3rd PPE student at UCL and former President of the EA Society there. I was an ERA Fellow in 2023 & researched historical precedents for AI advocacy.
Charlie Harrison
Karma: 510
haha, yes, people have done this! This is called ‘vignette-adjustment’. You basically get people to read short stories and rate how happy they think the character is. There are a few potential issues with this method: (1) they aren’t included in long-term panel data; (2) people might interpret the character’s latent happiness differently based on their own happiness
All good. Easy to tie yourself in knots with this …
Hi Zachary, yeah, see the other comment I just wrote. I think stretching could plausibly magnify or attenuate the relationship, whilst shifting likely wouldn’t.
While I agree in principle, I think the evidence is that the happiness scale doesn’t compress at one end. There’s a bunch of evidence that people use happiness scales linearly. I refer to Michael Plant’s report (pp20-22 ish): https://wellbeing.hmc.ox.ac.uk/wp-content/uploads/2024/02/2401-WP-A-Happy-Probability-DOI.pdf
Thanks for this example, Geoffrey. Hm, that’s interesting! This has gotten a bit more complicated than I thought.
It seems:
Surprisingly, scale stretching could lead to attenuation or magnification depending on the underlying relationship (which is unobserved)
Let h be latent happiness; let LS be reported happiness.
Your example:
So yes, the gradient gets steeper.
Consider another function. (This is also decreasing in h)
i.e., the gradient gets flatter.
2. Scale shifting should always lead to attenuation (if the underlying relationship is negative and convex, as stated in the piece)
Your linear probability function doesn’t satisfy convexity. But, this seems more realistic, given the plots from Oswald/Kaiser look less than-linear, and probabilities are bounded (whilst happiness is not).
Again consider:
T=1: LS = h ⇒ P(h) =1/LS
T=2: LS = h-5 ⇔ h = LS+5 ⇒ P(h) = 1/(LS+5)
Overall, I think the fact that the relationship stays the same is some weak evidence against shifting – not stretching. FWIW, in the quality-of-life literature, shifting occurs but little stretching.
Sorry – this is unclear.
“If people are getting happier (and rescaling is occuring) the probability of these actions should become less linked to reported LS”
This means, specifically, a flatter gradient (i.e., ‘attenuation’) – smaller in absolute terms. In reality, I found a slightly increasing (absolute) gradient/steeper. I can change that sentence.
I could imagine thinking about “people don’t settle for half-good any more” as a kind of increased happiness
This feels similar to Geoffrey’s comment. It could be that it takes less unhappiness for people to take decisive life action now. But, this should mean a flatter gradient (same direction as rescaling)
And yeah, this points towards culture/social comparison/expectations being more important than absolute £.
Thanks a lot for this. I hadn’t actually come across these terms; that’s super useful. I’ll have to read both these articles when I get a chance, will report back.
Hi Geoffrey,
Thank you!
It’s possible that these 3 exit actions have gotten easier to do, over time. Intuitively, though, this would be pushing in the same direction as rescaling: e.g., if getting a divorce is easier, it takes less unhappiness to push me to do it. This would mean the relationship should (also) get flatter. So, still surprising, that the relationship is constant (or even getting stronger).
Hey Eugene, interesting stuff!
1) Long-term AI is very likely a complement; short-term, it may be a substitute”
I wonder why you think this?
2) “Good evidence suggests AI benefits the already skilled”
I feel like the evidence here is quite mixed: e.g., see this article from the economist: https://www.economist.com/finance-and-economics/2025/02/13/how-ai-will-divide-the-best-from-the-rest
If we treat digital minds like current animal livestock, the expected value of the future could be really bad.
Great analogy Alistair
Hi Aaron,
I’m sorry it’s taken me a little while to get back to you.
In hindsight, the way I worded this was overly strong. Cultural explanations are possible, yes.
I guess I see this evidence as a weak update on a fairly strong prior that the burden of knowledge (BOK) is increasing – given the range of other variables (e.g., age of innovation, levels of specialisation), and similar trends within patent data. For example, you couldn’t attribute increasing collaboration on patents to norms within academia.
I’d be interested to compare # researchers with # papers. The ratio of these two growth rates is key, for the returns to research parameter from Bloom et al. Do send me this, if you have remembered in the intervening time.
Charlie
Thanks for writing this Gideon.
I think the risks around securitisation are real/underappreicated, so I’m grateful you’ve written about them. As I’ve written about, I think the securitisation of the internet after 9/11 impeded proper privacy regulation in the US, and prompted Google towards an explicitly pro-profit business model. (Although, this was not a case of macrosecuritization failure).
Some smaller points:
Secondly, its clear that epistemic expert communities, which the AI Safety community could clearly be considered
This is argued for at greater length here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4641526
but ultimately the social construction of the issue as one of security is what is decisive, and this is done by the securitising speech act.
I feel like this point was not fully justified. It seems likely to me that whilst rhetoric around AGI could contribute to securitisation, other military/economic incentives could be as (or more) influential.
What do you think?
Hi, thanks for this! Any idea how this compares to total costs?
thanks, Charlie! Seems like a reasonable concern. I feel like a natural response is that hedonic wellbeing is only one factor within life satisfaction. Though, I had a quick look online, and one study suggests they’re pretty strongly correlated (r between 0.8 and 0.9) https://www.sciencedirect.com/science/article/abs/pii/S0167487018305087
Thanks, glad you enjoyed 👍
At least, from one of showing the plot. I’m more skeptical of the line, the further out it goes, especially to a region with only a few points.
Fair.
This data is the part I was nervous about. I don’t see a great indication of “leveling off” in the blue lines. Many have a higher slope than the red lines, and the slope=0 item seems like an anomaly.
To be clear – there are 2 version of levelling off.
Absolute levelling off: slopes indistinguishable from 0
Relative levelling off: slopes which decrease after the income threshold.
And for both 1) and 2), I am referring to the bottom percentiles. This is the unhappy minority which Kahneman and Killingsworth are referring to. So: the fact that slopes are indistinguishable after the income threshold for p=35, 50, 70 is consistent with the KK findings. The fact the slope increased for the 85th percentile is also consistent with the KK findings. Please look at Figure 1 if you want to double check.
I think there is stronger evidence for 2) than for 1).At percentiles p=5, 10, 15, 20, 25, 30 there was a significant decrease in the slope (2): see below. I agree that occurrences of 1) (i.e. insignificant slopes above £50k) may be because of a lack of data.
I also agree with you that the 0 slope is strange. I found this at the 10th and 30th percentiles. I think the problem might be that there wasnt many unhappy rich people in the sample.
Thank you! I haven’t used GitHub much before … Next time 🫡
Hey Ozzie!
1) Thank you!
2) < Or, the 15 percentile slopes are far higher than the other slopes > Agreed, this is probably the most robust finding. I feel pretty uncomfortable about translating this into policy or prescriptions about cash transfers, because this stuff was all correlative; and unearned income might affect happiness differently from earned income.
3) < 50k threshhold seems arbitrary > This is explained in the second footnote. It is worth >$100 USD now, I believe.
< I’d also flag that it seems weird to me to extend the red lines so far to the left, when there are so few data points at less than ~3k > Do you mean from an aesthetic point of view, or a statistical one? The KK (2022) paper uses income groups – and uses the midpoints for the regressions – which is why their lines don’t extend back to very low income.
< I’m skeptical of what you can really takeaway after the 50k pound marks. There seems to be a lot of randomness here >
I think this depends on what claim you are making. I think there is pretty strong evidence for relative levelling off – i.e. significant decrease in the slope for lower percentiles. You can look at the Table for t/p values.
[Editted: didn’t phrase this well]. Though, I agree with you that there is less evidence for absolute levelling off (i.e. 0 slopes above 50k). The fact that the slopes for lower percentiles weren’t significantly positive might be because of a lack of data. 0 slopes for p=10, 30 seems to corroborate this.
Although, if the problem was a generic lack of observations above £50k, then we wouldn’t see significant positive slopes for the higher percentiles. Perhaps, the specific problem was that there wasnt many unhappy rich people in the sample. I will add something to the summary about this.
I haven’t checked for outliers via influence plots or the like.
4) Yeahh, I feel like that would be cool, but would be better to do on the bigger dataset that Killingsworth used. The usefulness here was to use the same methods on different (worse) data.
Thanks Larks!
Hey Mo, thanks so much!
I don’t have a particularly strong view on this.
I guess:
First, there are differences in the metrics used – the life satisfaction (0-10) is more granular than the 4 category response questions.
Additionally, the plot from OWID, a lot of the data seems quite short-term – e.g., 10 years or so. Easterlin always emphasises that the paradox is across the whole economic cycle, but a country might experience continuous growth in the space of a decade.
My overall view – several happiness economists I’ve spoken to basically think the Easterlin Paradox is correct (at least, to be specific: self-reported national life satisfaction is flat in the long-run), so I defer to them.