loading
loading
Check out the full 2026 Developer Survey results. Connect with Erin on LinkedIn and read her past articles on the Stack Overflow blog. Congrats to Stellar Question badge winner Franck Freiburger for winning the badge on their question Biggest integer that can be stored in a double. TRANSCRIPT [intro music] Ryan Donovan: Hello, and welcome to the Stack Overflow podcast, a place to talk all things software and technology. My name is Ryan Donovan, and today we are talking about the results of the 2026 developer survey. My guest for that is Erin Yepez. She's Market Research and Insights here at Stack Overflow. So welcome back to the show, Erin. Erin Yepez: Thank you for having me. RD: Of course, of course. So this is the 16th developer survey we've done, and I think it's the third one with AI on it. And this year sort of feels like the year that everybody's sort of standardizing using AI. I keep hearing people are not writing code anymore. What did we learn from the survey about AI this year? EY: So it's actually the fourth year that we've had AI on the survey. And we have some... continuity questions. In all of our surveys, we've asked about usage in general of AI tools. This year and last year, we specifically also asked about using AI agents. We also asked about how people are liking them. Do they find AI tools in their workflow to be favorable? We asked about what they're using them for specifically at work. This year, we asked quite a few new questions as well, just in terms of where do you find impact? We kind of tweaked some of the wording from last year. To be a little bit more specific, we asked also about whether anyone is avoiding using AI tools and why. And also, if anyone's building software with AI tools, what are they building? And if they're switching AI tools, what caused the switch? Things like that. Overall, though, I mean, we've seen compared to last year, usage still pretty high. We're at 83%. It's actually just about the same as last year for usage. But we added on to that. So now we specifically also know frequency of usage. Are you using AI tools at work daily, weekly, monthly or less? And then we also found out whether everyone's using AI tools daily, how many hours a day. Some great insights there. RD: Yeah. So you said 83% of developers use AI at work. How many of them are using it every day and how much every day? EY: All right. So overall, we have 73% of users indicate they're using AI tools daily for a variety of reasons, mostly using them for AI coding assistance or general purpose AI chatbots. RD: And, you know, you mentioned AI agents. How many people are using AI agents? EY: So we show when we break it up, 66% of respondents indicated using AI agents, 63% general AI chatbots. RD: So some of what we do is kind of technology leaderboards and the AI section has its own kind of technology leaderboards. Any surprises on who's top of the LLM or the AI agent leaderboards? EY: Surprises? I would say no. I think we still see the frontier companies have most market share, the most usage amongst our respondents. We asked about a lot of different companies though. So in particular, I know I was interested to see how some of the open weights companies were doing. I looked at some comparisons for, for year over year values as well. They're staying steady. So they're not decreasing, not really increasing either. The big one being DeepSeek that I'm thinking of. Which I know DeepSeek is not completely open weights now, but that's kind of the first one most people think of when you mention open weights. RD: Right. Yeah. I thought it was interesting that, you know, the top two, the Anthropic and OpenAM models are both 74%. So there's a lot of crossover there. A lot of people using both. EY: Yes. RD: So I think we've heard a lot about kind of tokenomics question. Are people, you know, is the data reflecting that? Are people switching? Are people you know, adjusting how they behave based on token costs? EY: No, we asked about specifically if in our question about switching AI tools, you know, if you have switched using tools in the past year, what made you switch? And people indicated that cost was part of that reason, but it was one of the lower ranked reasons. So only 6% indicated switching because of costs. We also asked another question specifically about tokens and if people limited themselves personally while using AI tools due to costs at work. And 33% indicated that they self-monitor their own usage, but we found out that 75% of people are changing their prompts for quality reasons. And then only 38% of users have indicated that they're changing their prompts due to cost. RD: Yeah. So we have a question about whether AI is making everybody productive, right? Like, is there some magic bullet tool? We looked at how many AI tools people use. Is it narrowing how many tools they're using? EY: Well, we haven't asked about how many AI tools people are using before in the dev survey. So we don't have anything to compare it to. We did ask last year how many software tools overall people are using. I know that was pretty high. On average, it was eight. And for more early career respondents, that number, the average number was higher than eight. So for most users in this year's survey, we did specifically ask about using AI tools. And we saw that upwards of 60% were using between two and five. RD: It's pretty clear that there's this pretty broad AI adoption right now. But every time I go to a conference, there's half the companies are trust and context and governance. Do people trust AI any more this year? EY: I think that there's a lot of conditions to it. So we asked a few different ways about how individuals decide that they trust the tools that they're using. Specifically, we asked, how much do you trust the output from AI tools or agents as part of your workflow? It's slightly a different variation of the trust question that we've asked in the past. In the past, we've asked just in a typical Likert sense, how much do you trust the accuracy of the AI tools you use at work? I trust it very much. I sort of trust it. Neutral. I don't trust it. I definitely don't trust it. So some variation of that, but this year we were a little bit more specific asking how much do you trust the output? And so select one of the following, uh, when I can verify it was what most people selected. So 48%, just about half of users trust AI tools when they can verify it. And opposed to that, we see that 7% have trust in AI tools for many tasks and decisions. RD: Right. For all the important stuff, they'll give it to AI. EY: Yeah. Yeah. So not very many people, but there were some people that indicated that. 1% indicated that they weren't sure or that they didn't use AI. So that's always in there. Even though that question was a conditional question in the survey, it wouldn't have showed up if you already indicated you don't use AI. RD: Right. One of the early kind of paradigms for AI was the retrieval augmented generation to sort of get some sense of information source. Is that something that's still important today? EY: Yeah, so 79% indicated that source attribution was either important or very important when evaluating an AI-generated answer. RD: And, you know, the RAG was sort of like to eliminate or mitigate the hallucinations that the AI was doing. Is there still a question about that? I mean, I know we asked about what were the sort of top pitfalls. EY: Yeah. Actually, we asked about, in the past year, has your company experienced a disruption due to AI? And if so, what would you indicate was the one that cost the most time slash money? And the top reason for that AI disruption was plausible but wrong code, which tracks with some of the survey results that we got last year, too, we asked the question in a different way, but we saw, again, top two responses as far as what caused people pause when using AI tools or what was giving them friction at work was almost right code. And then second was security concerns. RD: You know, I know one of the primary use cases of it is generating code. And it feels like the AI-generated code is putting a lot more pressure on code review. I think there was a fantasy or a dream that AI-assisted coding could be complete without human validation, but that doesn't seem to be quite the case. EY: Yeah, I would say, if anything, one of the bigger themes that I found in this survey, which tracks with some of the other research that we've done, is that In order to get ahead of those hallucinations and get ahead of the problems that come with your agent or coding assistants not being able to get fully correct code is the documentation, the context that comes with documentation. We see that coming up a lot in the survey. We have similarly asked about documentation in last year's survey too. So we can see that people are using AI a little bit more this year for documenting code. We saw that 51% are currently using AI for documenting code. Last year, we asked about it as well. We saw 33% mostly are partially using AI for documenting code. Then we also asked in a poll survey earlier in 2026 about what was affecting developers' time and what was sort of affecting developers' frustrations and some typical workflows that went along with that. Documentation was something that most spent the least amount of time on, but found to be one of the more frustrating parts of their job. RD: So I'd be remiss if I didn't talk about ourselves here. One of the stats I saw that was interesting is that 16% of folks say they're visiting Stack Overflow more to verify the AI output. Does that feel like a trend? EY: Yes, we do see a trend. Specifically, we asked last year about if anyone was visiting Stack Overflow specifically because they ran into a problem with AI. More people indicated that they were last year. 35% said that they did. So comparatively, we would see that if this was a trend, it's declining just year-over-year comparisons there. RD: Right. But there's still a need for verification. I think overall, you said there's trust issues except on code that they can easily verify. What are we verifying against? EY: Yeah. Just like we mentioned earlier, that's those attributable sources, the context that powers the AI engines to come up with the responses that they come up with and documentation. RD: I know we've here at Stack Overflow talked about getting context and documentation outside of AI agents. What did the survey say about how much time they spend looking for information? EY: Yeah. So we asked specifically about spending time searching for answers. 75% of developers said they spend five or more hours a week looking for information. A third of respondents also indicated more than that, five to nine hours. 20% indicated spending 10 to 14 hours a week. 10 to 14 hours a week. That's a day or two loss to just looking for information. Yeah. And we also asked about a similar question a while ago in the 2023 survey. And we didn't ask in hours and we didn't ask in weeks. We asked in minutes and days. So comparatively, we see that there might be a slightly less time spent searching for answers. We saw that 63% of respondents in 2023 spent anywhere between 30 and over 120 minutes a day searching for answers. RD: It still feels like a lot. EY: Feels like a lot. I think it might be slightly getting better, but it's still a lot. RD: Yeah. Is this some information that's just like, you know, they look for it, they find it, they get their thing done. Did we ask anything about that? EY: Well, we did ask about after you started working on a project, how often, yeah, if you needed context for that project, how often does it arrive after you've already started? So 79% of users indicated that they found important context after they already started. 43% said that sometimes that happens. Sometimes they start, they have everything they need. 12% said very often they found important context after they had started a project. RD: I know we asked about how people are doing context management. Is it something that people have tools for at this point? EY: I think they must have tools, but they might not be using them very regularly. We see that context is still pretty disorganized, as we would expect, given the little amount of time people are spending on documentation and the growing, slow-to-grow acceptance of using AI to help with documentation as well. 54% of users this year said they keep separate notes for important context. Whenever they start a new AI session, 45% of users said they also just start a fresh AI session to reset context. And then 35% said they manually go find files that they need every time that a specific question comes up that they know they have the answer to. In those files, 33% of users are copying and pasting in between tools. RD: Oh, wow. That sounds super frustrating. I know we asked about other frustrations with the context itself. What were people saying about that? EY: So top context problems, information quality. 63% said they struggle with incomplete information. 61% said knowledge is trapped in people's heads, so it's not even documented. They know they have to go to a person to get it. 57% say they find documentation that's outdated. And 53%, similar to the copying and pasting functions, are just their documented context is in separate tools. They have to go find them to bring them together. RD: Wow. So it sounds like people are spending five to 14 hours just searching for information, even with AI tools. Do you feel like AI tools are going to make the problem a little less bad? EY: I really do like the organized style that… AI prompting has forced us all to accommodate. And it goes right back to the issue of documentation. Documentation has always, I think, been sort of a, what do you call it? RD: It's been a chore for a lot. A lot of developers think it's a chore. EY: Yeah. I was thinking of whatever that, the cliche of something like the monster on your back. I don't even remember what it is. RD: Oh, yeah. EY: So it's, yeah, it's more than a chore. It's just like, I definitely don't have time for this, you know, amongst all of the other things I have to do. Spending the least amount of time on it, regularly finding it the most frustrating. So, but for... AI workflows everything becomes so much smoother if you have all these organized files. You organize your markdown files with that updated context, you have a method of finding out how something is outdated, if it has been the source it can you can attribute the source to a subject matter experts or that you have reasons to trust the source. Like that is the number one source for getting that information and that, yeah, you have easy ways of keeping it up to date and also, yeah, you're just keeping files updated as you work. So as things change, the files are changing too, so you have that regular refreshed context documentation organized file structure ready to go and then you're also able to use that whenever you need to trace back to: why did we make these decisions when I was deep in the flow state of work? And just making sure that everything was coming out correct. You also have the luxury of traceability with all of that documentation and that good context, which is also important. We see in some of the questions that we asked about in the survey. RD: Yeah, I know that traceability has always been important. But as a community that provides trusted knowledge, I know we asked about ourselves a little bit in here. There's been a fair amount of talk about the drops in traffic around the internet. What did we find about that sort of drop in traffic? EY: Yeah. So, well, we specifically asked about, are you visiting the site? If you're already visiting Stack Overflow, if you are, what are you doing on the site? 78% of users indicate that they're still coming to the site to read posts and discussions. Only 31% are coming to Stack Overflow to ask questions. 34% coming to Stack Overflow to answer questions. So I think while a majority of users are not taking part in those engaging activities, the passive activities of learning through reading through questions and answers and comments is still happening. RD: Yeah. I remember there's 1% who are asking and answering more because of AI. They're the real heroes. This is obviously the answer is AI, but why are people not visiting as much? They're not posting as much. EY: So 65% indicate that they are able to find answers without posting. So that might be coming from a Google search, might be coming from an agent. So whenever they're attributing sources, they're clicking on links and they're seeing what they need to see. Like, okay, I feel confident about whatever this AI assistant has told me is the right thing to do. 45% of users are using those AI tools to just get enough of the answer that they can move forward with. So no need to go do the extra verification, which saves a step. I totally understand. There's some things that I think a lot of us who visit Stack Overflow, we're going to visit some of the same problems. Like, I know I've gone to Stack Overflow to find this answer before. I just need you to remind me what it was. Here I am again asking the same question. And then besides that, 36% have also indicated they don't have the time to go to Stack Overflow to find, just sort of read through all of that, all of the context that goes along with the questions and answers. And then we also saw that 21% do not want to engage with posting because of fear of being judged or criticized. RD: Oh, all you big meanies out there, look at what you did. EY: Actually just pulled up a YouTube video that I was using for an example to do some content. And someone posted a great video about, it was titled the best question ever asked on Stack Overflow. Have you seen this? RD: I don't think I have. What is it? EY: Well, it's a very technical question. So I didn't get into the specifics of the technical question. But what this user was mostly concerned about was that it had so many views. It was an older post, older question, and it was closed. And he was just kind of going through all of the reasons why this was the wrong thing to do. There were some great discussions in there. I'll share that video with you. I think it's very interesting how much content you can find about Stack Overflow. RD: Yeah, everybody has their opinions. And if you've ever dipped into meta, you know that as much. So for the folks that are learning, did you find that they're learning about AI or general coding? EY: We see that users are still learning to code. We asked about if users are learning to code. We've asked this question in the survey quite a few years. So yes, they are still learning. For those that are learning to code or learning a new technology skill, we see the typical responses. They're using tech– technical documentation. We've asked this now for a couple of years. Are you using AI tools to learn to code? So yes, they are. 53% of users are now using AI tools to learn to code. We also specifically asked are you learning to code for AI, which is a little bit different. So of those, we see even more people, well, about the same amount, 51% are using AI tools to learn to code for AI. So letting the tool teach you how, what is the right way to use it might not be the best way to do it, but it is a way. So AI is being involved in the learning process either way. RD: Yeah. I mean, people are definitely learning from AI, but there's still a little bit of human there, right? There's a little bit of human in the loop. There's still direction from humans on what to learn. Why do you think AI hasn't replaced the human signals there? EY: Well, as far as the direction goes, when we asked in the survey, what resources do you use when you want to decide what it is that you want to learn next? What skill or coding language you'd like to learn next? So the top response for that, 44% indicate that they learn about what it is they want to learn from coworkers and peers. And then 33% from their managers. So there's a lot of work interaction going on there. RD: Yeah. Well, I know a lot of folks come to the survey to see what tech is hot, how loved is Rust at this point. So what's the, you know, start with the sort of languages. Who's topping the leaderboards there? EY: So this year, not a surprise, JavaScript is, again, the most used programming language now for the 12th straight year. 62% indicate having using it. We also see the typical leaders that we have been seeing for the past couple of years. Python is up there. 58% have indicated using it. HTML, CSS is up there as well. We also saw, yeah, speaking of most loved languages, most desired languages, Rust is back. And we see an increase in the usage this year too. They're up to 18% are now using Rust as per the survey. RD: There's a bunch of other technologies on there. Any changes in the database leaderboards? EY: Yes. We saw that Postgres is now the top database. I believe it was also the top database last year, too. RD: Yeah, they're pretty solidly number one. EY: Yeah, 58% have indicated using Postgres as a database in the past year. And we actually have featured on the survey this year a great time series that shows the databases over the years what has been the usage stats. So now you can see the diametrically opposed trends of MySQL and Postgres over the years back to 2018. It's very interesting. RD: Yeah. I mean, Postgres, I talked to a lot of folks on the podcast or repping Postgres all day. So how about in the other technology areas? What else did you see in there? EY: We asked about a new section this year, co-testing tools. For co-testing tools, we saw that ESLint was the number one tool that people are using. Only by 37%. So I guess co-testing is not for everybody. RD: That one's a linter. I think, I don't know how many people are still linting. I mean, I guess a lot, right? EY: A lot. I mean, compared to the other tools that we have here. We had, and I mean, considering, I know that ESLint is tied to the JavaScript Foundation. That makes sense to me that it's going along with JavaScript being the most used, that it would be the most used code testing tool. We also saw Postman and Newman on there right under ESLint for most usage. And PyTest, PyTest solid number three, again, with all of those Python users that we know that indicate they're using Python in the past year. RD: And do we have some AI technologies on this leaderboard? I think we had orchestration and such, right? The sort of things that developers would build with. EY: So for AI orchestration tools, top of the list is Langchain, 30% indicated having used it last year. Right under that, for the top three, we have number two OpenAI Agents SDK, which makes sense since we see that OpenAI's LLM is used just about as much as Anthropic’s. ChatGPT is still one of the top agent tools out there on the survey site as well. But number three for orchestration tools is Terminal. RD: Yeah. You don't mean just like a regular Terminal program? EY: Terminal. Yep. RD: All right. So the testing tools are new this year. Obviously, testing isn't new. Why did we decide to add a new section for testing? EY: Yeah, I wanted to add this section. I just thought it was interesting with, as we've seen in the survey results, that there's fluctuating trust and favorability with AI tools. We know that there's problems with almost but not quite right code responses from AI tools. It should go hand in hand that we're seeing more testing going on. I know that we also ask about testing with the AI itself. You can add that as part of your workflow or set up a separate agent. Ask the agent you're working with, hey, let's set up some tests and make sure that's part of our process here. I think that most people would want to use– my hypothesis is that people want to use something that's separate from AI to test code more now than ever to make sure that they are catching the things that AI is sort of overlooking because of whatever is causing hallucinations. They're suggesting these issues, so maybe they wouldn't be the best tools to catch the issues. And a lot of these co-testing tools are still part of open source projects that are maintained by individuals who want to make sure that they survive AI, they survive whatever changes happen with these new emerging technologies. So I definitely wanted to make sure that was covered in here so we could see maybe a little bit of that diversity of some of the non-AI tools people are using. RD: So if you had to pull out one stat or chain of stats that you think is the big takeaway for you, what would it be? EY: Yes, everything and anything to do with documentation. We have quite a few stats in the survey this year about documentation that reference documentation and context. Also, all of the stats that we have that reference attributable sources and how to find quality checks amongst your sources. I think that just proves a whole point about where the problems come from. It's not necessarily that AI itself is causing the problem. It might be something more foundational that we've always had as a problem, which is documentation. RD: That's right. Well, as a former technical writer, I love to hear the greater focus on documentation, you know? EY: Yeah. And I mean, technical documentation is still the number one tool for learning how to code. RD: There you go. [music] Well, it's that time of the show where we shout out somebody who came on to Stack Overflow, dropped some knowledge, shared some curiosity, and earned themselves a badge. Today, in honor of us asking questions in the survey, we are shouting out the winner of a stellar question badge. Somebody who asked a question that was saved by 100 users. And today, shouting out Frank Freeburger, who asked, biggest integer that can be stored in a double? If you're curious about that, we'll have the answer for you in the show notes. I'm Ryan Donovan. I host the podcast, edit the blog here at Stack Overflow. If you have questions, concerns, topics to cover, et cetera, please email me at podcast@stackoverflow.com. And if you want to reach out to me directly, you can find me on LinkedIn. EY: My name is Erin Yepez. I'm market research and insights at Stack Overflow. I would say the same thing. You can find me posting about surveys on the Stack Overflow blog or on LinkedIn. RD:And where can folks check out the full survey results? EY: Yes. So the full survey results will be on our survey website. That's where all of our survey results are. So if you go to insights.stackoverflow.co, C-O, not C-O-M, you'll see every single year's survey results and 2026 will be there. Just click on that. RD: All right. Thank you for listening, everyone. And we'll talk to you next time. [outro music]
Check out the full 2026 Developer Survey results. Connect with Erin on LinkedIn and read her past articles on the Stack Overflow blog. Congrats to Stellar Question badge winner Franck Freiburger for winning the badge on their question Biggest integer that can be stored in a double. TRANSCRIPT [intro music] Ryan Donovan: Hello, and welcome to the Stack Overflow podcast, a place to talk all things software and technology. My name is Ryan Donovan, and today we are talking about the results of the 2026 developer survey. My guest for that is Erin Yepez. She's Market Research and Insights here at Stack Overflow. So welcome back to the show, Erin. Erin Yepez: Thank you for having me. RD: Of course, of course. So this is the 16th developer survey we've done, and I think it's the third one with AI on it. And this year sort of feels like the year that everybody's sort of standardizing using AI. I keep hearing people are not writing code anymore. What did we learn from the survey about AI this year? EY: So it's actually the fourth year that we've had AI on the survey. And we have some... continuity questions. In all of our surveys, we've asked about usage in general of AI tools. This year and last year, we specifically also asked about using AI agents. We also asked about how people are liking them. Do they find AI tools in their workflow to be favorable? We asked about what they're using them for specifically at work. This year, we asked quite a few new questions as well, just in terms of where do you find impact? We kind of tweaked some of the wording from last year. To be a little bit more specific, we asked also about whether anyone is avoiding using AI tools and why. And also, if anyone's building software with AI tools, what are they building? And if they're switching AI tools, what caused the switch? Things like that. Overall, though, I mean, we've seen compared to last year, usage still pretty high. We're at 83%. It's actually just about the same as last year for usage. But we added on to that. So now we specifically also know frequency of usage. Are you using AI tools at work daily, weekly, monthly or less? And then we also found out whether everyone's using AI tools daily, how many hours a day. Some great insights there. RD: Yeah. So you said 83% of developers use AI at work. How many of them are using it every day and how much every day? EY: All right. So overall, we have 73% of users indicate they're using AI tools daily for a variety of reasons, mostly using them for AI coding assistance or general purpose AI chatbots. RD: And, you know, you mentioned AI agents. How many people are using AI agents? EY: So we show when we break it up, 66% of respondents indicated using AI agents, 63% general AI chatbots. RD: So some of what we do is kind of technology leaderboards and the AI section has its own kind of technology leaderboards. Any surprises on who's top of the LLM or the AI agent leaderboards? EY: Surprises? I would say no. I think we still see the frontier companies have most market share, the most usage amongst our respondents. We asked about a lot of different companies though. So in particular, I know I was interested to see how some of the open weights companies were doing. I looked at some comparisons for, for year over year values as well. They're staying steady. So they're not decreasing, not really increasing either. The big one being DeepSeek that I'm thinking of. Which I know DeepSeek is not completely open weights now, but that's kind of the first one most people think of when you mention open weights. RD: Right. Yeah. I thought it was interesting that, you know, the top two, the Anthropic and OpenAM models are both 74%. So there's a lot of crossover there. A lot of people using both. EY: Yes. RD: So I think we've heard a lot about kind of tokenomics question. Are people, you know, is the data reflecting that? Are people switching? Are people you know, adjusting how they behave based on token costs? EY: No, we asked about specifically if in our question about switching AI tools, you know, if you have switched using tools in the past year, what made you switch? And people indicated that cost was part of that reason, but it was one of the lower ranked reasons. So only 6% indicated switching because of costs. We also asked another question specifically about tokens and if people limited themselves personally while using AI tools due to costs at work. And 33% indicated that they self-monitor their own usage, but we found out that 75% of people are changing their prompts for quality reasons. And then only 38% of users have indicated that they're changing their prompts due to cost. RD: Yeah. So we have a question about whether AI is making everybody productive, right? Like, is there some magic bullet tool? We looked at how many AI tools people use. Is it narrowing how many tools they're using? EY: Well, we haven't asked about how many AI tools people are using before in the dev survey. So we don't have anything to compare it to. We did ask last year how many software tools overall people are using. I know that was pretty high. On average, it was eight. And for more early career respondents, that number, the average number was higher than eight. So for most users in this year's survey, we did specifically ask about using AI tools. And we saw that upwards of 60% were using between two and five. RD: It's pretty clear that there's this pretty broad AI adoption right now. But every time I go to a conference, there's half the companies are trust and context and governance. Do people trust AI any more this year? EY: I think that there's a lot of conditions to it. So we asked a few different ways about how individuals decide that they trust the tools that they're using. Specifically, we asked, how much do you trust the output from AI tools or agents as part of your workflow? It's slightly a different variation of the trust question that we've asked in the past. In the past, we've asked just in a typical Likert sense, how much do you trust the accuracy of the AI tools you use at work? I trust it very much. I sort of trust it. Neutral. I don't trust it. I definitely don't trust it. So some variation of that, but this year we were a little bit more specific asking how much do you trust the output? And so select one of the following, uh, when I can verify it was what most people selected. So 48%, just about half of users trust AI tools when they can verify it. And opposed to that, we see that 7% have trust in AI tools for many tasks and decisions. RD: Right. For all the important stuff, they'll give it to AI. EY: Yeah. Yeah. So not very many people, but there were some people that indicated that. 1% indicated that they weren't sure or that they didn't use AI. So that's always in there. Even though that question was a conditional question in the survey, it wouldn't have showed up if you already indicated you don't use AI. RD: Right. One of the early kind of paradigms for AI was the retrieval augmented generation to sort of get some sense of information source. Is that something that's still important today? EY: Yeah, so 79% indicated that source attribution was either important or very important when evaluating an AI-generated answer. RD: And, you know, the RAG was sort of like to eliminate or mitigate the hallucinations that the AI was doing. Is there still a question about that? I mean, I know we asked about what were the sort of top pitfalls. EY: Yeah. Actually, we asked about, in the past year, has your company experienced a disruption due to AI? And if so, what would you indicate was the one that cost the most time slash money? And the top reason for that AI disruption was plausible but wrong code, which tracks with some of the survey results that we got last year, too, we asked the question in a different way, but we saw, again, top two responses as far as what caused people pause when using AI tools or what was giving them friction at work was almost right code. And then second was security concerns. RD: You know, I know one of the primary use cases of it is generating code. And it feels like the AI-generated code is putting a lot more pressure on code review. I think there was a fantasy or a dream that AI-assisted coding could be complete without human validation, but that doesn't seem to be quite the case. EY: Yeah, I would say, if anything, one of the bigger themes that I found in this survey, which tracks with some of the other research that we've done, is that In order to get ahead of those hallucinations and get ahead of the problems that come with your agent or coding assistants not being able to get fully correct code is the documentation, the context that comes with documentation. We see that coming up a lot in the survey. We have similarly asked about documentation in last year's survey too. So we can see that people are using AI a little bit more this year for documenting code. We saw that 51% are currently using AI for documenting code. Last year, we asked about it as well. We saw 33% mostly are partially using AI for documenting code. Then we also asked in a poll survey earlier in 2026 about what was affecting developers' time and what was sort of affecting developers' frustrations and some typical workflows that went along with that. Documentation was something that most spent the least amount of time on, but found to be one of the more frustrating parts of their job. RD: So I'd be remiss if I didn't talk about ourselves here. One of the stats I saw that was interesting is that 16% of folks say they're visiting Stack Overflow more to verify the AI output. Does that feel like a trend? EY: Yes, we do see a trend. Specifically, we asked last year about if anyone was visiting Stack Overflow specifically because they ran into a problem with AI. More people indicated that they were last year. 35% said that they did. So comparatively, we would see that if this was a trend, it's declining just year-over-year comparisons there. RD: Right. But there's still a need for verification. I think overall, you said there's trust issues except on code that they can easily verify. What are we verifying against? EY: Yeah. Just like we mentioned earlier, that's those attributable sources, the context that powers the AI engines to come up with the responses that they come up with and documentation. RD: I know we've here at Stack Overflow talked about getting context and documentation outside of AI agents. What did the survey say about how much time they spend looking for information? EY: Yeah. So we asked specifically about spending time searching for answers. 75% of developers said they spend five or more hours a week looking for information. A third of respondents also indicated more than that, five to nine hours. 20% indicated spending 10 to 14 hours a week. 10 to 14 hours a week. That's a day or two loss to just looking for information. Yeah. And we also asked about a similar question a while ago in the 2023 survey. And we didn't ask in hours and we didn't ask in weeks. We asked in minutes and days. So comparatively, we see that there might be a slightly less time spent searching for answers. We saw that 63% of respondents in 2023 spent anywhere between 30 and over 120 minutes a day searching for answers. RD: It still feels like a lot. EY: Feels like a lot. I think it might be slightly getting better, but it's still a lot. RD: Yeah. Is this some information that's just like, you know, they look for it, they find it, they get their thing done. Did we ask anything about that? EY: Well, we did ask about after you started working on a project, how often, yeah, if you needed context for that project, how often does it arrive after you've already started? So 79% of users indicated that they found important context after they already started. 43% said that sometimes that happens. Sometimes they start, they have everything they need. 12% said very often they found important context
document fetched 2026-10-24 10:03 · canonical https://stackoverflow.blog/2026/10/06/tales-from-the-2026-developer-survey-results