32% of Senior Developers Say Half Their Shipped Code is AI-Generated (infoworld.com) 57
In July 791 professional coders were surveyed by Fastly about their use of AI coding tools, reports InfoWorld. The results?
"About a third of senior developers (10+ years of experience) say over half their shipped code is AI-generated," Fastly writes, "nearly two and a half times the rate reported by junior developers (0-2 years of experience), at 13%." "AI will bench test code and find errors much faster than a human, repairing them seamlessly. This has been the case many times," one senior developer said...
Senior developers were also more likely to say they invest time fixing AI-generated code. Just under 30% of seniors reported editing AI output enough to offset most of the time savings, compared to 17% of juniors. Even so, 59% of seniors say AI tools help them ship faster overall, compared to 49% of juniors. Just over 50% of junior developers say AI makes them moderately faster. By contrast, only 39% of more senior developers say the same.
But senior devs are more likely to report significant speed gains: 26% say AI makes them a lot faster, double the 13% of junior devs who agree. One reason for this gap may be that senior developers are simply better equipped to catch and correct AI's mistakes... Nearly 1 in 3 developers (28%) say they frequently have to fix or edit AI-generated code enough that it offsets most of the time savings. Only 14% say they rarely need to make changes. And yet, over half of developers still feel faster with AI tools like Copilot, Gemini, or Claude.
Fastly's survey isn't alone in calling AI productivity gains into question. A recent randomized controlled trial (RCT) of experienced open-source developers found something even more striking: when developers used AI tools, they took 19% longer to complete their tasks. This disconnect may come down to psychology. AI coding often feels smooth... but the early speed gains are often followed by cycles of editing, testing, and reworking that eat into any gains. This pattern is echoed both in conversations we've had with Fastly developers and in many of the comments we received in our survey...
Yet, AI still seems to improve developer job satisfaction. Nearly 80% of developers say AI tools make coding more enjoyable... Enjoyment doesn't equal efficiency, but in a profession wrestling with burnout and backlogs, that morale boost might still count for something.
Fastly quotes one developer who said their AI tool "saves time by using boilerplate code, but it also needs manual fixes for inefficiencies, which keep productivity in check."
The study also found the practice of green coding "goes up sharply with experience. Just over 56% of junior developers say they actively consider energy use in their work, while nearly 80% among mid- and senior-level engineers consider this when coding."
"About a third of senior developers (10+ years of experience) say over half their shipped code is AI-generated," Fastly writes, "nearly two and a half times the rate reported by junior developers (0-2 years of experience), at 13%." "AI will bench test code and find errors much faster than a human, repairing them seamlessly. This has been the case many times," one senior developer said...
Senior developers were also more likely to say they invest time fixing AI-generated code. Just under 30% of seniors reported editing AI output enough to offset most of the time savings, compared to 17% of juniors. Even so, 59% of seniors say AI tools help them ship faster overall, compared to 49% of juniors. Just over 50% of junior developers say AI makes them moderately faster. By contrast, only 39% of more senior developers say the same.
But senior devs are more likely to report significant speed gains: 26% say AI makes them a lot faster, double the 13% of junior devs who agree. One reason for this gap may be that senior developers are simply better equipped to catch and correct AI's mistakes... Nearly 1 in 3 developers (28%) say they frequently have to fix or edit AI-generated code enough that it offsets most of the time savings. Only 14% say they rarely need to make changes. And yet, over half of developers still feel faster with AI tools like Copilot, Gemini, or Claude.
Fastly's survey isn't alone in calling AI productivity gains into question. A recent randomized controlled trial (RCT) of experienced open-source developers found something even more striking: when developers used AI tools, they took 19% longer to complete their tasks. This disconnect may come down to psychology. AI coding often feels smooth... but the early speed gains are often followed by cycles of editing, testing, and reworking that eat into any gains. This pattern is echoed both in conversations we've had with Fastly developers and in many of the comments we received in our survey...
Yet, AI still seems to improve developer job satisfaction. Nearly 80% of developers say AI tools make coding more enjoyable... Enjoyment doesn't equal efficiency, but in a profession wrestling with burnout and backlogs, that morale boost might still count for something.
Fastly quotes one developer who said their AI tool "saves time by using boilerplate code, but it also needs manual fixes for inefficiencies, which keep productivity in check."
The study also found the practice of green coding "goes up sharply with experience. Just over 56% of junior developers say they actively consider energy use in their work, while nearly 80% among mid- and senior-level engineers consider this when coding."
AI coding (Score:5, Insightful)
That said, Google's AI is remarkably bad as a search engine replacement. It's bad enough that my reflexive reaction is that it is lying to me (a reflex learned unfortunately from experience).
Re: AI coding (Score:3)
Re: (Score:1)
That said, Google's AI is remarkably bad as a search engine replacement.
Googles search engine replacement is already remarkably bad compared to their search engine.
Their AI is just garbage icing on the garbage cake.
They have been removing search flags at an alarming rate and I'd swear they gimped how double quotes used to work to prevent any literal searches.
More and more sentences will have words actively ignored by google as it invents some query to use instead that often has little to do with my query.
Pasting an error code is laughable. No amount of quotation will produce u
Re: AI coding (Score:5, Insightful)
Re: (Score:2)
Yeah, this started long before AI, and the AI just made it worse.
Re:AI coding (Score:4, Interesting)
They have been removing search flags at an alarming rate and I'd swear they gimped how double quotes used to work to prevent any literal searches.
I'm glad it's not just me - I've caught it ignoring double quotes but assumed that it was an anomaly.
I only use Google when I'm not satisfied with DDG results. And that's fairly often - sometimes I can get ten or more relevant hits on Google when DDG produces none. But lately, the amount of crap I'm getting from Google has me doubling down on DDG. Google really is on the verge of becoming irrelevant when it comes to any results that don't involve buying something.
Re:AI coding (Score:5, Interesting)
This is how I use AI, to help me find answers
I recently started using a new embedded processor. It had a 2000 page datasheet
I got answers faster by asking Perplexity than I could by manually searching the doc
I still read the doc and wrote all of the code myself, but the AI helped a lot
Another area where AI helps is implementing a new function. I ask for sample code, read it, understand it, then write my real code
Re: (Score:2)
Re: (Score:2)
Probably a middle of the road STM32 or something similar .
A modern SoC like what's in your smartphone doesn't have a 2000 page datasheet. It doesn't have a datasheet period - Qualcomm and such work with people who develop products and they have a whole collection of documents from register lists, to subsystem documents, to various technical addenda like boot processes, ROM elements, etc.
A low end microcontroller probably has a datasheet on the order of a few hundred pages that documents ever
At the API level and not at the app level (Score:2)
Same. AI helps in calling library API boilerplate code. Less so, by a long shot, at application level changes.
Re: (Score:2)
I don't so much see AI as a Stack Overflow replacement, per se; but as more of a research assistant that will correlate things like Stack Overflow with the official documentation, my own code, and the code of the other developers at the company. It's fantastic at pointing me in the right directions. But I find that I have to repeatedly tell it to review and cite its sources. And the code it generates can be a decent starting point, but has always needed a lot of work to make it production-ready. It's far
Re: (Score:2)
for me AI is replacing the search engines, they have made the search so bad full with ads and irrelevant answers that I don't know if AI is that much better or search is that much worse. Probably a bit of both, but I am confident that the enshittification of AI will occur. We are currently improving AI but it won't be long before its entire purpose will be to sell me stuff I don't need.
Re: (Score:2)
60% of the time, it works *every* time.
As with any tool ... (Score:4, Interesting)
... you learn the nuances of your chosen tools. You learn which uses enhance productivity, and which are more likely to detract.
In any case, it all gets tested by me and then QA-ed by others before seeing the production light of day. What is this "trust" thing people keep babbling about?
Re: (Score:2)
I would say they are testing every feature at once and we have to fight with being lab rats... maybe forever while they desperately try to train it off our interactions on how to replace us. Privacy, cost, and vendor lock-in is a big concern as well; oh, and don't forget how relaxing simple coding can be between harder problems and how it's going to only leave the difficult bits and us having to understand it's increasingly complex slop needing corrections.
The biggest PAIN is API and syntax errors. Most th
Re: (Score:2)
Tools generally don't change erratically, allowing you to learn how they work. AI is just a huge hot mess that changes at the whims of the latest fashion trend. I find it hard to take AI seriously as a "tool". It reminds me more of those "daily random" web sites for people with writer's block.
Why would they say that? (Score:4, Interesting)
It's a liability of possibly insecure/plagiarized code that is directly on them, shouldn't they keep it quiet?
Re: Why would they say that? (Score:4, Funny)
Because they are bad engineers?
Re: (Score:1)
That was my first thought. Immediately followed by "if they can't write code as easy as writing proper sentences in their native language, then they don't deserve to be senior anything."
Re: (Score:3)
They are saying that as surveyed by Fastly. (They aren't Surely, Securely, or Reliably - infer from that what you will.) The responses may not be truthful or representative of all developers.
Re: (Score:2)
Re: (Score:2)
Internal Suggestion Engine (Score:1)
No doubt it is effective (Score:2)
Re: (Score:3, Interesting)
I used to keep a library of PowerShell admin scripts so I wouldn't have to rewrite them when I needed them again. Now I can ask ChatGPT to write them for me faster than I can find them in my library.
As you say, though, small chunks. When I've tried more complex tasks I end up spending more time debugging the AI's code than the task is worth.
Re: No doubt it is effective (Score:2)
Developers are efficient (Score:5, Interesting)
Artists may still be debating if they think art is defined by a lot of work. Developers are taking "shortcuts" since many decades. Copy & Paste code I wrote for another project? Of course! Refactor code into functions to be reusable? Not doing so is a code smell! Using libraries maintained by others? Yes, thank you! Standard snippets, code completion, refactoring features in the IDE? Why not? Using high level programming languages with a large STL? Yes, please!
Ignoring tools that may make you more efficient would be stupid. You don't have to use all or use them all the time, but of course people will evaluate when they are an advantage and then use them. 32% sounds like the number one can expect from people using AI as an advantage but not shoehorning it into everything just because. You won't get to 100% now or any time soon, but a third is what you can achieve from using it for productivity.
Re: (Score:3, Interesting)
Re: Developers are efficient (Score:4, Insightful)
Re: (Score:2)
That depends on who your customers are. The general public won't recognize quality and proper engineering if it bit them in the ass (Microsoft, I'm looking at you). But other customers, like airlines or the Department of War have their own capable staffs looking over suppliers shoulders.
In some cases, DoD contracts are granted in part based upon the competence of the suppliers staff. Contracts have been known to follow these people between companies.
Re: (Score:2)
In some cases, DoD contracts are granted in part based upon the competence of the suppliers staff. Contracts have been known to follow these people between companies.
That's exactly the problem they are trying to avoid. Imagine you are the CEO of a company. Do you want to lose your customer just because someone quits/dies? No you don't.
Re: (Score:2)
Lol, so many of the posts here reek of bitter developers who have clearly never worked in the halls of quality software engineering
Multiplication (Score:2)
So, if 35% of developers say half their code is AI generated, multiplying that out accounts for 17.5% AI generated. Another 17.5% is not AI generated.
What about the remaining 65%?
Re: (Score:2)
Correctly done today's code ... (Score:2)
... is little more than a formalized model and API description anyways, so I don't see this as too much of a big deal. Well established systems use CASE tools and code generation regularly. So much so that using AI may even be a superfluous intermediate step.
That AI is likely to take over the grunt work and bulk of coding shouldn't be news anymore to anyone paying attention.
Last week 100% of my code was AI generated. (Score:4, Interesting)
Last week 100% of the code I produced was AI generated. A couple of thousand lines of Rust resulting from a two hour discussion with Warp on Friday afternoon. I wanted a real-time animated data visualisation GUI and absolutely no desire spend anytime in researching the best Rust GUI library to do it in and then learn how to do it with that library. The end result is a working visualisation and looking over the code I now know how to drive that GUI library.
Can't trust dev estimates (Score:5, Interesting)
The problem with this survey is we can't trust developer estimates of how long it took them or how much time they saved. The METR report and Mike Judge's write-up [substack.com] show that quite clearly. Talk to me when Fastly includes actual timings of how long developers actually took to do the job with AI vs. without showing a statistically significant difference.
BALDERDASH! (Score:2)
The problem with this survey is we can't trust developer estimates of how long it took them or how much time they saved.
How dare you! I know in my gut that using AI to write a fancy five-function script save me at least 20 man-years or work! /s
Other things developers are bad at: designing sane GUIs and naming things. As a developer, I name this phenomenon YATDANGA (Yet Another Thing Developers Are Not Good At). ;)
Re: (Score:2)
One would think, right? Yet there's a constant stream of "new" done-to-death games in the Play Store that exist solely to appear at the top of the listings (because they're newer) and attract clicks to the ads in them. The people who write those games absolutely would use AI to do it if it'd let them do it faster, and we'd see that in the number of new releases (those lists don't care about how substantial the software is). It'd also make it less boring to create Yet Another X Clone. So, as Mike asks, where
AI generated... (Score:2)
...Human approved. Typing is for suckers.
Age (Score:2)
I'm 56. My code is of higher quality than when I was 25, but I've lost a lot of stamina and I now need more breaks when coding. So while there is possibly some truth in the statement that senior developers are better to catch and correct AI's mistakes, I guess another factor is that some senior developers have become slower than junior developers to create code.
Re: (Score:2)
Re:Age (Score:4, Insightful)
I guess another factor is that some senior developers have become slower than junior developers to create code.
I concur with that (I'm 52, no more 10-hour marathons at the keyboard for me!).
Note however that the proper metric to measure isn't "time elapsed to create code", but "time elapsed to get that code sufficiently bug-free that you can ship it to customers without it causing a mess".
Any high-school student (or AI) can write code that looks reasonable and passes the basic acceptance tests. The real trick is getting that last 1% correct, so that the code can "just run" indefinitely, unsupervised, for all use-cases, without crashing or misbehaving or otherwise requiring a human being's time to manage it.
Re: (Score:2)
I haven't seen much if any slow-down as I age, and I'm 60. What I have seen is that I spend more time thinking so I write less code to get the same result and need to do less debugging to get it working correctly. I also have a bigger library of code I can use without having to write it all from scratch so again I end up writing less code. This last is especially true for tests, and I already know the corner cases and odd cases out that many of my co-workers don't even realize need tested. But the correct m
In related news ... (Score:2)
Reminds me of an old Texas joke (Score:2)
A Texas rancher was talking with his friend from Vermont, bragging about how big his ranch was. "My ranch is so big, if I get in my car in the morning and drive all day, when the sun goes down, I'm still on my ranch." The Vermonter though for a minute, and said, "Yeah, I used to have a car like that too!"
These survey respondents said AI generates 50% of their code. Yeah, I've known some developers like that too.
Not a "senior coder", I use it "sometimes." (Score:2)
The big thing for me is that AI doesn't "write the code I put in production" - it provides guidance on techniques to use, or solves bugs I have written.
The same as StackOverflow for me. Just more personalized to my exact situation.
"I'm writing a shell script to ssh into a remote system and run some commands, I have to use some environment variables defined locally on the system I'm executing the script on, and other environment variables that are defined on the remote system I'm connecting to, and I can't r
New tools take time to learn (Score:2)
I posit that the greatest gains are achieved by the most coherent communicators.
I have seen massive gains from using AI. I turn on the microphone and dictate what I want. I break tasks into small achievable tasks with clearly defined goals and strict testing requirements. I make the AI commit changes to git and monitor the CI/CD pipeline. I constantly review code and train the models (I generally use three in parallel and cho
24 senior developers use code by Ai (Score:2)
Iteration and exploration (Score:2)