DISCLAIMER: Like the author I have skin in this game, in that I’m a professional policy advocate, and I currently work on public interest AI policy advocacy. So credit or discount as necessary.
Hi Charbel, overall agree with your thesis. In particular, I think we have enough evidence about risks and interventions that we should move the locus of activity towards advocating for those interventions and preventing those risks.
However, I want to underscore and elaborate on your Section 4-D: I don’t think research on audiences or messaging is a rival to the advocacy rebalancing you’re calling for, and I don’t want it to be tarnished as so. (Which I think you agree with.) I think it’s a needed capacity for the movement to mature and become more effective.
In other words: some of what could be filed under research is work that increases the effectiveness of marginal advocacy-hours. I agree 100% that it’s about repeating messages, but some messages to some people are ineffective and some are polarizing, and in general you want to be saying the right messages to the right people in the right contexts. (Some messages and messengers are net negative with some audiences.)
I come from corporate lobbying, and we had an enormous amount of audience data available to us, which meaningfully impacted when and how we engaged on policy issues, and I know from colleagues from US national political campaigns and issue campaigns that data can contribute meaningfully to success vs failure.
More generally (beyond persuasion), we need to understand where the public is on issues: (US) policymakers are money and vote maximizers (votes ~ your public incentive funnel), and we’re going to lose the money battle (your Commission lobbying figures are a proxy for this), so we need to make as much progress on the vote battle as possible.
In fact, to the extent I have any disagreement at all with your thesis and argument, it is likely on policymaker beliefs. I think that is an enormously difficult row to hoe if neither donors nor the public align with your asks to US policymakers, especially as issue saliency increases and the question climbs out of the staffers and advisors and into electoral politics.
Which is partly why I’m helping a new team start a high-quality biweekly tracker of Americans’ opinions on AI, including affect, agency, community and family impact, government performance, and AI futures. It’s called Americans On AI (home, substack post on first wave).
I plan to write up some background for the LessWrong community in a few weeks, once we have more data and more of our infrastructure online, but for this group the key pieces are probably that we are working with a high quality partner (NORC at the University of Chicago) for fielding on a probability-based panel, and we are including and already publishing data on AI futures relevant to the discussion of catastrophic risks.
From the most recent (first) wave:
77% say AI behaving in ways its developers didn’t intend is already happening or likely,
71% say the same about AI becoming smarter than humans at most tasks, and
71% about AI being used to build weapons of mass destruction.
It’s a bit like the METR-graph-for-policy but focused on the public, as well as various subgroups that we oversample to hit minimum CIs (sex, age bands, party ID, race/ethnicity, education, etc. - not all of which we publish on yet). We’re also experimenting with different ways of measuring salience / issue priority and will likely begin publishing on that later this year.
Returning to the messaging question, the persuasion evidence is still nascent. Seismic did good work, but Social Change Lab (SCL) did a high-powered RCT (N ~3,500) earlier this year which tested eleven harms on willingness to act, not just on concern.
SCL’s takeaways complicate the pure bundling story:
Salience doesn’t predict action. Job loss is the harm people raise first on their own, and one of the weakest at getting them to do anything. Bundling through what people already care about can route through the least useful entry point.
Extinction resonated least, but reading about AI-enabled warfare raised concern about extinction more than reading about extinction did. My guess is that you need to make the mechanisms of catastrophic risk legible to elicit concern.
It appears anger predicts action but anxiety doesn’t. Proximity and concern for other people also predict it, and again fear and fatalism don’t. These all go to core framing questions.
If anyone’s interested in the tracker in the meantime, happy to connect and chat about methodology, results, etc.
And there’s more work being done on the messaging front this year, so happy to take comments to pass on to our team and various others that are working on this.
DISCLAIMER: Like the author I have skin in this game, in that I’m a professional policy advocate, and I currently work on public interest AI policy advocacy. So credit or discount as necessary.
Hi Charbel, overall agree with your thesis. In particular, I think we have enough evidence about risks and interventions that we should move the locus of activity towards advocating for those interventions and preventing those risks.
However, I want to underscore and elaborate on your Section 4-D: I don’t think research on audiences or messaging is a rival to the advocacy rebalancing you’re calling for, and I don’t want it to be tarnished as so. (Which I think you agree with.) I think it’s a needed capacity for the movement to mature and become more effective.
In other words: some of what could be filed under research is work that increases the effectiveness of marginal advocacy-hours. I agree 100% that it’s about repeating messages, but some messages to some people are ineffective and some are polarizing, and in general you want to be saying the right messages to the right people in the right contexts. (Some messages and messengers are net negative with some audiences.)
I come from corporate lobbying, and we had an enormous amount of audience data available to us, which meaningfully impacted when and how we engaged on policy issues, and I know from colleagues from US national political campaigns and issue campaigns that data can contribute meaningfully to success vs failure.
More generally (beyond persuasion), we need to understand where the public is on issues: (US) policymakers are money and vote maximizers (votes ~ your public incentive funnel), and we’re going to lose the money battle (your Commission lobbying figures are a proxy for this), so we need to make as much progress on the vote battle as possible.
In fact, to the extent I have any disagreement at all with your thesis and argument, it is likely on policymaker beliefs. I think that is an enormously difficult row to hoe if neither donors nor the public align with your asks to US policymakers, especially as issue saliency increases and the question climbs out of the staffers and advisors and into electoral politics.
Which is partly why I’m helping a new team start a high-quality biweekly tracker of Americans’ opinions on AI, including affect, agency, community and family impact, government performance, and AI futures. It’s called Americans On AI (home, substack post on first wave).
I plan to write up some background for the LessWrong community in a few weeks, once we have more data and more of our infrastructure online, but for this group the key pieces are probably that we are working with a high quality partner (NORC at the University of Chicago) for fielding on a probability-based panel, and we are including and already publishing data on AI futures relevant to the discussion of catastrophic risks.
From the most recent (first) wave:
77% say AI behaving in ways its developers didn’t intend is already happening or likely,
71% say the same about AI becoming smarter than humans at most tasks, and
71% about AI being used to build weapons of mass destruction.
It’s a bit like the METR-graph-for-policy but focused on the public, as well as various subgroups that we oversample to hit minimum CIs (sex, age bands, party ID, race/ethnicity, education, etc. - not all of which we publish on yet). We’re also experimenting with different ways of measuring salience / issue priority and will likely begin publishing on that later this year.
Returning to the messaging question, the persuasion evidence is still nascent. Seismic did good work, but Social Change Lab (SCL) did a high-powered RCT (N ~3,500) earlier this year which tested eleven harms on willingness to act, not just on concern.
SCL’s takeaways complicate the pure bundling story:
Salience doesn’t predict action. Job loss is the harm people raise first on their own, and one of the weakest at getting them to do anything. Bundling through what people already care about can route through the least useful entry point.
Extinction resonated least, but reading about AI-enabled warfare raised concern about extinction more than reading about extinction did. My guess is that you need to make the mechanisms of catastrophic risk legible to elicit concern.
It appears anger predicts action but anxiety doesn’t. Proximity and concern for other people also predict it, and again fear and fatalism don’t. These all go to core framing questions.
If anyone’s interested in the tracker in the meantime, happy to connect and chat about methodology, results, etc.
And there’s more work being done on the messaging front this year, so happy to take comments to pass on to our team and various others that are working on this.