Where's The Evidence That AI Increases Productivity? (msn.com)
- Reference: 0180801478
- News link: https://it.slashdot.org/story/26/02/16/0316228/wheres-the-evidence-that-ai-increases-productivity
- Source link: https://www.msn.com/en-us/money/savingandinvesting/ai-isn-t-really-stealing-jobs-yet-that-doesn-t-mean-we-re-ready-for-it/ar-AA1KXQAi
> This decoupling — maintaining high output with significantly lower labour input — is the hallmark of productivity growth. My own updated analysis suggests a US productivity increase of roughly 2.7% for 2025. This is a near doubling from the sluggish 1.4% annual average that characterised the past decade... The updated 2025 US data suggests we are now transitioning out of this investment phase into a harvest phase where those earlier efforts begin to manifest as measurable output.
>
> Micro-level evidence further supports this structural shift. In our work on the [2]employment effects of AI last year, Bharat Chandar, Ruyu Chen and I identified a cooling in entry-level hiring within AI-exposed sectors, where recruitment for junior roles declined by roughly 16% while those who used AI to augment skills saw growing employment. This suggests companies are beginning to use AI for some codified, entry-level tasks.
Or, AI "isn't really stealing jobs yet," [3]according to employment policy analyst Will Raderman (from the American think tank called the Niskanen Center). He argues in Barron's that "there is no clear link yet between higher AI use and worse outcomes for young workers."
> Recent graduates' unemployment rates have been drifting in the wrong direction since the 2010s, long before generative AI models hit the market. And many occupations with moderate to high exposure to AI disruptions are actually faring better over the past few years. According to recent data for young workers, there has been employment growth in roles typically filled by those with college degrees related to computer systems, accounting and auditing, and market research. AI-intensive sectors like finance and insurance have also seen rising employment of new graduates in recent years. Since ChatGPT's release, sectors in which more than 10% of firms report using AI and sectors in which fewer than 10% reporting using AI are hiring [4]relatively the same number of recent grads.
Even Brynjolfsson's article in the Financial Times concedes that "While the trends are suggestive, a degree of caution is warranted. Productivity metrics are famously volatile, and it will take several more periods of sustained growth to confirm a new long-term trend." And he's not the only one wanting evidence for AI's impact. The same weekend [5] Fortune wrote that growth from AI "has yet to manifest itself clearly in macro data, according to Apollo Chief Economist Torsten Slok."
> [D]ata on employment, productivity and inflation are still not showing signs of the new technology. Profit margins and earnings forecasts for S&P 500 companies outside of the "Magnificent 7" also lack evidence of AI at work... "After three years with ChatGPT and still no signs of AI in the incoming data, it looks like AI will likely be labor enhancing in some sectors rather than labor replacing in all sectors," Slok said.
[1] https://www.ft.com/content/4b51d0b4-bbfe-4f05-b50a-1d485d419dc5
[2] https://www.aaas.org/sites/default/files/2025-10/BharatChandar_CSGWest_AAASEPI_ShapingAIFutureInYourState_Oct2025.pdf
[3] https://www.msn.com/en-us/money/savingandinvesting/ai-isn-t-really-stealing-jobs-yet-that-doesn-t-mean-we-re-ready-for-it/ar-AA1KXQAi
[4] https://www.employamerica.org/labor-market-analysis/dont-blame-ai-for-the-rise-in-recent-graduate-unemployment/
[5] https://finance.yahoo.com/news/ai-everywhere-except-data-suggesting-230358135.html
Perhaps too early to tell? (Score:1)
We already see the effect from Microslop on how bad the software quality is when AI is involved (to mention one example), but I don't think we have seen what the cost to overall productivity is... or even how to relate and quantify it in terms of productivity. My gut feeling is that a lot of people will soon be required to fix all the damage caused by AI. We may only be at the very early stages of that phase. Either way - this must surely be a big productivity killer (wasted effort).
Re: (Score:3)
There is also going to be lag while companies offload their talent onto AI. If those companies are expecting their human talent to retain their abilities, they might be wrong in that expectation. Telling their talent to siphon off their work to AI is only going to cause them to lose their intellectual edge. Use it or lose matters when it comes to brains and their use.
For a test, take yer basic entry level person and give them a math book. Tell them to learn the first 10 chapters but they should use AI as mu
Re: (Score:2)
Intense propaganda and physical abuse by data-manglers is supposed to convince every Joe-Peanut that the lower standards of service/performance/reliability provides by *.ai/LLM is the STANDARD to be maintained. Kinda like convincing ice-cream eaters to buy air-foamed store brand ice-creame instead of buying heavy creame, vanilla-beans and eggs and churning it yourself. When crap becomes the standard of excellence, excellence ceases to exist.
We just started. (Score:2)
I remember very similar articles to this when the tech bubble burst saying things along the lines of , "Computers increased productivity, but not as much as people expected." That was after 20 years of figuring out how computers fit into the workplace. AI has been shoehorned in over two years and we are expecting to be able to have ANY idea how this is going to play out?
Re: (Score:2)
Predicting the future isn't easy.
The .com bubble was pretty much "Oh hey, computers have been around for a while and they're great, now the internet is arriving so let's make sure we don't miss out. Let's pump all of the money into any company that has a web page or might consider making one, in case they conquer the internet."
LLMs feel to me more like, "Oh hey, someone added an LCD screen to a toaster and they claim next year they'll make the airline industry irrelevant, let's put all of the money into toa
Depends on the topic (Score:5, Interesting)
I've used LLMs for generating web interfaces and it does pretty good. Try to get it to code in assembly or have a container log traffic between other containers and the LLM will keep chasing it's tail trying to fix bugs.
Re: (Score:3)
Yeah thats my observation.
Give it a react web interface, or maybe a wordpress template, and it'll do fantastically. Its eaten every website on the planet, it knows what that looks like.
Ask it to optimize a lockless high throughput low latency kafka pipeline that needs to maintain an optimal memory pressure across all core counts, without race conditions or i/o contention, and its going to shit the bed and not know where to start. Actually, so will the first year uni graduate. But at least he can *learn*.
Re: (Score:2)
I think there are many people much more advanced than the first year graduate who cannot do your task. You may be a bit biased, remember: [1]https://xkcd.com/2501/ [xkcd.com]
You are comparing someone who worked in a very niche of the field (or used time to learn that niche) to a generalist. If you need your high-throughput kafka pipeline and want to compare to a kafka expert, you would need to compare to a specialist model fine-tuned on kafka and related tech.
[1] https://xkcd.com/2501/
Re: (Score:2)
Almost ANY building task involves "specialty" or "hi-IP" input.' from CS-grads building computer code to Amazon savages building water buckets. Yesterday I watched a film on members of the latter group ... savages if you will ... who raised water from river to tree-top housing by rope+winches with JUST the number of holding "knots" supportable by the vine-diameter from which the rope was created. Amazing "specialist" creation for an everyday task at the river; wife will be p*ssed without t
Re: (Score:2)
For shits and giggles, I tested it against x86-64/linux (as) and it did fine. Can't remember which I used... probably one of the Qwens.
Now granted, it wasn't super complicated, it just basically called stat(2) and write(2) to display the output... but I mean, it did do it, no libraries.
Call me old fashioned but (Score:2)
maybe aim for same payroll AND more productivity?
Re: Call me old fashioned but (Score:4, Interesting)
I have seen it work in both directions. I was a technology wrangler when desktop publishing took off. It was great that you no longer needed letraset and actual paste, but much of it turned into, if you can do this on your computer then you no longer need an assistant (then secretary). But that also took another valuable brain and set of eyes out of the process. Conversely, we worked on some multi year projects with LEGO, and watched as they automated more and more of their US plant. Adding computer control to sorting and packing lines, and automating such mundane tasks as making sure a minifig heads were on straight. They prided themselves for never losing a person from this, they would assign them to a new project or product. This was about the time they were turning the corner on adding outside IP to their lines. It allowed them to use experienced people to staff these new initiatives, and from all indications, it kinda worked.
I don't think they care if it does (Score:5, Insightful)
The possibility of replacing every white collar worker is just too tantalizing. Even if it costs more replacing those workers moves more power to the top.
We are past the point where billionaires are just trying to make more money. We are at the point where they want more power. More power means being able to decide who gets the function in society and that means controlling who gets to work. The best way to do that is to limit the amount of available work.
It breaks a dependency the billionaires have on us working stiffs. Over and over again when we catch billionaires candidly they show complete disgust for us. So if they have to spend an extra 20 or 30% of their already limitless wealth to no longer have to interact with or depend on us that would be a small price to pay
Basically it's the end of capitalism just not the way that the blue haired girls keep telling you we should do it
Hogwash (Score:2)
Our strong GDP is evidence of the extremely wealthy swapping around companies, buying and selling paper wealth with paper money faster than ever before. That's it. Wealth creation and transfer is not captured in the metric of GDP.
Re: (Score:2)
Um, if wealth creation and transfer is not captured in the metric of GDP then how can our strong GDP be evidence of the extremely wealthy swapping around companies?
Re: (Score:2)
> if wealth creation and transfer is not captured in the metric of GDP then how can our strong GDP be evidence of the extremely wealthy swapping around companies?
More money is changing hands without more work being done. The GDP can increase without more wages being paid. Therefore it can be measuring something which says absolutely nothing about the economic health of the nation, which is not defined by the numbers in billionaires' bank accounts, but by those of the lower class... many of whom don't even have accounts except maybe Cash App or similar. You need that for begging these days, nobody carries cash.
Re: (Score:2)
I think what's more accurate to say is that GDP is not a good measure of "productivity". It measures the spending we've made (purchases, investments, etc.), but that's the nominal value of things, which yes can be companies buying and selling paper wealth. It doesn't measure how that wealth is redistributed, but more importantly, wealth is not productivity. If I increase the price of my product by 50% and sell it, I've increased GDP, but I'm not more productive if I produce the same amount, despite what the
Is this raw? (Score:2)
> ... in 2025's payroll growth ...
I'm guessing this is raw data: It hasn't been adjusted for unemployment or inflation.
Slow growth can mean employees are being dismissed: Which has happened much in the last 12 months. While entry jobs seem to be taking less of a hit, suggesting AI isn't so useful, senior employees represent a large portion of the payroll: Losing a few of them will have a large impact. Then, there's the dismissal of professionals because so much health, safety and auditing, is no longer a must-have.
Inflation has lim
Productivity (Score:4, Informative)
Productivity always rises in a recession. [1]https://www.bls.gov/opub/mlr/2... [bls.gov]
[1] https://www.bls.gov/opub/mlr/2014/beyond-bls/does-the-productivity-of-individual-workers-increase-during-recessions.htm
Sir the Dow is 50,000! (Score:1)
Now don't you feel silly? Also drug prices have dropped over a 1000%. People just get paid to take drugs now like how your grandma warned you what happened in the 80s. Pay no attention to the 41% increasing coffee prices...
Marketing Ramp Up (Score:1)
This looks like another round of the recent hype articles about AI. They need to make money so the field is being flooded with information about progress. Claude AI is so much better (it's not). Spotify programmers don't code anymore! Mathew Broderick is skipping work today! A Harvard study found computer grad employment down 9% or so since 2022 and Stanford estimates 20%. Most articles over the past two years are about the same. I'm not sure the Niskanen Center is on top of this.
AI's #1 economically valuable use case... (Score:1)
...is enabling bosses to imply/threaten summary replacement by it, to get the proles to work twice as hard for the same pay.
Simple empirical evidence (Score:2)
Ask average joe to write a script that shows the time elapsed between two given dates. One time joe is allowed to use AI, one time not.
You will see a 100% increase in productivity for that task.
Fax machines, computers, or internet.. (Score:4, Insightful)
Go read some history, you'll find these same articles.
I just completely automated most of our IC2 and IC3 network engineering processes with GPT agents.
Not redid; the AI executes the same workflows and tools better, cheaper, and faster. Compressing months of work to two weeks.
It's real.
Re: (Score:2)
I have 30 years experience and in the last year I've written only a few lines of code. It's real for me too. I was able to convert all our unit tests to a new library in a few days it would have take me a month or more manually.
Benefits (Score:3)
The benefits will be sure to start rolling any day now. Aaaaaaaaaaaaaaaany day.
"It productivity researcher" (Score:4, Interesting)
Who works at Stanford's HAI and only seems to write about AI. "Advancing AI research, education, and policy to improve the human condition."
"This week America's Bureau of Labor Statistics showed a 403,000 drop in 2025's payroll growth - while real GDP "remained robust, including a 3.7% growth rate in the fourth quarter.""
This gibberish doesn't even warrant the customary correlation != causation.
[1]https://fred.stlouisfed.org/gr... [stlouisfed.org]
"This decoupling - maintaining high output with significantly lower labour input - is the hallmark of productivity growth."
Wait.. what? Significantly lower labor input? "Drop in payroll GROWTH" != "significantly lower labor input"
From page 4.
"
Jan 2025 158,268
Dec 2025 158,497
"
[2]https://www.bls.gov/news.relea... [bls.gov]
"This is a near doubling from the sluggish 1.4% annual average that characterised the past decade... "
Just look at the chart, the numbers are all over the place and we have genius's comparing averages with a snapshot to advance a narrative. The weasel words and manipulation of both the data and the reader are absurd.
[1] https://fred.stlouisfed.org/graph/?g=33N6
[2] https://www.bls.gov/news.release/pdf/empsit.pdf
I'm getting lots extra done (Score:2)
But all of it is work that I never would have done in the first place.
I had Gemini write. me some elisp so that I can bring up a buffer with all my CLs in it. I wanted something I could use inside emacs rather than moving 2 virtual desktops over to p4v.
Clearly, this work is not critically important. But it does make my life marginally better. But it's so low priority that it wouldn't have been worth my time to figure out how to interact with p4 and dump everything into an org buffer.
I greatly suspect it's w
So tentative answer (Score:5, Interesting)
So tentative answer is increase in productivity in some sectors but not all, and no widespread unemployment, but also not a clear massive boost in productivity. So both the largest worries about AI in terms of replacing people and the largest claims that this technology is all junk turn out to be likely not correct, but with more data still needed to be sure. So the question then becomes will this evidence impact at all the positions of either the AI-hypesters or the anti-AI groups at all, or alternatively will both groups just ignore it or try to spin it to fit their preexisting position?
Re: (Score:2)
After careful consideration, I'm going to say "meh" to everything you just said.
What I'd say is that the quoted articles are designed to pour oil on the boiling water of investor's fears while the AI investment bubble explodes.
Re: (Score:2)
Aren't bubbles usually imploding?
Re: (Score:2)
Maybe metaphorically, but an actual bubble explodes when it pops. The gas inside a bubble is at a higher pressure than the gas that surrounds it. See: [1]https://phys.libretexts.org/Bo... [libretexts.org]
[1] https://phys.libretexts.org/Bookshelves/Classical_Mechanics/Classical_Mechanics_(Tatum)/20%3A_Miscellaneous/20.02%3A_Surface_Tension/20.2.01%3A_Excess_Pressure_Inside_Drops_and_Bubbles
Re: (Score:2)
> So both the largest worries about AI in terms of replacing people and the largest claims that this technology is all junk turn out to be likely not correct,
Yet.
> but with more data still needed to be sure.
An important caveat.
Re: (Score:2)
The differences in comparison to previous industrialization, like automated looms, trains, wood/metal lathes, automated phone systems, is that those machines all produced deterministic results. They could be understood by most mechanical or electrical engineers. You could see the obvious benefit of Ford's assembly line, and use that new fangled film technology to record what it looked like inside the new motor plants.
The LLMs have a lot of emergent properties. They are somewhat deterministic, as far as s
Re: (Score:2)
Yes, it's clear that some tasks can be automated well with AI when using specially trained models. Classification tasks and such comes to mind.
It's also clear that there is some productivity increase overall possible by smart use of AI. I found for example that instead of reading Microsoft's documentation, it's often more efficient to ask Copilot about it. Not surprisingly Microsoft's AI is pretty decent at knowing Microsoft's own stuff.
My worry is having to waste time fixing someone else's AI vibe coded sl
Re: (Score:2)
Like Windows 11?