Only 4 percent of jobs rely heavily on AI, with peak use in mid-wage roles Workers in just four percent of occupations use AI for three quarters of their tasks, according to research from Anthropic that explores how its Claude model is used. The research found that roughly 36 percent of occupations incorporate AI for at least 25 percent of their tasks. These findings align with previous reports …
4% is far far higher than I would have imagined.
so far the only people using AI that I've seen in industry have been a certain subset of job candidates trying to cheat their way through the take-home technical tasks. Easily caught because the tasks I've given are designed to fool AI and be done properly by a human and the cheaters never even check to see if their code executes properly.
one of ours luckily has only 1 relevant comparable source in its training data from some other company's coding task, which helpfully specifies they have to name a variable a certain name (which then doesnt get used) so i look out for that variable
Another has some automation required and the AI just assumes the names of locators which clearly don't exist but it's not opening an application to find locators it's a glorified chatbot, so it looks fine but won't run
Other ones i use ambiguous writing which a human would understand and I got a bit cheeky with a rogue instruction with the font set to the same colour as the background which gets blindly copy-pasted into the prompt
Even if they got past all that (they don't) the next interview is for them to walk me through the code so it's a nice timesaver but they'd get caught at the next stage anyway (and it prevents me from being rude when I inevitably unmask them after taking time out of my busy day)
Pseudo-AI is useless in customer-facing jobs, and even more in highly specialised/critical jobs like surgury.
So pseudo-AI can basically only be used in administrative tasks and, even then, it's not used all that much.
Well gosh, I do believe I've read that some lawyers found out that AI won't write their judicial papers all that well.
I'm waiting for the report that says that companies have banned the use of AI for internal reports.
Ah, who am I kidding ? That will never happen.
> Pseudo-AI is useless in customer-facing jobs
Alot of customer service time is spent by cosplaying as an FAQ system, basically reading back to customers what they could have found themselves if they bothered to open the company website and locating the "Questions" sub-page. LLM based systems can take a lot of this boring work off service agents, allowing them to lower waiting time for customers who actually require their assistance.
> if they bothered to open the company website and locating the "Questions" sub-page
Or if the companies would put useful information front and centre on their home pages, instead of glossy photos and mindless videos.
LLMs: making it easier to justify aggravating web design.
(I may, possibly, be frustrated after yet another week spent trying to dig for actual specs to get my random assortment of boxes to talk to each other but being continually told Acme Inc has The Solution For You - fine for you, but what about me? Mutter, swear)
> LLM based systems can take a lot of this boring work off service agents, allowing them to lower waiting time for customers who actually require their assistance.
... which is great, until your LLM hallucinates a legally binding answer
https://www.bbc.co.uk/travel/article/20240222-air-canada-chatbot-misinformation-what-travellers-should-know
When I call customer service, I **NEED** a human being. Talking to an AI (or a human reading from a script) is a waste of time, as I've already looked at the company website, tried to resolve the issue myself, and found I can't, either due to lack of access or lack of information. (And having the phone system tell me that I should take a look at their website for their FAQ is more wasted time, as I've already done so.) Can we please implement the shibboleet protocol?
In related news, I recently told Verizon that incoming calls with caller id that isn't in the Verizon-preferred format are silently sent to voicemail - no ringing on either end, no missed call notification for the recipient. The tech support droid ended up marking the ticket as "no troubleshooting performed" because I refused to wipe my wifi and bluetooth network lists and passwords as part of "troubleshooting" - after I explained multiple times that this happens on both Android and Apple phones, and we "fixed" the problem by changing the caller ID on the calling phone. Never could talk to someone who actually knew what they were talking about, so they'll probably never fix it. (+1 234 555 6789 fails, 234 555 6789 succeeds.)
Exactly. AI can and probably will, replace most customer service operations. Assuming it's cheaper than humans, that is. And the results will surely be dreadful at least initially. Abandon all hope of intelligent response, ye who enter here.
On the other hand, there is some chance that the quality of AI response will improve over time (as long as improvement doesn't cost much money) and there seems to be no hope whatsoever that the quality of customer service at companies like Comcast will ever improve.
Well, Zoom and similar conferencing tools have automatic minuting now. I troubled to read through some and while it's mostly correct, the most common mistake was for the transcript summary to attribute 'next step' actions to the wrong person. This will undoubtedly be improved on in the future - so long as anyone bothers to complain to the tool makers about it.
Well, since ASR is basically a seq-2-seq encoding, and nowadays mostly done by transformers, which are an ML model architecture...yes, speech recognition is very much AI.
And the above post also isn't about the speech recignition itself, but the summarization of an entire meeting worth of transcribed audio to minutes, which is a summarization task, and that is definitely done using LLMs these days.
The physical job replacement by technology has already happened / continues to happen. It has been many years since a team of men with shovels has been replaced by one man in a JCB. The analogue for physical jobs is mechanisation rather than AI. To take it a step further and replace the man driving the JCB might be 90% possible with perhaps simple feedback automation not necessarily AI, but there will always be that last bit that is the hard bit.
Like automatic landings for aircraft, been happening since the 50s, or for civil aircraft the 1960s. Still 60 years on we have 2 human brains watching its every move.
So AI finds a place in admin, but there is still that last 10% that is hard to fix - bosses may consider that the 10% error is worth it for a large reduction in staff costs. Customers are likely to disagree because that's a lot of customers for a big business and a lot of unwanted unhappy noise flying around the internet.
As President Elongated Muskrat is finding out. One of his precious Cybertruk thing just totalled itself when on Autopilot.
https://www.pcmag.com/news/tesla-cybertruck-crashes-into-pole-while-running-latest-version-of-fsd
didn't he promise that it would be fully certified in 2018?
Maybe, but there are now a whole bunch of Tesla hearses available to profit from that bonus situation (NO 2019, UK 2021, NL 2022, NL 2025), except maybe in France where they are slightly losing favor due to current geopolitics.
I'd bet that's why the French are building this giant solar-powered wanksy, clearly visible from StarLink orbit, to show'em who's the bigger boss!
From the article and the occupations cited, it would seem those occupations probably represent, at best 2% of the total workforce.
It would be useful if the original research had done the mapping of occupations to actual workers, as that really would of put a hole in the AI hype balloon…
remember when cell bricks hit the market and only your real-estate agent, and wealthy aunts had them in their Lincoln and cadillac?
Language models have already taken over a large part of my world, and that includes the food industry. Every fast food restaurant uses AI now to take and place orders. Doordash is AI, shopify as well. It's beginning to make its way into inventory lists and such. All online ordering is pretty much AI now. tax collection have gone AI. King frumpy the second just dictated that aviation and financial services be turned into ai run by musk, and even visa has inked a contract to turn over visa transactions to musk and company.
They have the more advanced models, but they have to get people using it, and they cannot power it yet. To make ai as obligatory as they want it to be, they will have to build a cottage industry around it. That means they need very, very large databases collecting you, they need raw elements to build the things, and they need to power it. It will first be powered by vast amounts of oil and gas. A LOT of oil, augmented by solar, wind, and so on. The first nuke stations have been inked to help power these things. Next will be nuclear fusion, which many are saying is right around the corner. The east and west are in a race to be the first. That could, potentially save humanity from its self.
Right now I, and many others are resisting the AI takeover with everything we have, but there will come a day soon. I will no longer have a choice. It will be decreed to me to use it, and I will sit in front of a computer like homer simpson and weep. I will cry on my keyboard with pride in my eyes as I speak to the computer for the first time, and it tells me to kill myself.
I don't think you know what AI is.
Your example "Shopify is AI"
Bullshit.
Shopify has a few elements where you can get ChatGPT to write some text, maybe generate some images, but that's it.
Shit like remembering customer preferences, abandoned cart retrieval, single sign on is NOT AI.
int main(enter the void)... The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
