DEV Community

Cover image for Why AI Can't Replace Developers: The Real Limits of AI in Coding
Michael Amachree
Michael Amachree

Posted on

Why AI Can't Replace Developers: The Real Limits of AI in Coding

AI Can't Write Better Code Than Devs

As artificial intelligence (AI) continues to evolve, it's becoming an increasingly popular tool among developers. However, while AI-powered coding tools like GitHub Copilot and ChatGPT offer convenience and efficiency, there's an important truth that every developer should keep in mind: AI can't write better code than you.

In this article, we’ll explore why AI can’t replace developers, why it’s still an invaluable assistant, and how we can use AI to enhance our coding workflow. Drawing inspiration from recent research, including Apple's GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models, we’ll dive into the technical limitations of AI when it comes to complex coding tasks.


The Limits of AI in Coding: Why AI Can't Replace Developers

AI has made significant strides in automating many aspects of software development. Whether it's generating boilerplate code, assisting with documentation, or offering solutions to common errors, AI can save developers time and energy. However, it's crucial to understand that AI is still far from perfect and cannot handle everything a developer can do.

One of the main issues with AI is its inability to reason through complex problems. AI tools can handle simple tasks with ease, but when it comes to more complicated issues that require understanding of a project’s specific context, human developers are still far superior.

I’ve worked with various AI-driven code assistants like GitHub Copilot, Gemini Code Assist, Codeium, and others. While these tools are useful for many tasks, I’ve encountered numerous situations where the AI simply doesn’t meet expectations. For example, when a developer asks AI to fix a problem in code, the issue often spirals into a bigger problem. Debugging AI-generated code is also much more difficult than dealing with code written by a human who understands the project’s nuances.

When AI gets things wrong, the results can be catastrophic. AI struggles to learn from its mistakes in the same way humans can, often failing to take into account the broader context or even the smallest details that matter in a codebase.

Moreover, AI tools are still severely limited in their ability to understand and reason. While they can provide answers based on data they've been trained on, their ability to handle complex tasks or innovate is still lacking. This is something many developers fail to realize, especially when they expect AI to act like a senior developer or creative partner.


The GSM-Symbolic Paper: Highlighting the Limitations of AI Reasoning

In the GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models paper, Apple researchers shed light on the shortcomings of AI’s reasoning capabilities. They show that while AI models can process patterns and provide answers based on their training data, they still struggle with true reasoning. AI doesn’t understand the context of a problem—it merely predicts based on patterns it has seen before.

The paper highlights that despite advancements in large language models (LLMs), these models may never achieve true reasoning as humans do. For example, when given simple mathematical tasks, LLMs can be extremely accurate. However, when minor changes are made to the problem (e.g., changing names, numbers, or adding irrelevant details), the models' accuracy drops significantly.

This is especially true when we consider coding tasks. AI can assist with simple tasks like generating boilerplate code or offering solutions to known problems, but it struggles with complex, multifaceted tasks that require understanding, innovation, and context—the kinds of challenges that real developers face every day.

Summary of Key Findings from the GSM-Symbolic Paper

The paper suggests that LLMs rely on pattern recognition rather than true reasoning. While AI has made impressive strides, there’s still a long way to go before it can match the reasoning and problem-solving capabilities of human developers.

Why Mathematical Reasoning is a Core Challenge

For tasks that require nuanced reasoning—like debugging complex code or designing new features—AI’s inability to reason like a human makes it an unreliable partner. It’s unable to understand why certain solutions might work better than others based on the context of the project.

Implications for Coding Tasks: Complexity, Context, and Correctness

AI is great at solving predictable problems with structured data, but when it comes to tasks requiring creativity or complex logic, it struggles. The paper suggests that LLMs are currently best suited for tasks like querying information and summarizing documents rather than handling intricate coding challenges that require creativity and context.


How to Use AI Effectively in Coding Tasks

AI can still be an extremely helpful tool for developers, but it must be used properly to avoid frustration and errors. Here are some tips for using AI effectively:

  1. Treat AI as an Assistant, Not a Replacement

    AI is a tool that can support you in your work, but it can’t replace the expertise, creativity, and problem-solving skills you bring as a developer. Always approach AI as a junior developer that you guide, giving it clear instructions and correcting its mistakes.

  2. Focus on Small, Specific Tasks

    AI is most useful for tackling small, specific tasks. Don’t ask it to build an entire application from scratch—break down tasks into manageable pieces and ask for help with individual parts of the code.

  3. Prompt AI Clearly

    AI is only as good as the prompts it receives. If you don’t ask it the right questions or provide the necessary context, you’ll get inaccurate or irrelevant results. Be clear and specific in your requests.

  4. Use AI for Repetitive Tasks

    AI is perfect for handling repetitive, time-consuming tasks like writing boilerplate code or checking for simple bugs. Save your time and energy for the complex, creative aspects of development that require a human touch.

  5. Know When to Stop Relying on AI

    If the AI is struggling to solve a problem, don’t keep pushing it in hopes of a better result. Recognize when it’s time to take over the task yourself or seek help from another human developer.


Common Pitfalls and How to Avoid Them When Using AI in Development

AI is a powerful tool, but it’s easy to fall into the trap of over-relying on it. Here are some common mistakes developers make when using AI and how to avoid them:

  1. Relying Too Much on AI for Problem-Solving

    Some developers, especially those with less experience, turn to AI to "fix" problems without truly understanding the issue. This is dangerous because it doesn’t encourage critical thinking or problem-solving skills. Solution: Use AI as a tool for learning, but always try to understand the problem and solution yourself.

  2. Expecting AI to Write Perfect Code

    AI is far from perfect. It can make mistakes, offer suboptimal solutions, or generate code that’s difficult to maintain. Solution: Always review and test AI-generated code thoroughly. It’s a starting point, not a final solution.

  3. Not Providing Enough Context

    AI needs clear instructions and context to provide accurate results. Solution: Be specific in your requests. If the AI doesn’t understand the problem, it won’t be able to provide a useful solution.

  4. Treating AI Like a Senior Developer

    AI tools can sometimes seem like they know everything, but they’re still not capable of handling complex coding tasks or offering creative solutions. Solution: Treat AI like a junior developer and guide it through the process.


TL;DR: AI Can’t Write Better Code Than Devs

AI is a powerful assistant, but it cannot replace the creativity, problem-solving, and reasoning abilities of human developers. While AI can help with simple tasks like code generation and bug fixing, it falls short when it comes to understanding complex problems, handling context, and making creative decisions. Developers should use AI to enhance their work, not replace their skills. AI should be treated as a tool—something that helps developers write better code more efficiently, but not something that can think, reason, or innovate like a human. The key to great software development remains the same: growth, learning, and creativity, areas where AI still has a long way to go.

Loved this article? 🚀 Check out my side project, Svelte Mini Apps!

Discover open-source tools built for everyday use.

Top comments (33)

Collapse
 
peter-fencer profile image
Peter Vivo • Edited

So true:

  • "The paper suggests that LLMs rely on pattern recognition rather than true reasoning."
  • "Treat AI like a junior developer and guide it through the process."

In my workflow, special when I work alone ( side project ), AI is a good to argue. So sometime capable to replace a talk with a collogues.

Collapse
 
peter_truchly_4fce0874fd5 profile image
Peter Truchly

Yes, in my experience the AI assisted coding is mostly reminding me times when I had to review and patch code after my younger colleagues. However, as it was with the colleagues, AI often knows some interesting library, technique or idea.
Certainly beneficial, does not solve for everything.

Collapse
 
dev_michael profile image
Michael Amachree

Yep, I use it a ton for my side projects, haven't used it much in large open source or collaborative projects yet, but I'm sure to try some day.

Collapse
 
cheetah100 profile image
Peter Harrison

You should add the caveat 'yet' to this article. Broadly speaking I agree that AI can be weak. I just had a bug in some code I'm writing and I asked Claude to identify the problem. The first response recommended making changes which didn't really seem to address the issue. When I challenged it it was able to re-evaluate and actually identified the issue. The thing is I wasn't able to understand where the issue was occurring just by eyeballing the code. The reason is that the bug was a second order effect in a browser. Claude did manage to identify it. But could it write a whole system? Obviously not. yet.

Collapse
 
dev_michael profile image
Michael Amachree

I’d love to agree with you, but as mentioned in Apple’s article, given the current design, AI lacks true reasoning capabilities and won’t achieve it without a significant breakthrough. So, I’d say replacing human developers is still a distant possibility. The idea that AI might someday get there seems like an obvious stretch—one the article already addresses.

Collapse
 
cheetah100 profile image
Peter Harrison

Oh man. Here we go again. Is there some kind of objective benchmark we can test for in this "true reasoning" capability? A benchmark that 90% of humans pass but which AI consistently fails? The main issue with current models and the mechanism of training them is they use back propagation. It is a very expensive approach which results in crystalline models. Switch to Hebbian learning - no small matter - and suddenly you get local learning in real time and adaptive networks. Is that far away? Probably not given the value of what has been achieved to date. I 'lost' a debate in 2017 about how we wouldn't see systems like we see today for at least my lifetime. By lost I mean most people agreed my estimate was unrealistic. My estimate was 2020. Well, 2024 and we have machines which pass the Turning Test so convincingly they need to condition them to be explicit about being AI the whole time. Every time some proclamation is made about how we won't make further progress we just blast on past.

Thread Thread
 
dev_michael profile image
Michael Amachree • Edited

Yeah, you might be right—maybe. But I never said AI will never replace developers. What I did say is that AI can't replace developers, at least not with its current capabilities. As I mentioned, unless AI undergoes some groundbreaking advancements, it simply won't be able to do it.

If AI can't consistently produce new and innovative solutions at a 90%+ success rate like human programmers, no one who values quality would choose AI over human developers. Sure, small ideas or companies might rely on AI to handle most tasks, but do you really think a company like Apple or Google would fire their entire staff and let AI run everything? I highly doubt it.

To add to that, AI can probably handle what most junior developers do—which is kind of sad because it raises the question: how will they get hired and eventually become senior devs?

But currently, AI can't replace a professional's job. And I keep stressing right now because I never said it won’t happen in the future. I’m not here to reassure anyone that they won’t lose their jobs down the line or anything like that.

Thread Thread
 
richard_treitel_10eb32171 profile image
Richard Treitel

I've heard that AI can give junior practitioners, of many fields, an easy path to tap the collected experience (and occasional collected stupidity) of seniors. IoW it can shorten their learning curve. Perhaps it will put a premium on cognitive ability, as opposed to wisdom.

That said, I still treat Copilot as a junior, albeit knowledgeable, developer.

Collapse
 
miketalbot profile image
Mike Talbot ⭐

Aren't many of the limitations being addressed by providing the ability to execute code and reason in steps? This is a very fast-moving field, and dual-mode thinking is on the cards right now. I find that context is the biggest issue to AI working well on my codebase.

Collapse
 
dev_michael profile image
Michael Amachree

My point might not have been clear: if AI can write most or all of your code, you’re probably not building anything new or impressive. Try creating something like Bluesky from scratch using only AI. Or better yet, build an entire platform like Twitter (sorry, X) from the ground up with nothing but AI. Then, ask it to ensure the performance surpasses what human developers achieved. You’ll quickly understand why companies still hire developers instead of relying solely on AI, despite it being cheaper in their own words.

True innovation and breakthroughs currently come from humans. That’s not to say AI can’t assist with complex codebases—it absolutely can. But can it truly replace a developer? Not yet.

Collapse
 
miketalbot profile image
Mike Talbot ⭐ • Edited

Fair enough, I agree with that. It can't replace a senior developer or a software architect; if it did, it would create commodities (because that's what people would ask it to do, just like they ask for the articles we see littering this place).

My initial response was to the impression that "if it does work, it's a bad idea because it removes some things humans enjoy doing", which is what I took from my first reading. I don't disagree with your examples. If we were going to globally vote on whether AI should replace those things, whether we should ban it from those tasks or create laws to reduce its remit, I'd be voting on the side of the humans. No such laws will come, so we must ride the wave and work out what the machines will do next.

I don't know how quickly AI will advance; I don't know what it will continually fail at. The AI music generation stuff from Udio shocked me with its advancement, so who knows where.

To be fair, in my comment, I was also commenting, having read other comments here.

So, I was wandering around my general findings:

  • LLMs solve real problems right now, today, very well.
  • LLMs have continued to improve; the multi-step reasoning of o1 certainly seems to have made a difference in coding tasks. I find it a useful, knowledgeable partner - certainly not perfect at coding.
  • Discussing architectural concepts with ChatGPT Advanced Voice has opened me up to other algorithms and approaches and helped me talk through my ideas in a way that I've never experienced before.

I think the moral arguments of the copyright and training dilemmas are a challenge. I learned everything I know by looking at other people's work. If we make machines capable of doing that faster, it's only a matter of time until they are trained on every piece of publically available material, and we have to start morally talking about what can be learned from observation, who has the power to learn, and why can't someone invent a machine that can learn. We can pass a law in one jurisdiction, but this is an arms race, and there are many jurisdictions...

For now and for the foreseeable future, though, these tools are only really useful to aid a human in moving faster, so the output is still reliant on the human in the mix, their skills and their vision. That may be as far as they go.

There are many historical references for the challenges of technological advancements, especially during the Industrial Revolution. Perhaps this is the first time such advancements have stretched beyond business boundaries, though.

Collapse
 
efpage profile image
Eckehard

Recently I had to transfer a fairly complex code from VBA to Arduino C++, which is different in many aspects. I was truely amazed that the result from ChatGPT did compile without changes. There had been some substantial changes in error handling, as the Arduino platform uses a serial output only. So, it seems that AI gets better in handling "complex" problems.

Collapse
 
dev_michael profile image
Michael Amachree

I can't stress this enough: this isn't the kind of complexity I was referring to in my article. AI can indeed handle a lot of tasks—I even converted my project from React to Svelte using AI.

But if your code was truly as complex as you claim, to the point where it would challenge a human, AI wouldn’t have fixed it without errors. AI can certainly assist with some codebases and issues, but when it comes to really complex code, it's just not capable right now.

Collapse
 
dev_michael profile image
Michael Amachree

To add to that, AI can probably handle what most junior developers do—which is kind of sad because it raises the question: how will they get hired and eventually become senior devs?

But currently, AI can't replace a professional's job. And I keep stressing right now because I never said it won’t happen in the future. I’m not here to reassure anyone that they won’t lose their jobs down the line or anything like that.

Thread Thread
 
efpage profile image
Eckehard

Things are evolving rapidly and I´m pretty unsure where this will end up. Finally, all predictions are not worth much. Either things happen - or they don't. Or they develop in a completely unexpected direction.

Maybe AI can´t replace a professional's job. But even now it can do things humans can´t. I was just amazed that - over half a year or so - AI solutions have grown so much better. Maybe in one year or in ten AI can replace a professionals job - or make it much more creative, as it frees people from doing the stupid things.

The only thing we can say for sure: AI will change many job profiles. For good or for bad? Up til now this is a human decision...

Collapse
 
mo-dev profile image
Mo Andaloussi

I think developers who do not use the AI tools are the ones who will get replaced because they will become slower.

On the other hand, working a lot with those tools, you will get rusty.

Collapse
 
darkwiiplayer profile image
𒎏Wii 🏳️‍⚧️

I think over-reliance on AI tools is the first sign that what a developer is doing is on its way to becoming a task AI can take over. In the short term, making use of AI to be more productive than the competition might work, but in the long term, the only solutions are a) seek higher grounds by deepening your skills and taking on more tasks that simple AI automation still cannot handle or b) overcome capitalism.

Collapse
 
dev_michael profile image
Michael Amachree • Edited

Yeah, you're kinda right, I mean I also use AI (for web dev) but depending on your line of work as a dev you probably can or can't use AI a ton, as such it's not fair to say everyone who doesn't use AI will become slower.

But depending on the task, and using AI can be a real help.

Collapse
 
mo-dev profile image
Mo Andaloussi

What I mean by slower is that every developer has to Google search for a solution or how to install a dependency... Small tasks like this are much faster with a prompt than doing a Google search.

Thread Thread
 
dev_michael profile image
Michael Amachree

I completely understand—that's why I mentioned it depends on your line of work. If your company uses proprietary software and doesn't allow its documentation to be indexed by Google or utilized by existing AI models, then using AI might be counterproductive because it lacks the necessary context.

As a web developer, there's already a wealth of publicly available information that AI has learned from, making it quite useful in our field. However, for areas like cybersecurity or network engineering, depending on the company's policies and the nature of the work, leveraging AI might not be as feasible or beneficial.

Collapse
 
bradstondev profile image
Bradston Henry

I think you're right that, right now, AI agents will not be replacing traditional coders. My experience is that even when give the proper context, LLMs can still give you code or responses that don't reflect your need, and a human touch is needed.

I think another big problem with LLMs in the long term when it comes to creative solutions (e.g. writing, coding, etc) is that their are fundamentally "non-original". They do not make new approaches or ideas, they only use old ideas to find solutions and attempt to give you the best answer from what already exists.

I think "creative codes", or coders solving problems in unique ways or solving new problems will always been needed.

Collapse
 
dev_michael profile image
Michael Amachree

Yeah, what I do believe is that they will be replacing most of the existing no-code website builders, though.

Collapse
 
hello-insurance profile image
Hello Insurance

yet

Collapse
 
incrementis profile image
Akin C.

Hello Michael Amachree,

thank you for your article.
It's easy to read and provides a good overview of why you think AI can't replace developers.

In my opinion, a great analogy (it confirmes my experiences):
"Always approach AI as a junior developer that you guide, giving it clear instructions and correcting its mistakes."

I also really like how you describe the problem and offer a solution to it.

Collapse
 
chamster profile image
Konrad

Excellent points. AI is good and will aid us. However, the hype and exaggerated expectations vested in it as-per-today are harmful. In the future, maybe... But that also means "maybe not". And somehow, those who believe in the "maybe" tend to lean towards "surely". On occasion, lean quite heavily... Like, 89 degrees, heavily. :)

On the other hand, when I was a student, I read an article saying that instead of large desktop tower computers, everybody will be equipped with a personal, pocket-sized device and have global access to the world's full information for free...

But they also suggested we'd have flying cars, so I'd go with "maybe", still.

Collapse
 
nyangweso profile image
Rodgers Nyangweso

Great perspective, "Treat AI as a Junior Dev". Thank you for sharing

Some comments may only be visible to logged-in visitors. Sign in to view all comments. Some comments have been hidden by the post's author - find out more