Most politicos want some kind of far-reaching change. Most swing voters are small-c conservative, have a steady job and a mortgage, and are suspicious of anything that sounds radical or that would disrupt their lives.
Americans, especially swing voters, have heterogeneous beliefs. Eg. about 20% of Republicans are pro-choice, while lots of black conservatives still vote Democratic. People who agree with a candidate on every issue are rare.
True marketing, in the sense of making an honest, thoughtful attempt to persuade someone different to change their mind (classically called "rhetoric"), is fairly rare in modern politics. Most marketing is what I call "meta-marketing", and is done by members of a group ("Xers") to persuade an audience of other Xers that they, the marketers, are skilled at persuading Xers of Xer beliefs. An easy example is eg. arguing that "abortion is a social justice issue" - someone persuaded by the phrase "social justice" is almost certainly pro-choice already.
For this reason, many stories that politicos tell about persuasion (eg. "group X hates/will hate Trump for reason Y") are imaginary, fabricated, projections or rationalizations.
Most political ideas, eg. "critical race theory", start with a small group and then spread outward into wider circles.
People will pick up ideas from sources they trust, whether that's a personal community, organization, celebrity, candidate or journalist they follow.
However, they aren't just gullible lemmings who believe everything they hear. Agreeing with someone's deeply held views and instincts builds trust, while trying to fight them lowers and eventually destroys it.
Americans are heterogeneous in who and what they trust. Swing voters generally don't trust either party, or they wouldn't be swing voters. Even people who are solidly D or R will trust different voices within their own coalition.
If people don't trust you, they'll often start believing the opposite of what you say. Eg. Joe Manchin now has high approval among Republicans, while Adam Kinzinger has high approval among Democrats. Reverse psychology can work in real life.
Issues vary in salience (what people are paying attention to). Salience in a campaign depends on the randomness of the news cycle, what both candidates run on, and also on what various pressure groups try to push. People will switch votes when the issues they agree with a candidate on become more salient.
Taking popular positions helps you win, and in a two-party system, a successful party must be a "big tent" to bring in 51% of voters. This is obvious but important and neglected, so I'll include it for the record.
Technical facts
The state-of-the-art in deep learning is advancing extremely fast, especially around natural language. It's now actively hard to invent easily measurable linguistic tasks that average people can solve and models can't. Many predictions that a particular task would be hard going forward have fallen embarrassingly quickly.
Large models can have much better memories than any person, remembering eg. details of the bios of millions of people from reading the whole Internet.
Things are moving faster than the (already fast) pace of publications would suggest. Many models are developed in secret, then not released or discussed publicly for a year or more. At deep learning events, there are constant, nervous hints of "oh, you haven't seen anything yet".
The barriers to entry for state-of-the-art deep learning are fairly high, it requires substantial cash and somewhat niche technical expertise. Many individuals and academics are being out-competed by a small number of groups with big resources, established infrastructure and deep moats (Google, DeepMind, OpenAI, Anthropic). There is ongoing consolidation in the industry, analogous to the consolidation of car or airplane manufacturers in the early 20th century.
These big groups are, at least for the moment, fairly bad at and/or uninterested in productizing their models (eg. OpenAI could easily sell DALL-E profitably, but they choose not to). Many of their leaders are former academics who see "research" and prestige as the ultimate goal, not practicality, or are people with a similar mindset. The huge profitability of FAANG core business models (ads, phones, online retail, B2B software sales) makes productizing new things less important.
- happens partly because big tech companies are, in many ways, constrained. They are large, hierarchical organizations with internal rules and systems, they are under political pressure from various angles, and they will take flak for doing something unpopular or controversial. (Either among the public, politicos, or their own employees, who are restive and sometimes revolt.)
One remaining gap in deep learning systems is that, unlike most software, they are somewhat stochastic and it's much easier to get them to solve something 99% of the time than 100%. This makes them tough to deploy in situations where there isn't human oversight, or where a single failure would be catastrophic (full self-driving cars are both, explaining their slow progress).
Possible systems to build
It's worth thinking very carefully about if we choose to build these tools - the most important question - and if so, exactly how. Once something is built, and deployed, and successful, it's hard to put the genie back in the bottle. Developing potentially misaligned AI is dangerous. Building AI capabilities is dangerous, although everything here is a fairly straightforward extension of existing stuff. It's implausible that important new tech can be kept exclusive indefinitely; sooner or later, bad guys will use it too. We might extend the exclusivity window, or lead, if we're skilled and lucky. If we're unskilled or unlucky, the window could be short.
However, on the flip side, it's also true that:
the existing situation (both politically and technologically) is fairly grim, and being conservative will likely cause a bad outcome by default;
the levers we have now (both political-campaign-wise and AI-coordination-wise) are deeply insufficient for the scope of the problem, we are metaphorically trying to lift a 500-ton rock with three guys and some hand tools;
given the rapid pace of deep learning tech, anything straightforward to build will likely be built eventually by someone else, no matter what we do. Making it even faster than it would have been is still bad, of course. But we shouldn't go too too far there; if eg. DeepMind and OpenAI hadn't existed, AI might be a year or two behind, but wouldn't be fundamentally, dramatically different. At least attempting to be secure, and not making technology public immediately, should help some.
there is, potentially, something to be said for "eating up overhang". Hardware and software will keep advancing. If in 2050, it's technologically feasible for any coder with a clever idea to build a dangerous system in an hour, and then someone discovers how, the development is likely to be explosive, uncontrolled and very bad. If we can do it safely, then slowly getting to a stable equilibrium point with a new tech might plausibly be better than leaving it lying around like an unexploded land mine. (But there are many possible factors here.)
It may be worth deliberately trying to make systems to check ourselves - eg. an "ethics committee" that is actually an ethics committee, not for PR purposes, but who really will tell us if we're doing something dumb.
We should also, of course, keep in mind our priorities, our capabilities, and what would get us the best payoffs in a tight time window. These are only some ideas that came to me in the last few months, I'm sure there are more out there.
Voter Vectors
Autoencoders and some other models map a complex set of data (images, songs, videos, etc) to an "embedding" or "latent space". Eg., a face GAN might map a face photo to a 256-dimensional vector, which is one point in "face space". These vectors are often semantically meaningful and obey simple arithmetic, eg. 0.5 <man's face> + 0.5 <woman's face> will produce a blend of plausible children (this is literal arithmetic, not a metaphor). <My face> + <age> will produce an older version of me, <my face> + <gender> will produce a more masculine or feminine me, and so on.
There are endless "political compass" memes and attempts to group voters by age, gender, race, ideology, etc., but this is mostly done ad-hoc and not from first principles. It might be possible to learn "voter vectors" by eg.:
gathering a big, heterogeneous set of discussions on policy-relevant topics, with author names or IDs attached;
training a standard LLM (large language model) to predict the next word in a discussion, in the usual way;
training a second LLM which takes, as input, a sampling of the person's other writings, maps the entire thing to a "voter vector", and is used as the input to the first LLM, with the word prediction loss function back-propagating through both.
A person's writings, from any source, could then give us a "voter vector" on how they think politically. Unlike standard data sets, the vector dimensions would be optimized to deliver maximum predictive power for what a voter would say, about any topic. Note that a model could learn many more features than a person's explicit opinions; someone's word choice, life history, hobbies etc. are all probably predictive.
My guess is this would be fairly easy to test as a proof-of-concept, by downloading the Reddit comment dataset and training a model on that. Reddit is a bad model of swing voters (skews hard towards the young and politicos), but is good for a tech demo because it's big, diverse, heterogeneous and can trivially be downloaded. Open-source medium-size language models are available off-the-shelf from EleutherAI (GPT-J-6B, GPT-NeoX-20B), as a starting point. Bigger models would be better, of course, but they're expensive (in money and time), and it would be good to have proof-of-concept before going all-out. The best large-scale public data source here is probably Facebook, since it's used heavily by swing voters, but it can't trivially be downloaded and we'd need some strategy for sampling it.
Voter Simulations
With a language model and a voter vector, it should be easy to simulate voters with any given set of views, then see how they would respond to a given question or line of reasoning. This allows for much faster iteration than focus groups, with the usual caveats about Goodharting the model. For further accuracy, we could calibrate ourselves by doing real-world advertising tests, then seeing how responses in the field matched up to our simulations, and iterating from there. With vector arithmetic, we could also simulate voters who have a particular set of beliefs and things they support, and then move to supporting new ideas or candidates while otherwise making as few changes as possible. This could also tell us who is relatively more persuadable and who is less.
Reward Scores for Text Generation
A standard generative LLM just tries to predict the next word; that's somewhat limiting, because what it writes isn't optimized for any particular human purpose. However, it's simple to couple an LLM with a reward model of what text is "good" or "bad", and use reinforcement learning to optimize the reward; OpenAI did this with "Learning Text Summarization from Human Feedback" and InstructGPT. In OpenAI's case, the reward was just "quality" as judged by human raters, but one could use the same tech to optimize any text for any purpose that one can make a numerical score for. (By reinforcement learning standards, this is fairly simple, just a few rounds of PPO.)
One could, for example, train a reward model on whether voters with particular beliefs click on an ad, whether they share a piece of writing widely, whether they respond in a positive or negative tone (sentiment analysis), or whether they rate something as persuasive to them in surveys. This reward model can then be used to tune a language model to write things that are ranked highly. One could write things on the same topic that are optimized to sound polite, angry, silly, formal, informal, from a poor audience, from a rich audience, etc., as long as one has some data on what writing is good or bad by that criterion.
Writing Assistants
As discussed above, for many reasons, politicos are an overwhelming force in the media compared to swing voters; the voices of swing voters are not heard. (As TvTropes says, "most writers are writers", and writing is a rare occupation for swing voters.) This makes swing voters distrust most media. In addition, because swing voters are heterogeneous and inconsistent, it would be hard for any one source to be trusted by them as a group, although some like Joe Rogan get closer than others.
This means there could be great prospects for an AI writing assistance tool, like Wordtune, built for swing voters on board with our campaign. For a casual layman who doesn't write professionally, writing something with AI help ("centaur AI") could be much faster and higher-quality than what either human or AI could do. This would give this group more reach, and increase their persuasiveness, compared to the professional media and political establishments.
Crucially, I think we should encourage assisted writers to be fully honest and write their true beliefs, even if they disagree with us on a few issues. There are strong pressures for everyone in a campaign, party, or "scene" to adopt the official "line" or "messaging". However, this limits their circle of trust; a writer who takes the "party line", even if not paid, will (rightfully) be seen as an extension of the campaign, just as (eg.) many conservatives saw public health officials as extensions of the Democrats during COVID. Since swing voters are heterogeneous, diverse, and inconsistent, there should be big advantages to assisting many writers who are trusted by different audiences. And the only way to have them not be distrusted as hacks is for them to actually not be hacks, to speak up when they disagree.
Note that in recent years, big campaigns have been flooded with way more volunteers than they can usefully handle. Writing would potentially be good work for volunteers to do. However, as above, volunteers are typically politicos and the wrong type of person to do this work; we should explicitly correct for this, by encouraging those with different views or backgrounds.
Empirical Voter Studies
The size and reach of social media datasets, and big deep learning models, lets us study things "in the wild" that were difficult to observe before. For example, one way to see what someone is thinking is by paying attention to which phrases, terms or concepts they use, which then tells you a lot about which authors they read and trust. One could plausibly construct a latent space of "meme vectors", embodying particular ideas like "wokeness" or "YIMBYism", or important facts or news stories, and then study how they spread over time. This would be a real-world study of memetics that would let us figure out how to promote our own ideas.
Another thing to empirically study would be how persuasion has worked in practice; given a voter vector, and given that that vector has changed over time (in the wild) to favor a particular candidate or position, what statistical correlates that might explain the change? Hard data on this would be super valuable, since (as discussed above) there are many myths about persuasion and even smart people often fall for them. People have tried to quantify this before, of course, but they've been limited to crude ad-hoc measures like certain demographics (eg. truck drivers) or who agree with one particular stance (eg. those who dislike immigration) being more likely to change their minds. A richer dataset could be very powerful.