Alex Imas
@aleximas.bsky.social
Economics + Applied AI, Prof at University of Chicago Booth School of Business. Formerly: Carnegie Mellon, UCSD, Northwestern.
Website: www.aleximas.com
Website: www.aleximas.com
Follow up surveys using both free response and direct questions revealed social desirability bias as the main reason given for the own-other gap. Specifically, the gap was attributed to a reluctance to report one's own AI use rather than an inflation of others' use.
April 29, 2025 at 9:36 PM
Follow up surveys using both free response and direct questions revealed social desirability bias as the main reason given for the own-other gap. Specifically, the gap was attributed to a reluctance to report one's own AI use rather than an inflation of others' use.
On extent of use, majority of students reported using AI 0-1 days a week, while the majority of their friends used AI 4-5 days a week.
April 29, 2025 at 9:36 PM
On extent of use, majority of students reported using AI 0-1 days a week, while the majority of their friends used AI 4-5 days a week.
Results were striking, while ~60% of people reported using AI at all themselves, they reported that ~90% of their friends used AI.
The most common response for own use was "not at all", while the most common response for others' use was "a moderate amount" followed by "a lot".
The most common response for own use was "not at all", while the most common response for others' use was "a moderate amount" followed by "a lot".
April 29, 2025 at 9:36 PM
Results were striking, while ~60% of people reported using AI at all themselves, they reported that ~90% of their friends used AI.
The most common response for own use was "not at all", while the most common response for others' use was "a moderate amount" followed by "a lot".
The most common response for own use was "not at all", while the most common response for others' use was "a moderate amount" followed by "a lot".
🚨New paper (link in reply)🚨
Are we underestimating AI use in self-report surveys?
Yes, by as much as 30 percentage pts. We find 60% self-reported vs. truth closer to ~90% (!!)
Why? Social desirability bias, people embarrassed/worried to admit AI use, so they underreport 🧵
Are we underestimating AI use in self-report surveys?
Yes, by as much as 30 percentage pts. We find 60% self-reported vs. truth closer to ~90% (!!)
Why? Social desirability bias, people embarrassed/worried to admit AI use, so they underreport 🧵
April 29, 2025 at 9:36 PM
🚨New paper (link in reply)🚨
Are we underestimating AI use in self-report surveys?
Yes, by as much as 30 percentage pts. We find 60% self-reported vs. truth closer to ~90% (!!)
Why? Social desirability bias, people embarrassed/worried to admit AI use, so they underreport 🧵
Are we underestimating AI use in self-report surveys?
Yes, by as much as 30 percentage pts. We find 60% self-reported vs. truth closer to ~90% (!!)
Why? Social desirability bias, people embarrassed/worried to admit AI use, so they underreport 🧵
The algorithm knows me better than I know myself
April 14, 2025 at 9:50 PM
The algorithm knows me better than I know myself
Promised my kid that I had saved all my amazing transformer toys at his grandparent’s house. Opened the box only to find this:
Wtf is a formulator force.
It also broke the second I touched it.
Tears were shed.
Wtf is a formulator force.
It also broke the second I touched it.
Tears were shed.
April 13, 2025 at 8:01 PM
Promised my kid that I had saved all my amazing transformer toys at his grandparent’s house. Opened the box only to find this:
Wtf is a formulator force.
It also broke the second I touched it.
Tears were shed.
Wtf is a formulator force.
It also broke the second I touched it.
Tears were shed.
The issue, as always, is getting people to click and engage. We have more knowledge freely accessible than any time in human history. Full MIT courses on literally everything, thu click of a button. Are people more educated now than before? (no)
April 13, 2025 at 1:26 PM
The issue, as always, is getting people to click and engage. We have more knowledge freely accessible than any time in human history. Full MIT courses on literally everything, thu click of a button. Are people more educated now than before? (no)
Would also add: Reasoners excel in STEM because of large amount of training data with very similar content. Reasoners still continue to fail in STEM when faced with new content (see recent results on 2025 Math Olympiad questions).
arxiv.org/abs/2503.21934
arxiv.org/abs/2503.21934
April 11, 2025 at 2:06 PM
Would also add: Reasoners excel in STEM because of large amount of training data with very similar content. Reasoners still continue to fail in STEM when faced with new content (see recent results on 2025 Math Olympiad questions).
arxiv.org/abs/2503.21934
arxiv.org/abs/2503.21934
From @andrewpiper.bsky.social’s excellent thread:
April 10, 2025 at 11:22 PM
From @andrewpiper.bsky.social’s excellent thread:
The finance paper writes itself: "Can (Fake) News Move Markets: Yes"
April 7, 2025 at 5:28 PM
The finance paper writes itself: "Can (Fake) News Move Markets: Yes"
Please submit your AI/ML + Econ work to the Machine Learning in Economics Summer Conference at Booth!
Deadline: May 9
Please RT.
www.chicagobooth.edu/research/cen...
Deadline: May 9
Please RT.
www.chicagobooth.edu/research/cen...
April 7, 2025 at 1:40 PM
Please submit your AI/ML + Econ work to the Machine Learning in Economics Summer Conference at Booth!
Deadline: May 9
Please RT.
www.chicagobooth.edu/research/cen...
Deadline: May 9
Please RT.
www.chicagobooth.edu/research/cen...
This, from my colleague
Brent Neiman says it all.
Trump admin used the results from his paper to justify tariff rates. Except they used the completely wrong number.
Staggering 🤯
www.nytimes.com/2025/04/07/o...
Brent Neiman says it all.
Trump admin used the results from his paper to justify tariff rates. Except they used the completely wrong number.
Staggering 🤯
www.nytimes.com/2025/04/07/o...
April 7, 2025 at 10:36 AM
This, from my colleague
Brent Neiman says it all.
Trump admin used the results from his paper to justify tariff rates. Except they used the completely wrong number.
Staggering 🤯
www.nytimes.com/2025/04/07/o...
Brent Neiman says it all.
Trump admin used the results from his paper to justify tariff rates. Except they used the completely wrong number.
Staggering 🤯
www.nytimes.com/2025/04/07/o...
Great point from @andreyfradkin.bsky.social: as a movement’s policies become more incoherent and harmful, the quality of “experts” who are willing to publicly defend it decreases precipitously.
April 6, 2025 at 2:14 PM
Great point from @andreyfradkin.bsky.social: as a movement’s policies become more incoherent and harmful, the quality of “experts” who are willing to publicly defend it decreases precipitously.
Ambitious and important JMP from Ruchi Mahadeshwar (with Alex Zhou) at Brown. They use experimental variation and a structural model to show that 1) on margin, disutility from sex work is *16* times that of non-sex work and 2) labor supply of sex work is v. elastic. drive.google.com/file/d/15oXa...
January 15, 2025 at 3:47 PM
Ambitious and important JMP from Ruchi Mahadeshwar (with Alex Zhou) at Brown. They use experimental variation and a structural model to show that 1) on margin, disutility from sex work is *16* times that of non-sex work and 2) labor supply of sex work is v. elastic. drive.google.com/file/d/15oXa...
Halfway though this and loving it.
January 14, 2025 at 4:51 AM
Halfway though this and loving it.
If you squint you can almost see our boy riding that worm. Incredible.
January 6, 2025 at 2:25 PM
If you squint you can almost see our boy riding that worm. Incredible.
Excellent post. One prediction I’d disagree w/ is corporate adoption. Many medium/large firms already adopting AI extensively, but using internal models trained in-house for specific tasks
My 2025 prediction: enterprise adoption will be more tailored/in-house models, less frontier closed models 1/2
My 2025 prediction: enterprise adoption will be more tailored/in-house models, less frontier closed models 1/2
January 3, 2025 at 3:00 AM
Excellent post. One prediction I’d disagree w/ is corporate adoption. Many medium/large firms already adopting AI extensively, but using internal models trained in-house for specific tasks
My 2025 prediction: enterprise adoption will be more tailored/in-house models, less frontier closed models 1/2
My 2025 prediction: enterprise adoption will be more tailored/in-house models, less frontier closed models 1/2
Great post. But this def of reasoning set bar too low imo. Take data on # of boys/girls in class. Write this prompt: calc fraction of boys v. Girls. If > .5, then 1, if </=.5, then 0. By this def, my “model” is reasoning. It is drawing conclusion (more boys than girls) by gen inference from data.
January 2, 2025 at 5:06 PM
Great post. But this def of reasoning set bar too low imo. Take data on # of boys/girls in class. Write this prompt: calc fraction of boys v. Girls. If > .5, then 1, if </=.5, then 0. By this def, my “model” is reasoning. It is drawing conclusion (more boys than girls) by gen inference from data.
Great summary right here on the state of AI in 2024. I particularly like the part on “LLMs need better criticism”. As Simon says, arguing there are no good use cases and that it’s all unreliable “plagiarism” slop is a pretty good tell on how well you understand the tech.
December 31, 2024 at 9:18 PM
Great summary right here on the state of AI in 2024. I particularly like the part on “LLMs need better criticism”. As Simon says, arguing there are no good use cases and that it’s all unreliable “plagiarism” slop is a pretty good tell on how well you understand the tech.
Good write up of Toner-Rogers result on introduction of AI to material science labs leading to increase in productivity, w/ almost all gains going to those who were already most productive
But perhaps most interesting result was buried: the scientists hated the tool 1/2
www.wsj.com/economy/will...
But perhaps most interesting result was buried: the scientists hated the tool 1/2
www.wsj.com/economy/will...
December 29, 2024 at 8:00 PM
Good write up of Toner-Rogers result on introduction of AI to material science labs leading to increase in productivity, w/ almost all gains going to those who were already most productive
But perhaps most interesting result was buried: the scientists hated the tool 1/2
www.wsj.com/economy/will...
But perhaps most interesting result was buried: the scientists hated the tool 1/2
www.wsj.com/economy/will...
Wow this paper is badass. But also the replies are hilarious.
December 28, 2024 at 1:44 PM
Wow this paper is badass. But also the replies are hilarious.
Interesting paper on the convergence of representation hypothesis.
In cases with high overlap in representation between different ANN models, these representations also aligned with human representations.
When models disagreed, none predicted human representation
www.biorxiv.org/content/10.1...
In cases with high overlap in representation between different ANN models, these representations also aligned with human representations.
When models disagreed, none predicted human representation
www.biorxiv.org/content/10.1...
December 27, 2024 at 9:40 PM
Interesting paper on the convergence of representation hypothesis.
In cases with high overlap in representation between different ANN models, these representations also aligned with human representations.
When models disagreed, none predicted human representation
www.biorxiv.org/content/10.1...
In cases with high overlap in representation between different ANN models, these representations also aligned with human representations.
When models disagreed, none predicted human representation
www.biorxiv.org/content/10.1...
This looks like a must-read for economists. Lots of implications for UBI, the potential disruptions from AI, etc. I’m excited to read this closely over break.
December 23, 2024 at 2:39 PM
This looks like a must-read for economists. Lots of implications for UBI, the potential disruptions from AI, etc. I’m excited to read this closely over break.
Love this at the end.
December 19, 2024 at 3:31 PM
Love this at the end.