Let’s play a fast recreation: I’ll present a graph and attempt to guess what it’s about.
No, it isn’t a crypto coin crashing just a few hours after being minted. And never, additionally it is not an oscillatory/wavy graph made with pure CSS, however a harsher fact.
I already gave it away with the title, but it surely nonetheless hits like a ton of bricks to know it’s the steep decline within the variety of questions requested on Stack Overflow. You may see its peak round 2014 with greater than 200,000 questions requested in a single month. However now in 2026, it’s struggling to even hit 3,000 questions a month.
We don’t need to be specialists within the subject to search out out the perpetrator. You guessed, it’s AI… principally.
Whereas AI is painted because the Stack Overflow killer, the reality is Stack Overflow’s downfall began lengthy earlier than ChatGPT’s launch in late 2022.
By neighborhood accounts and likewise from private expertise, moderation since its peak in 2014 has been (and nonetheless is) one of many main causes for the shortage of questions.
As the positioning grew, Stack Overflow wanted a greater option to reasonable the a whole lot of 1000’s of questions requested each month: the inevitable wall that forum-based communities hit once they scale past a sure level. There are a number of methods to attempt to resolve this, however the route Stack Overflow took won’t have been the very best:
On Stack Overflow, we shut or delete questions that may’t be answered right away – it’s not very sociable, but it surely scales splendidly.
It’s clear Stack Overflow wasn’t specializing in the amount of the questions however somewhat on the standard of them, whereas avoiding duplicates as a lot as doable. This sample was in favor of Google searches for questions that had been already answered and, therefore, dwelling on pre-answered questions as an alternative of on customers making new or duplicate ones.
It wasn’t useful both how the neighborhood appeared to shut upon itself, making it more durable for novices to even ask a query. And should you’re like me, you most likely wish to inquire with out being advised you’re silly, as if getting punished for eager to be taught.
Generative AI was the ultimate nail within the coffin. I can’t complain about this, as AI seemingly gives the identical solutions with out judgment (actually, perhaps an excessive amount of encouragement) nor delay, so I can see why folks may choose asking an LLM as an alternative.
Nonetheless, as I dug deeper into this, my concern was not about simply Stack Overflow, however the tech ecosystem at giant. Questions like, are we nonetheless asking questions? Are we nonetheless in search of to be higher? Or will we all depend on LLMs, and solely on LLMs, for recommendation? That stored ringing in my thoughts as I continued my analysis.
I imagine that, past the autumn of Stack Overflow, these questions linger greater than ever. How AI has usually impacted our workflow, how we will use it in problem-solving, and what we will do about this as builders.
Drawback-Fixing and AI
Is AI a greater programmer than you? What makes a programmer higher than others is as subjective because it will get, however some are desperate to say that AI can write code higher than you. In accordance with that analysis:
AlphaCode achieves human-level downside fixing expertise and code writing capacity as proven by efficiency in programming competitions.
At the very least that’s when it was examined in opposition to Codeforce’s (an internet code competitors website) issues, the place I admit it might and can carry out higher than your common programmer. However most builders don’t care about Contest issues past a technical interview; they know being a software program developer is a lot greater than that.
AI writing high quality code is an especially nuanced matter and lacks a decisive conclusion. Nonetheless, should you take the time to analysis, you’ll discover that AI-generated code has plenty of flagrant variations. In accordance with the analysis from Cornell:
AI-generated code is mostly less complicated and extra repetitive, but extra vulnerable to unused constructs and hardcoded debugging, whereas human-written code displays higher structural complexity and the next focus of maintainability points.
Okay, so it might generate easy code, however can it write good code? Even resolve issues higher than a software program engineer would?
In accordance with MIT analysis, AI can write good code, but it surely can’t probably assume and make choices like a software program engineer. AI can’t compete on that degree but, at the least with out operating into a variety of bugs.
Drawing on each first-hand expertise and suggestions, if all you do is copy-and-paste AI-generated code with out cautious consideration, you might be sure to hit critical bugs and probably even vulnerabilities. Actually, VeraCode printed an article stating that “[…] 45% of AI-generated code accommodates safety flaws,” after testing for safety vulnerabilities in 100 AI fashions. That’s a big share of code that’s flawed security-wise and would have price implications for any consumer who needs to “vibe-code” with out doing thorough checks.
Enjoyable truth: GitHub launched the outcomes of its AI in software program improvement survey in August 2024, and over 97% of its respondents have used AI outdoors or inside their work. That’s even except for the businesses implementing using AI in your present code workflow. It’s actually all over the place; there’s virtually no escaping its utilization
However, does that imply it’s all dangerous? The reply to that, in my view, isn’t any. In accordance with analysis carried out by Harvard Enterprise Evaluate, AI is efficient for serving to resolve issues (let’s not additionally ignore the trade-off from the examine that AI workflows lead to much less motivation). In essence, it’s maybe finest used to reinforce problem-solving effectiveness.
Which means that, as AI is taking up industries and being included into our every day work, it nonetheless gained’t exchange your creativity and problem-solving strategy, which you’d must deal with distinctive on a regular basis challenges. It’s tough to copy.
Like each different software, AI has its limits, and with out human craftsmanship behind it, the software is sort of ineffective. A superb craftsman makes use of all of the instruments at his disposal to realize his objectives, AI being simply one in every of them.
“The effectiveness of the software is set by the ability of the craftsman who created it and the ingenuity with which he makes use of it.”
Craig D. Lounsbrough
The large hazard is not only safety vulnerabilities, however over-dependence on the software, which I imagine will result in an eventual decline within the variety of code craftsmen within the coming technology. How ought to newer and skilled builders go about this?
Some Recommendation
Here’s a checklist of questions I ask myself when choosing up AI in my improvement work:
- Am I asking the LLM smaller, particular questions? This fashion, I can confirm every course of step-by-step somewhat than eyeballing the entire system code as a complete. I’m nonetheless a developer within the sense that I’m not leaving the LLM to do all the work.
- Am I evaluating the output when it’s completed? In different phrases, do I perceive what it did? Would I be comfy modifying the generated code if I do know a greater strategy, or when I’ve to keep up it sooner or later?
- Am I checking the software’s references? This can be extra geared in the direction of analysis as an alternative of straight code output. The place precisely are its solutions coming from? Are these good sources? Are there others? It’s essential to know the software just isn’t citing a fictional supply, however somewhat, arising with fashionable and tried-and-true approaches.
- Have I examined the work? Did the software perceive the duty and take into account all edge circumstances? That is maybe an important query as a result of figuring out how folks use your software is one thing a machine is much less inclined to know than a human.
What occurs after we cease asking?
Take into consideration this: if we cease asking questions, how will AI be educated sooner or later? Applied sciences change and enhance over time. What’s up to date now will quickly change into old school. Take CSS, for instance. With the latest CSS updates (nesting, view transitions, container queries, and so on.), we’re writing CSS vastly totally different than even just a few brief years in the past. You wouldn’t wish to be caught with an outdated and clumsy resolution educated from code written many years in the past. If we cease asking questions and answering them, don’t you assume that may make the LLMs lag behind? That’s simply me speculating, however I feel it’s straightforward to think about that being the case.
We can’t deny Stack Overflow’s service over time. It bought us asking. It bought us answering. It bought us pondering. The query we must always all ask ourselves is,Will LLMs do the identical?
I’ll go away you with this quote from Stack Overflow co-founder Jeff Atwood:
Stack Overflow is you. That is the scary half, the nice leap of religion that Stack Overflow is based on: trusting your fellow programmers. The programmers who select to take part in Stack Overflow are the “secret sauce” that makes it work.
