Re: bets—I am highly uncertain about AI timelines myself. Amongst the EA/AI community, I think my timelines would be considered “long”—I expect major advancements in robotics will be needed for AI to displace more than about 20% of the current jobs in developed countries, and I don’t expect that could happen within the next 10 years.
Tax Geek
Karma: 209
The data in this area is pretty shaky. On the graph you have shared specifically:
Although the graph goes back to 1820, there is actually very limited data before WWI, as most countries didn’t have income taxes back then (which is the main source of data). In fact, when you click on the individual countries, most of them do not go back before 1900. The US starts in 1913, China starts in 1979, India in 1923, Indonesia in 1984 and Nigeria in 1985. Japan’s series doesn’t start until 2008!! Clicking through the countries, it looks like countries with considerably less than half of the current world’s population had any data before 1900, and I am not sure what extrapolations they had to make to get the “World” figures they use for teh graph.
I would also be very sceptical of income data for developing countries, which tend to have large “informal economies”.
China is an interesting case. There is no doubt that China has experienced enormous economic growth since its economy opened up in the late 1970s. But the graph for China shows the income share of the top 10% growing from 28% of China’s income to 41.7%! Caveats about data reliability still apply, but that is quite a shocking increase in inequality.
More generally—there is some evidence that median wages have started to decouple from general productivity growth, with the labour share of income declining over the past ~20-30 years. Economists argue about the data, the causes of this, and whether it is a temporary or lasting change. But some think it is because “superstar” firms with low labour shares have taken a greater market share thanks to globalisation and digitsation. See e.g. Autor et al (2020), and this 2018 OECD working paper.
In any case, I don’t think we are going to resolve this disagreement by exchanging data because: (1) data is limited; (2) there are so many different ways to interpret the data and experts genuinely disagree on this; and (3) we can also argue over the extent to which the past predicts the future. I think there is only so much we can learn from the past as I am concerned AI can be different in important ways—i.e. I think it is very plausible that the distributional impacts of automating cognitive labour will be quite different to the impacts of automating physical labour or routine, administrative tasks, but I won’t be able to prove this with data.
The main reason I replied wasn’t to get into an argument over data as that can go on for days with no resolution. Rather, I replied because I saw this post already had a fair number of “disagree-votes” and only one comment explaining why, so I thought I should at least explain my alternative view.
This argument seems to be based on the view that “jobs” are just tasks that need to be done in the economy, and that it’s better if we have AI/robots that can do those tasks more efficiently. But “jobs” have a dual role—for most people, they are also the main pathway for upward social mobility.
Financial capital follows a power law distribution. Human capital is normally distributed. Jobs are therefore the main mechanism we have in society for those who were not lucky enough to be born into financial capital to nevertheless earn a living and get a share of the vast wealth that “the economy” produces. If that mechanism is obliterated—what other ways are there to ensure people not born into financial capital to get wealth? Will they have to rely on altruism alone?
I have heard people worry that the new wealth produced by AI will all be concentrated in the hands of, say, five people, and the rest of us will be left to starve.
I think this is a strawman of actual wealth concentration concerns. Most reasonably informed people I know who hold these concerns aren’t expecting AI’s wealth to be concentrated in the hands of 5 people. But AI’s wealth could easily be concentrated in the hands of the people who hold the majority of AI-related capital—say, the top 10% globally − 800 million people (and their descendants) - with few pathways for those not born into capital to get a share of that wealth. Even the top 1% globally is 80 million people.
If these 80 million people hold the vast majority of wealth and power, it is foolish to think that the rest of society can “just keep doing what they’re currently doing”, completely unaffected. For example, if the top 80 million people got incredibly wealthy to the point that they each got private jets (according to Google, there are currently only about 25k private jets), that alone could raise air pollution or carbon emissions in a way that definitely impacts the other 99% of the world.
Good interview. On why there is such vast disagreement about AI’s potential economic impacts—I do think Ajeya’s hypothesis about different base rates has merit, and I also agree that many economists are simply not entertaining the premise where AI actually can beat the top humans at all cognitive tasks.
But I think one grossly underrated reason why economists and AI futurists talk past each other on this point is that most futurists don’t understand what “economic growth” and “GDP” actually measure. Presumably, when futurists talk about “economic growth”, they use it as a shorthand to refer to real output rather than nominal output, since the latter can be increased by simply printing money. However, we don’t have ways to directly measure real economic growth — rather, all our measures are based on market transactions at nominal prices, and then we apply some “deflator” to account for inflation. Some adjustments are also attempted to account for changes in quality and consumer tastes (i.e. the consumption basket today is vastly different from the consumption basket 100 years ago), but this is a very messy and imperfect process.One of the reasons why GDP is so bad at capturing technological changes is that productivity improvements are, all else equal, deflationary. Consider AI-generated video and art—there is ~infinitely more of this today than there was 10 years ago, but because most of that is generated at such low costs, the amount that that video and art contribute to GDP is trivial.
At the same time, AI will not eliminate positional and status goods. So even if AI eliminates all physical scarcity, positional goods are likely to then take up more and more of the overall “economy”. This suggests that the correlation between GDP and the well-being metrics that people actually care about will weaken, if not decouple entirely.
I suspect much of the gap between mainstream economists and AI futurists would shrink (though not entirely vanish) if futurists just stopped referring to “economic growth” and instead referred to how they expect physical output or energy usage to change over time. Both the Agricultural and Industrial Revolutions can be explained just as well, if not better, with energy usage as with GDP. I think the same will be true of TAI. Most economists would readily concede that AI could increase energy usage by 1000x or more if AI managed to solve nuclear fusion (for example) - they’d just be (rightly) sceptical of any suggestion that that translates to a 1000x increase in measured GDP.
Thanks for the post. I generally agree with your arguments but thought I should respond as someone currently doing research on a non-alignment problem. While I want a global pause, I have no idea what I personally can do to help achieve that. Whereas I at least have some idea of actions I can take that might help reduce the “massive increase in inequality/power concentration” problem.
“Solve philosophy” is not the same thing as “implement the correct philosophy”, and we need the AI to bridge that gap. There is a near-consensus among moral philosophers that factory farming is wrong, yet it persists.
This is a great point and I just wanted to call it out. I do think research is most likely to make a difference when it is produced with some thought about implementation—i.e. who the relevant audience is, how to get it to them, whether the actions you are recommending they take are actually within their power, etc.
Thank you! Appreciate the feedback :)
Yes, I think “AI as normal technology” is probably a misnomer—or at least very liable to being misinterpreted. Perhaps this later post by the authors is helpful—they clarify they don’t mean “mundane or predictable” when they say “normal”.
But I’m not sure a world where human CEOs defer a lot of decisions, including high-level strategy, to AI requires something that is approximately AGI. Couldn’t we also see this happen in a world with very narrow but intelligent “Tool AI” systems? In other words, CEOs could be deferring a lot of decisions “to AI”, but to many different AI systems, each of which has relatively narrow competencies. This might depend on your view of how narrow or general a skill “high-level strategy” is.
From the Asterisk interview you linked, it doesn’t sound like Arvind is expecting AI to remain like narrow and tool-like forever. Just that he expects it will take longer to reach AGI than people expect, and only after AIs are used extensively in the real world. He admits he would significantly change his evaluation if we saw a fairly general-purpose personal assistant work out of the box in 2025-26.
I sort of see where both criticisms are coming from. The lowest-common-denominator, community-related posts get the highest engagement (including from people like the OP) because they require little context. The high context technical stuff is much harder to parse, and necessarily has a smaller audience, so gets less engagement (perhaps with the exception for AI Safety, which is currently experiencing a “boom”).
There will naturally be fewer technical posts in the areas I’m interested in and, like Michael_PJ, I have no desire to read long technical posts in areas I’m not interested in, so I end up engaging disproportionately with community-related posts.
Fiddling with the forum filters helps—I personally have downweighted posts tagged “Community” and “Building effective altruism”—but I suspect few people do this.
Hi Dushan. I cover this at a high level under “But impacts will be uneven” heading. I agree with you that countries in the supply chain will benefit and others less so.
Thank you! Yes, totally fair point. I am not trained in development economics so was very uncertain about this post, and expected there to be large differences between countries that I wouldn’t pick up. It’s disappointing to hear that the development econ mainstream has not been engaging with this topic.
I had in mind the lower-income countries (mostly in Africa) when writing most of this. Your point about how, without TAI, these countries might be able to develop export industries and climb the development ladder is an interesting one. I had thought of that briefly, but was unsure how likely that was to happen, given we haven’t seen any African country do it yet (to my knowledge). But perhaps it’s just something that takes time and can only really happen after the middle-income countries become rich.
AI agrees with you on cars and yachts, but says the majority of TVs, gaming rigs, and bikes consumed in HICs are made in LMICs.
Fair enough. I think most of these are made in Asia and I do expect Asia (particularly China) to fare better than most other LMICs or developing countries.
I should note that “transfers” is not limited to unemployment benefits. For OECD governments, the biggest class of transfer by far is currently public pensions.
There are all sorts of good reasons why the elderly should be happy with lower public pensions (elderly poverty rates tend to be lower than for children or working-age adults, life expectancy has increased far more than retirement ages have). But that still doesn’t happen for political economy reasons. Perhaps that will change with TAI—the elderly tend to own more capital so they should see massive returns in general. Maybe they’ll be happy with <10x pension increases even as wages increase 10x. I just wouldn’t take that for granted.
Agree 100% that governments would need to tap into capital gains somehow, or capital more broadly. I also like that Capital dividend fund idea—thanks for sharing.
Thanks for the post! I share much of the concerns you raise, particularly your conclusion that benefits of AI will not be distributed equitably through natural market mechanisms.
There will still exist a sizable gap between the development of these systems and their diffusion into the broader economy, but this gap will be on the order of years, not decades.
I am curious about why you think this. And by “the broader economy” are you talking about the global economy or only the US? I don’t have any firm views on speed of diffusion but I find decades plausible, at least when it comes to the global economy. Especially if diffusion involves widespread deployment of robotics.
Thanks. I’m a bit sceptical of that 10x estimate and will have a closer look at that paper.
However, even assuming wages for non-automatable roles goes up ~10x before full automation, that won’t help governments if their costs rise more than 10x. In developed countries, government costs mostly consist of social protection transfers and wages themselves. In the case where wages rise 10x, transfers could rise more than 10x if (1) transfers are linked to wages (which they often are); and/or (2) the share of people receiving transfers rises (because unemployment rises).
It is possible that transfers could be de-linked from wages somewhat, but political economy can make that difficult and, to the extent that people’s welfare depends on the parts of the economy that are not rapidly growing (e.g. healthcare, housing, childcare), that could have negative welfare impacts.
So I’m not saying governments are doomed—as I point out, TAI should be creating value and the challenge is ultimately one of distribution. But governments still have to worry about revenue, because it’s not the size of GDP that matters so much as the composition of government income and spending.
Thanks for the feedback and you raise some interesting points.
the demand for manual labor will increase dramatically, because the people making a lot of money from AI will want bigger houses, more roads, and more physical goods in general
I’m not sure about this. I think we have different intuitions on how broad the class of “people making a lot of money from AI” vs how broad the class of knowledge workers losing out from TAI. My sense is that the former will be smaller than the latter, just based on data I’ve seen showing how little most people in developed countries have saved for retirement in the form of financial assets like shares (i.e. I haven’t researched this, and it likely depends on the country).
I guess this will also depend on the specific capabilities of TAI and how much labour it will displace.
I appreciate this challenge and it would be good to see some modelling on it (I might have a quick go at some stage).
Some of the production of goods will be done in the HICs (perhaps increasing remittances), but I think a lot will be done in LMICs, which would a boon for those countries.
Interesting point. Some of the goods you mention people getting rich from AI will want are bigger houses and roads. These are not globally traded so, to the extent TAI increases demand for construction work, I would expect this demand to increase in HICs rather than LMICs.
To the extent that people getting rich from AI want to increase their consumption, I expect much of that will be channelled through higher demand for in-person services—e.g. personal chef, gardener, masseusse, and (as you point out) construction.
Some increased consumption might also be channeled through physical goods, but physical goods are already incredibly cheap relative to incomes in HICs. The main bottlenecks to consumption are likely to be physical space (your house needs to be big enough for your toys) and leisure time. I expect TAI could remove the leisure time bottleneck for existing capitalholders, which could boost demand for leisure complements. But there is still only so much time in the day, and many higher-tech leisure goods like TVs, gaming rigs, bikes, cars, yachts don’t rely much on LMIC labour anyway. (Some things like clothing and tents may be more exceptions.)
I do agree with you that some LMICs could see a boon—but I don’t expect this effect to be widespread across LMICs in general.
Also, in this scenario, government revenue would be large because they would be taxing the high wages of manual workers, and taxing massive corporate/capital gains.
I disagree with this because most governments tax labour more heavily than capital. So if all the income currently earned by labour just shifts to capital, all else equal tax revenues will decline. (At the same time government spending will likely have to increase to pay for increased unemployment. I go into this more here.)
If governments are able to shift their tax bases towards capital we might see tax revenues remain stable.[1] But there are major challenges to taxing capital: capital is mobile and more easily hidden, there are serious valuation issues when you try to tax unrealised gains[2], and it is distortionary because it is very hard to properly account for inflation. Plus there are a bunch of political challenges. All these challenges are why tax bases rely more heavily on labour to begin with.
- ^
I find it highly unlikely tax revenues would increase in the near term.
- ^
Most countries tax realised gains instead, but this leads to undertaxation because taxpayers can choose when to realise those gains, which is usually only when they need to fund personal consumption. There are also tax avoidance opportunities and distortionary lock-in effects that crop up when taxpayers can choose when to realise their gains—which is part of why capital is almost always taxed at flat rates rather than on a progressive schedule.
- ^
Thanks for this post. As someone who has only recently started exploring the field of AI safety, much of this resonates with my initial impressions. I would be interested to hear the counterpoints from those who have Disagree-voted on this post.
Do you think the “very particular worldview” you describe is found equally among those working on technical AI safety and AI governance/policy? My impression is that policy inherently requires thinking through concrete pathways of how AGI would lead to actual harm as well as greater engagement with people outside of AI safety.
I have also noticed a split between the “superintelligence will kill us all” worldview (which you seem to be describing) and “regardless of whether superintelligence kills us all, AGI/TAI will be very disruptive and we need to manage those risks” (which seemed to be more along the lines of the Will MacAskill post you linked to—especially as he talks about directing people to causes other than technical safety or safety governance). Both of these worldviews seem prominent in EA. I haven’t gotten the impression that the superintelligence worldview is stronger, but perhaps I just haven’t gotten deep enough into AI safety circles yet.
On the “AI as normal technology” perspective—I don’t think it involves a strong belief that AI won’t change the world much. The authors restate their thesis in a later post:
There is a long causal chain between AI capability increases and societal impact. Benefits and risks are realized when AI is deployed, not when it is developed. This gives us (individuals, organizations, institutions, policymakers) many points of leverage for shaping those impacts. So we don’t have to fret as much about the speed of capability development; our efforts should focus more on the deployment stage both from the perspective of realizing AI’s benefits and responding to risks. All this is not just true of today’s AI, but even in the face of hypothetical developments such as self-improvement in AI capabilities. Many of the limits to the power of AI systems are (and should be) external to those systems, so that they cannot be overcome simply by having AI go off and improve its own technical design.
The idea of focusing more on the deployment stage seems pretty consistent with Will MacAskill’s latest forum post about making the transition to a post-AGI society go well. There are other aspects of the “AI as normal technology” worldview that I expect will conflict more with Forethought’s, but I’m not sure that conflict would necessarily be frustrating and unproductive—as you say, it might depend on the person’s characteristics like openness and willingness to update, etc.
I’m considering writing a post on how I think governments are likely to respond to the tax, spending, and inequality pressures that transformative AI will bring.
Others have already pointed out that if TAI displaces a lot of workers, this will reduce tax revenue (as most countries’ tax bases rely heavily on labour income) while increasing government spending (to pay for benefits to support people out of work). I’ve also read some articles suggesting we’ll need a high degree of global coordination in order to make sure AI’s benefits are widely distributed.
I agree global coordination on tax would be the first best solution, but I also think it is highly unlikely to happen. However:
I think there are things individual countries can (and hopefully will) do to mitigate national inequality;
I’m not sure TAI will worsen global inequality (I think it probably will, but not by as much as I initially thought); and
I don’t think governments are going to go broke everywhere (I don’t this is actually possible). We will likely see significant economic disruptions and maybe a few defaults, but the fiscal situation may not be as bad as some people seem to think.
I’m not sure exactly where I’ll end up with it, but my hope is to outline a few realistic pathways that will help people (including myself) decide where best to focus their efforts.
Please let me know if you’d be willing to read drafts or act as a sounding board — I would very much appreciate the help.
Thank you very much for this post. I agree that privacy seems to be a massive blindspot for EA. I think it is a common blindspot for well-meaning people/groups, because they find it hard to imagine how adversarial actors might actually act.
As a counterbalance to the “convenience” benefits offered by tech cloud platforms, I want to point out that there are also people who are concerned about data privacy who may hesitate to participate in EA activities/workshops/bootcamps or apply to EA grants/orgs if it would require handing over copious amounts of personal information or even having a LinkedIn profile (given LinkedIn’s terrible reputation with privacy). I am one such person.
I know there are always trade-offs between convenience and security, and that LinkedIn is widely used in recruitment spaces. But I would like to see more considered discussion of these trade-offs within EA. Currently the default just seems to rely on a high degree of trust, but even well-meaning people might can inadvertently leak other people’s information (e.g. feeding example profiles into an AI to help them write their own profile). My initial impression is that there is much more that could be done to protect privacy without significant convenience cost.
For example, I attended EAGx Virtual conferences in 2023 and 2024. The Google sheets with over 1000 attendees’ personal data (which is maintained by CEA) was shared with all attendees, and I still have access to these in 2025. I don’t see much benefit for these documents to exist at all, let alone remain publicly accessible years after the conference has ended. This likely also breaches data privacy laws in many jurisdictions, including the GDPR, which has a principle of data minimisation and requiring data to be kept no longer than is necessary.
At the very least, the retention policies for widely shared documents should be seriously considered.
You may be interested in the Windfall Trust’s work—they’ve published a paper mapping the tax risks from labour-displacing AI and have also floated a proposal to raise the global minimum corporate tax rate.
Prominent economists have also discussed expanding the use of consumption taxes to plug the gap.
I am sceptical of automation/token taxes in general as they are narrow by design. It is hard to predict in advance where in the value chain economic rents will accrue (e.g. what if AI becomes commoditised and the rents accrue to those who hold proprietary data instead?). An automation/token tax implicitly assumes that the value will be captured at the computing layer.