Stack Overflow developer survey AI

Stack Overflow developer survey AI

Here’s a weird thing about software development right now. Ask a room full of programmers whether they use AI tools, and most hands go up. Ask the same room whether they trust what those tools produce, and the hands drop fast. Somebody laughs. Somebody shakes their head.

That awkward split is the whole story of Stack Overflow’s 2025 Developer Survey, which is the most recent full set of results available as I write this. More than 49,000 developers from 177 countries filled it out. If you’ve been trying to figure out whether AI coding tools are a revolution or a headache, the honest answer from the people who actually use them is: both, at the same time.

I’ve spent a good while with the results and the follow-up analysis Stack Overflow has published since. I want to walk you through what stood out, and I’m skipping the hype version and the doom version. This is just what working developers said when someone asked them straight.

The adoption numbers are huge

Let’s start with the part nobody argues about. AI tools are everywhere in programming now.

According to the survey, 84% of respondents said they’re using or planning to use AI tools in their development process. A year earlier that number was 76%. Among professional developers, a little over half said they use these tools every single day. Not “tried it once.” Every day.

Stack Overflow’s own look-back at the data shows how fast this happened. In 2023, 44% of developers said they were using AI tools. That climbed to 62% in 2024 and about 79% in 2025. Think about that for a second. Three years ago, a lot of developers were still rolling their eyes at the idea. Now it’s closer to the default setting.

The tools people reach for tell you something too. ChatGPT is the most-used by a wide margin, and GitHub Copilot sits at around 68% among people who use AI tools at all. Those two are basically the front door. Most developers start with a chat window or an autocomplete in their editor, and then they branch out.

Students are even more into it. Among student respondents, AI use went from 63% in 2024 to 73.3% in 2025. If you’re wondering what the next wave of developers looks like, they’re growing up with an assistant in the room.

So adoption isn’t the question anymore. People are using it. The interesting part is how they feel about it.

Where the trust went

This is the number that made me stop scrolling. Roughly 46% of developers said they actively distrust the accuracy of AI tool output. Only around a third said they trust it. And just 3% said they “highly trust” it.

Three percent. Out of tens of thousands of people who use these tools constantly.

The mood has shifted over time, too. In earlier years, favorable feelings toward AI tools ran above 70%. In the latest data that’s slid to around 60%. That’s still a majority, so nobody’s calling it a revolt. But the direction is clear. The more people use these tools, the more clearly they see the cracks.

I think that makes complete sense, and it’s a pattern you’ve probably seen with other technology. The first time a new tool wows you, you overlook the flaws. Six months later, after it’s confidently handed you the wrong answer a dozen times, you start checking its work. That’s not hostility. That’s experience.

It also helps to separate two ideas that often get mashed together: “I use it” and “I trust it.” They aren’t the same. You can use a calculator you don’t fully trust, as long as you know what to double-check. Developers seem to be treating AI tools the same way, as a fast, occasionally unreliable coworker whose work always needs a second look.

So why keep using something you don’t trust?

Fair question. If nearly half of developers distrust the output, why is daily use so high?

Because distrust and usefulness can live in the same tool. Think about a very fast intern. They can draft a function in ten seconds. They can write a test file while you get coffee. They can explain an error message you’ve been staring at for twenty minutes. They’re also wrong a meaningful chunk of the time, and they say wrong things in the same confident tone as right things. You’d still keep that intern around. You’d just never merge their work without reading it.

That’s pretty much where developers have landed. The speed is real. The first draft really is faster. What’s less certain is whether the finished, reviewed, tested result is faster, and I’ll come back to that in a minute.

There’s also pressure from above. A lot of companies are rolling out AI tools and measuring adoption. When your manager’s dashboard counts how often you use the assistant, you use the assistant. Some of that high usage number is people who were told to use it, not people who fell in love with it. The survey can’t fully separate those two groups, so it’s worth keeping in mind.

When AI isn’t enough, people turn to people

One of my favorite findings in the whole survey is also the most human one. When developers don’t trust an AI answer, about three quarters of them said they’d ask another person. A coworker. A friend. A forum.

Another chunk, around 35%, said they visit Stack Overflow specifically after running into trouble with AI-generated code. Yes, the company running the survey benefits from that stat, so take it with the right amount of salt. But it matches what a lot of developers say in their own words: AI gets them 80% of the way, and then they need a real person who’s actually hit this exact problem.

That’s a reassuring thing, honestly. For all the talk about machines replacing programmers, the thing developers reach for when the machine fails is another human who has seen the problem before. Human knowledge, written down by humans, still sits underneath the whole thing. The AI tools learned from it, and developers still go back to it when the AI gets it wrong.

The beginner problem

Here’s where I get a little nervous. Remember those students and early-career developers using AI at record rates?

An experienced developer can look at generated code and feel a little itch. Something’s off about that loop. That variable name hides a bug. This approach will fall over when traffic spikes. That itch comes from years of seeing things break. A beginner doesn’t have it yet.

So the people most likely to lean hard on AI are the people least equipped to catch its mistakes. That’s not a knock on beginners. It’s just how learning works. You can’t spot a bad answer if you’ve never built the skill to tell the difference.

If you’re new to coding, I’d suggest a simple habit. Before you accept any AI-generated code, try to explain to yourself, line by line, what it does. If you can’t, don’t paste it in yet. Ask the tool to explain it, then ask a follow-up, then try to rewrite a piece of it from scratch. It’s slower. It’s also how you end up with skills that last longer than a subscription.

What about AI agents?

Agents are the newer, flashier thing. Instead of suggesting a line of code, an agent takes a goal and goes off to do several steps on its own: edit files, run commands, fix errors, try again.

The survey found that adoption here is a lot more cautious. A bit over half of developers said they either don’t use agents at all or stick to simpler AI tools. About 38% said they have no plans to adopt agents. Among the folks who do use them, about 69% agreed they’ve increased their productivity, which is a strong number for the group that’s actually tried it.

It’s a classic early-adopter shape. The people using agents like them. A big share of everyone else is waiting to see what happens. Stack Overflow’s recent notes on its 2026 survey say agent usage has roughly doubled since then, but that concerns about AI-generated code have grown as well. That’s an interesting combination. More use, more worry. Sound familiar?

What this means if you write code for a living

A few practical takeaways, and none of them are glamorous.

First, treat AI output like a pull request from someone you’ve never worked with. Read it. Run it. Test it. Don’t assume.

Second, get good at the part AI is bad at. Debugging, reading unfamiliar code, understanding how a whole system fits together, knowing why a decision was made. Those skills matter more when someone, or something, is generating code faster than you can read it.

Third, keep your human network alive. The survey is pretty clear that people still turn to other people. Don’t let your best problem-solving partners drift away just because a chatbot is always available at 2 a.m.

And fourth, be honest with your team about where AI helps and where it doesn’t. A lot of the trust gap comes from people quietly cleaning up AI mistakes while leadership assumes everything’s smooth.

What this means if you manage developers

Don’t read “84% adoption” as “84% happier and faster.” The same survey says the top frustration is output that’s almost right but not quite. Fast drafts can create slow reviews.

Measure outcomes, not logins. How long does it take to ship a finished, tested feature? How many bugs come back? How do your newer developers do on their own? Those answers matter more than how many people clicked the assistant this week.

And give people permission to say “I didn’t use AI for this because it wasn’t helping.” You’ll learn more from that sentence than from any usage dashboard.

What’s coming next

Stack Overflow developer survey AI

Stack Overflow opened its 2026 Developer Survey in June, and as of early October the results aren’t out yet. The questions this year are interesting. They’re asking where AI fits across the whole software development lifecycle, and whether developers are moving from “human in the loop” to “human on the loop,” overseeing a bot that does most of the driving. They’re also asking whether the industry is moving from experimenting for the sake of it to actually measuring return on investment.

My guess? Adoption will tick up again. Trust will stay stubbornly low, or at least stay complicated. And the most interesting numbers will be about how work is changing, who’s getting hired, and what junior developers are learning.

When those results land, I’ll be reading them closely. If you’re a developer, a student, or someone thinking about hiring one, you should too.

The bottom line

Developers aren’t ignoring AI, and they aren’t worshipping it. They’re doing something more grown-up. They’re using it every day, double-checking it constantly, and going to other humans when it falls short.

That’s not a failure of the technology. It’s a pretty healthy relationship with a powerful, imperfect tool.

If there’s one thing I’d take from the survey, it’s this: the winners won’t be the people who use AI the most. They’ll be the people who know exactly when to believe it and when to push back.

So where do you land? Are you in the 3% who highly trust it, the 46% who don’t, or somewhere in the messy middle with everyone else?