You might have seen the State of JavaScript survey before, but did you know there's now a State of Web Dev AI survey as well?
I just published the results, and the data revealed quite a few interesting trends.
The AI Takeover is Here
A few years ago, NFTs (remember them?) were all you'd hear about. But that technology had very little practical use, and once the hype wore out it wasn't long before the world more or less forgot about it.
It seemed at first like AI might follow the same path, with many dismissing LLMs as "fancy autocorrect" and predicting an eventual return to the statu quo.
But it should be clear by now that AI is anything but a flash in the pan, and the survey data confirms it.
For example, the proportion of code generated by AI tools has jumped from 28% in 2025 to 54% this year:
It's especially striking to see that the segment of respondents who use AI to produce nearly all of their code is now the single largest bucket–when it was the second smallest last year!
The frequency at which developers reach for AI tools has also increased drastically, with developers reporting they use AI “constantly” going from 11% to 21%.
All those signs point in the same direction: not only is AI usage increasing, but it's doing so at a speed rarely seen before when it comes to adopting new workflows.
The Claude Code Connection
So what's behind this shift? While it's impossible to prove any causal link with survey data alone, one factor worth considering is the growing impact of Claude Code, and agentic coding in general.
Claude Code is the most-loved coding assistant:
And in terms of raw usage, Claude Code is second behind GitHub Copilot with 62.9% of respondents having used it–while OpenAi Codex comes in a distant third with only 34.5%.
But more crucially, both Claude and Claude Code top the ranks of tools developers are actually paying for:
Anthropic likes to position itself as the underdog fighting against OpenAI's dominance, but these results indicate that when it comes to developers at least, OpenAI is the one fighting an uphill battle.
A Threat to Developers
So it looks like developers are embracing their new AI overlords en masse, and the few lone voices warning us about Skynet can probably be dismissed as delusional crackpots, right?
Well… not so fast. While it's true that developers tend to be more concerned with the next line of codes than with doomsday predictions, it doesn't mean they're blind to the very real issues created by this sudden AI surge.
For starters, developers are quite conscious of the fact that their jobs may now be threatened by the very machines they helped train, as is becoming apparent in the recent Meta layoffs:
And as a commenter astutely pointed out, even if AI can't do your job, all you need to be fired is that your boss thinks it can.
But while job loss tops the list of AI issues developers worry about, it isn't the only one by any means:
Military use of AI scored quite high as well, which makes sense since the survey was filled out at a time when the Pentagon's use of AI was in the news.
And of course, AI's environmental impact is another very real worry, with the construction and operation of new datacenters putting extra stress on an already-struggling planet.
Preparing for an AI Future
Taken together, these survey results point a nuanced picture: yes, developers have embraced AI and its productivity gains–but they're also quite aware of the risks and issues presented by AI adoption, and many of them aren't quite convinced that the juice is worth the squeeze.
Believe it or not, this is only a tiny fraction of the data collected, and I encourage you to check out the full survey results to get a broader picture of the AI developer ecosystem in 2026, as well as read the conclusion by the one and only Primeagen.
And hey–if you'd like to brush up on your CSS knowledge one last time before Claude Code takes your job, why not go fill out this year's State of CSS survey and see how many new CSS features you know?
Top comments (38)
The data from the State of Web Dev AI survey really highlights the shift in developer sentiment. We've moved past the "AI will take my job" fear and straight into the practical concerns of unmaintainable, AI-generated technical debt. When it's easier to generate a new feature than to read the code running it, we have a serious maintainability crisis on our hands.
I find it a bit concerning how few people are seeing the angle that the ability to build software is moving from someone you can learn to something you have to buy. Maintainability concerns are valid, but the broader cultural effects are relevant on a much larger scale.
I read that the public sentiment on AI is still on a nosedive. This is more apparent on new college graduates. Let's see how they handle it these coming years.
I suspect many of the "anti-AI" college graduates still use ChatGPT to help with assignments…
These numbers are truly astounding 😵 Especially the proportion of code generated by AI tools and potential threats give some food for thought. Call me naive, but I never considered military use of AI as a threat... I don't even want to think about how much damage AI could do in the wrong hands of some powerful officials 🫣
Seriously? It's already happening in Gaza and West Asia. Wrong officials already have access to AI enhanced weapons... Just Europe isn't the target (yet).
How exactly is AI used in these cases?
The risk angle I'd add: vendor concentration. Most production LLM workloads in 2026 hit just 2-3 APIs (OpenAI, Anthropic, Google) — if any of them tightens pricing or terms mid-year, half the SaaS layer built on them collapses overnight. The mitigation isn't building local infrastructure (too expensive for most teams), it's keeping wire compatibility across providers so you can swap in a day, not a quarter. Underdiscussed compared to AGI doom takes.
If the AI bubble bursts, they'll all increase prices at the same time in an attempt to prove profitability. There's no way to mitigate that other than to simply be prepared to pay more or switch back to non-AI solutions where it's no longer profitable to use AI.
Fair point on the bubble scenario — if everyone hikes together, compatibility doesn't save you on absolute cost. The case for portability is more about non-price risk: sudden ToS changes, deprecations, geographic restrictions, and rate-limit policy moves which are per-vendor and don't track macro pricing. You can't escape the bubble, but you can escape any single vendor's bad decision.
My AI fear is the pandemic of "Cognitive Atrophy" that is about to hit those whose job are not taken by AI but impacted by +50%. We could well see humans losing the ability to think for themselves let alone reason and problem solve.
For me as a software developer, I am concerned that AI (LLMs) will take the part of my job I find most rewarding, leaving me with the parts I hate. When than happens I will be looking for a career change.
They used to say, they don't pay software developers to code (but they will pay tech bros!) but to think. That is not quite true. They should be paying software developers to grain an understanding of the problem and formulate a compatible solution. I am not sure they can say that about the LLMs! Their only real skill is pattern matching.
The same problem is mirrored on the larger scale too: As the workforce gets used to AI solving most problems, it forgets how to build things from scratch. Junior developers no longer turn into senior developers, and eventually, too few people are left to generate new original code to train AI on and things get ugly.
I feel the same way. The implications of a society that cannot think for itself are dark. It makes us easier to manipulate and control.
curious what the survey split looks like between teams that have any written constraints on what agents can do vs teams running agents purely on vibes. that gap feels like the actual risk, not AI capability itself.
The jump to 54% of all code being AI-generated in the survey data is wild, but the psychological shift is the most concerning part.
The 28%→54% jump is the stat everyone will quote, but the concerning part isn't the volume — it's that review didn't scale with it. We benchmarked it: 63% of AI-generated functions had a security finding, and the misses cluster in the "negative space" a feature prompt never mentions — missing auth, no input validation, unpinned crypto. Frontier models avoid the catastrophic stuff (injection, eval) but consistently skip the hardening. So the real 2026 risk isn't "AI writes insecure code" — it's "AI writes plausibly-secure code fast enough that nobody re-checks what it silently skipped." The AI-code-balance chart is the one I'll be citing.
Glad to hear that, you can message me thru LinkedIn 🙂
Spot on. The most immediate risk of 2026 isn't evil AI; it’s a silent, compounding degradation of data integrity. When models train on synthetic data, and agents ingest unstructured, ambient web text, the signal-to-noise ratio plummets.
This is exactly why architectural patterns are shifting toward strict zero-trust data boundaries. If we aren't aggressively filtering and validating input footprints before they hit an execution runtime, we're just subsidizing the entropy. Thanks for highlighting the real battleground.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.