Survey of 2023–2024 data finds that AI created more tasks for 8.4 percent of workers.
See full article...
See full article...
Time saved by AI offset by new work created, study suggests
- Thread starter JournalBot
- Start date
I mean, did we ever think it was supposed to be anyone but the people paying for the chatbots? Companies aren't known for increasing wages when those workers have been more productive.
Upvote
259
(260
/
-1)
Not surprising when you consider that the rollout strategy for most corporations has been to license the tools first and try to find the applications for them after the fact.
Upvote
212
(212
/
0)
Ah, there's the problem, they listened to people who weren't software developers about what's capable of replacing a software developer.
Because the answer continues to be another software developer.
Upvote
202
(205
/
-3)
For the sake of the global environment, I hope usage of this power hungry tech decreases soon.
Upvote
88
(103
/
-15)
Upvote
181
(181
/
0)
As a lawyer, ChatGPT has created more work for me because if a client sends me some legal document generated by (or partially generated by) ChatGPT I spend more time revising it and fixing the problems than if they'd just asked me to draft the damn thing in the first place.
Upvote
285
(286
/
-1)
TheWerewolf
Ars Scholae Palatinae
This has been my experience as well, and why I don't usually rely on genAI for critical tasks.
Like coding with Visual Studio.
It's very much a sense of "Oh wow... how did it get that right?" when it does predict what I want to do, but far, far more often it's "For the love of God shut up - that's NOT what I want!" along with a lot of it getting some repetitive tasks right, but then finding out it gets a few of them wrong somewhere in the bulk, or just ignoring some cases forcing me to review every line of code again - something I don't have to do as much with my own hand-written code (since, you know, I wrote it and was there while it was being written).
I've reached the point where I've turned off most of the predictive coding features entirely. I can code faster and more reliably than it can.
Like coding with Visual Studio.
It's very much a sense of "Oh wow... how did it get that right?" when it does predict what I want to do, but far, far more often it's "For the love of God shut up - that's NOT what I want!" along with a lot of it getting some repetitive tasks right, but then finding out it gets a few of them wrong somewhere in the bulk, or just ignoring some cases forcing me to review every line of code again - something I don't have to do as much with my own hand-written code (since, you know, I wrote it and was there while it was being written).
I've reached the point where I've turned off most of the predictive coding features entirely. I can code faster and more reliably than it can.
Upvote
225
(226
/
-1)
Keynes thought that the advent of new technologies will make workers so efficient that we’d be working less and less and have more time for leisure and egalitarian pursuits.
But I guess he hadn’t anticipated “infinite growth” or “middle managers” and so we just create more work for ourselves. In some cases so much more that we’re working more than before and still not getting anything done.
Do you want more meetings? Because this is how you get more meetings.
But I guess he hadn’t anticipated “infinite growth” or “middle managers” and so we just create more work for ourselves. In some cases so much more that we’re working more than before and still not getting anything done.
Do you want more meetings? Because this is how you get more meetings.
Upvote
151
(152
/
-1)
Honestly, I've been avoiding AI tools like the plague in anticipation of these exact problems. I'm going to be better off writing functions and macros that will do all of the automation deliberately than I would be trying to coax an AI answer into working.
The efforts spent developing coding assistants would be better spent on making bespoke code generating functions on common repetitive tasks. You know what's faster than having an LLM spit out a comprehensive switch-case structure for a given enumeration? Turning that into a right-click->generate-code feature that I don't have to double check.
Upvote
125
(128
/
-3)
Right, it's like debugging someone else's code, i.e. the worst part of programming.
Upvote
197
(197
/
0)
There's one use case I've found for ChatGPT where it's really rather good.
ChatGPT is good at identifying what kind of grammatical mistake has been made in a given sentence. (In English, at least).
One place where conventional search falls down is when you have a sentence with what you think is a grammatical error -- but you don't know what the name of the error is. It is not always easy to work backwards from "Hrm, this arrangement of words and punctuation doesn't look quite right" to "Ahah! That's an X error." But ChatGPT can frequently tell you what kind of grammatical mistake has been made and give you the name of it.
It's always important to then take ChatGPT's answer to an actual grammar site and read about the rule in question, just to make sure that the app isn't hallucinating, but I've found it quite useful.
ChatGPT is good at identifying what kind of grammatical mistake has been made in a given sentence. (In English, at least).
One place where conventional search falls down is when you have a sentence with what you think is a grammatical error -- but you don't know what the name of the error is. It is not always easy to work backwards from "Hrm, this arrangement of words and punctuation doesn't look quite right" to "Ahah! That's an X error." But ChatGPT can frequently tell you what kind of grammatical mistake has been made and give you the name of it.
It's always important to then take ChatGPT's answer to an actual grammar site and read about the rule in question, just to make sure that the app isn't hallucinating, but I've found it quite useful.
Upvote
32
(44
/
-12)
It was never about a brighter future that frees mankind from the chores of production.
It was always about increasingly concentrating capital in a few hands through the work of alienated serfs.
It was always about increasingly concentrating capital in a few hands through the work of alienated serfs.
Upvote
103
(109
/
-6)
ubercurmudgeon
Ars Praefectus
Great, so studies have statistically shown that AI represents all that is soulless and wrong. Now what?
Upvote
33
(38
/
-5)
This has always baffled me. Lawyers already use document templates and cut and paste from previous documents.
The difference is that if you copy a real estate contract from a template, you do have to check over and update it, but you don't have to check to make sure the LLM didn't splice in some soft core erotica or add a clause that they building must be painted in green polka dots or the contract is voided.
Upvote
100
(101
/
-1)
Just think of all the lawyers caught submitting AI slop and then think about how much of that is slipping through and becoming legal record that a judge can later cite or be used to influence a major court decision.
Upvote
96
(96
/
0)
I find that cursor works reasonably well for many tasks, mostly for repetitive front end stuff though. Low stakes and quite verbose. But I must say the tools are getting better by the month.
My biggest gripe with AI for software dev longer term is the upward pressure it creates on writing more code. The best code is often the code not written at all, and AI coding so far seems at odds with that.
Upvote
63
(64
/
-1)
I'm reminded of a rather large leap back in time to "Roots". Kunta Kinte has just been whipped and is being given advice about how to be a slave. It went something along the lines of "Keep your head down, and don't stop moving, but don't don anything too fast or do too much. Be slow enough to keep your strength up, but fast enough to avoid the whip."
Sad that advice likely necessary before the 18th century founding of our country (I actually typed company, first - Freud, go away!) is now modern-day critical, need to know information for employees in the 21st century.
The illusion that human workers aren't commodities seems to ave been shattered again.
Sad that advice likely necessary before the 18th century founding of our country (I actually typed company, first - Freud, go away!) is now modern-day critical, need to know information for employees in the 21st century.
The illusion that human workers aren't commodities seems to ave been shattered again.
Upvote
44
(50
/
-6)
What I've found is that the useful task for GenAI is to identify the repetitive tasks and make suggestions on how I can automate them. It does a pretty good job at that, and I can immediately look at the response and know if it's correct or not. So I can quickly whip up scripts to automate the most repetitive tasks I do, saving time in the future in a reliable manner.
I've done some exploratory experiments at using an LLM to actually do the work, and it usually takes 3-6x more effort on my part, with less useful learning taking place that I can re-use in the future.
Upvote
31
(31
/
0)
Wow, if this result holds up in other studies then we might be getting close to the bubble bursting on the LLM industry.
It feels true, certainly. While LLMs have occasionally saved me hours, other times I have had to spend significant time double checking and correcting their outputs. A few times hallucinated answers sent me on a wild goose chase that wasted a lot of time trying to make a false approach work. You have to be careful and selective in using these things and know when to try something else.
I think LLMs and other current generative AI will still find use-cases and be somewhat useful, and sometimes very harmful, but they are vastly overvalued. Image generation will probably find more "valuable" use than LLMs. Increasing efficiency and advertising will help tide over the industry as it is forced to sell a flawed product at modest prices, until the "real thing" happens some time between 5 and 20 years from now. But that won't be an LLM dominated system.
It feels true, certainly. While LLMs have occasionally saved me hours, other times I have had to spend significant time double checking and correcting their outputs. A few times hallucinated answers sent me on a wild goose chase that wasted a lot of time trying to make a false approach work. You have to be careful and selective in using these things and know when to try something else.
I think LLMs and other current generative AI will still find use-cases and be somewhat useful, and sometimes very harmful, but they are vastly overvalued. Image generation will probably find more "valuable" use than LLMs. Increasing efficiency and advertising will help tide over the industry as it is forced to sell a flawed product at modest prices, until the "real thing" happens some time between 5 and 20 years from now. But that won't be an LLM dominated system.
Upvote
45
(46
/
-1)
VelvetGlove
Ars Tribunus Militum
LLM's are the programmer's dream: doesn't everyone want to spend less time coding so they can spend more time debugging?
\s
*edit: ninja crepuscularbrolly said it better
\s
*edit: ninja crepuscularbrolly said it better
Upvote
108
(108
/
0)
The tech bros push out more vaguely named models that are all slightly different so the researchers are always a step or two behind the latest and greatest and they can just say, but the research that shows minor improvements was only for the previous model, this one is better.
Upvote
31
(31
/
0)
Mungus the Unhyphenated
Ars Tribunus Militum
In my tech job, AI is of little benefit. It takes enough time to script an automation routine or perform a large task in PowerShell, for instance -- I don't need to spend additional time correcting AI output. About all it's good for is occasionally coughing up some generic canned script snippets when it takes too long to sift through all my files to find the same thing due to phone calls, emails and other distractions. So it can be a useful go-fer in a way. But that's it. I have too many things that demand detail and precision to risk leaving to an AI going astray without systematically checking its output. And at that point I could have done everything myself.
Upvote
25
(25
/
0)
I like to say "work always expands to fill the time given to it"
Upvote
53
(53
/
0)
Not very much of it. You'd need both an incredibly inattentive judge, and an incredibly inattentive opposing counsel, to let this stuff get through.
The fact that it's so easy to spot and the consequences are so serious for doing it makes it incredibly baffling that there are lawyers who are still doing it. But there are tens of thousands of lawyers practicing in North America and I'm aware of it happening like, half a dozen times, so it's not exactly an epidemic. But honestly, the guys who are getting caught for this are almost certainly lawyers who are being negligent in myriad other ways in their practice and were probably going to get nailed for something much less headline-grabbing sooner or later.
Upvote
36
(36
/
0)
Anyone with a brain would guess that but plenty of people are blinded by almost religious techno optimism. Having data like this is important
Upvote
33
(34
/
-1)
Earlier today, I took a company required training on ChatGPT where an HR-lackey was trying to demonstrate using ChatGPT for data analysis. That quote from Inigo Montoya kept playing in my head. Anyways, I watched as the trainer had to keep clarifying to ChatGPT what was wanted and seeing it provide dull responses full of errors/hallucinations. Meanwhile, I could do create the analysis in either Excel or R faster than the trainer could get ChatGPT to do bad work. My favorite part was that the trainer would stress, multiple times, the need for a human expert in the loop but then later exclaim how he didn't need to be an expert at analysis because of AI. [This timeline is infuriatingly stupid.]
Upvote
111
(111
/
0)
Upvote
26
(29
/
-3)
A few people will flag that this study is based on "old" results (a whole year ago? gasp), but that doesn't change the fact that the fundamental equation hasn't changed.
LLM models are too unreliable to use for high-value activities.
Equally, they're too expensive to use for low-value activities where accuracy tolerances are wider.
(In truth, I'm not sure quite how much of a market there is for the latter, other than in the seedier side of things. E.g. spam, social engineering exploits, etc)
Fundamentally, anything with a commercial or legal aspect which is generated by an LLM needs to be validated by a human with domain knowledge.
And that negates the cost-savings from using the LLM. Because otherwise, you're exposing you and/or your company to too much risk, as several lawyers and a few overly keen early adopters of LLMs-as-first-line-support have found out.
It's the same issue as per self-driving cars; until they're provably 99.99% better than a human in all scenarios, the commercial risk of deploying them is simply too high.
LLM models are too unreliable to use for high-value activities.
Equally, they're too expensive to use for low-value activities where accuracy tolerances are wider.
(In truth, I'm not sure quite how much of a market there is for the latter, other than in the seedier side of things. E.g. spam, social engineering exploits, etc)
Fundamentally, anything with a commercial or legal aspect which is generated by an LLM needs to be validated by a human with domain knowledge.
And that negates the cost-savings from using the LLM. Because otherwise, you're exposing you and/or your company to too much risk, as several lawyers and a few overly keen early adopters of LLMs-as-first-line-support have found out.
It's the same issue as per self-driving cars; until they're provably 99.99% better than a human in all scenarios, the commercial risk of deploying them is simply too high.
Upvote
68
(69
/
-1)
David Mayer
Ars Centurion
I wouldn't say they said it better, you certainly were more succinct.
Upvote
11
(12
/
-1)
Ohh, this just happened to me last week. We've deployed GitLab Duo across the firm so I thought I'd give it a try and asked it to implement some Spring function I'd never used before. It spit out something that looked competent but when I integrated it and tried to run it, it just gave back empty results.
Turned out there were two ways to implement it but you had to stick to one way or the other. The AI conflated the two and mixed up the implementation. Added a good half day of debugging and I found the answer at Stack Overflow anyway.
When we're all using the AI tools and no one's answering questions at Stack Overflow, where will the AIs crib their answers from?
Turned out there were two ways to implement it but you had to stick to one way or the other. The AI conflated the two and mixed up the implementation. Added a good half day of debugging and I found the answer at Stack Overflow anyway.
When we're all using the AI tools and no one's answering questions at Stack Overflow, where will the AIs crib their answers from?
Upvote
68
(68
/
0)
SubWoofer2
Ars Tribunus Militum
As did C. Northcote Parkinson, credited with Parkinson's Law, and several similar dicta besides.
And yes, AI has shifted my job towards proofreader/QC/fact checker.
Upvote
10
(11
/
-1)
Ouch. Yeah, the problem is that LLMs have no understanding. Of anything. They are very clever search and summary engines. That is all that they are
Upvote
22
(25
/
-3)
"You can see the computer age everywhere but in the productivity statistics." - Robert Solow, 1993
I have to laugh at the fact that this is still true. I don't think it will be true forever, I don't think it can be, but I think it's genuinely fascinating how true this has remained over the last 30+ years.
I have to laugh at the fact that this is still true. I don't think it will be true forever, I don't think it can be, but I think it's genuinely fascinating how true this has remained over the last 30+ years.
Upvote
25
(27
/
-2)