The speed of development teams using AI-based programming automation tools was almost twice as fast as those that did not, according to a GitHub experiment conducted a few years ago. Not all CEOs and CTOs have noticed this, but the future is here.
The gap between the understanding of the situation by managers, developers, and the real capabilities of AI is not a fiction. In the Stack Overflow 2023 survey, only 70% of the professionals surveyed had tried AI, and of those, many had only used it for minor performance improvements.
What can and can’t AI do in development automation right now? How will the situation change in the near future? Anton An, who is an experienced and recognized IT developer, answered all these questions.
In addition to GitHub Copilot, the market includes assistants like Codeium and Tabnine, and Transformer-based models like ChatGPT, CodeLlama, and StarCoder. A more detailed review can be found here. AI-powered software also automates testing. Beyond inserting simple templates, AI can now perform several key functions:
AI developer assistants handle prototyping, including for UI/UX. From textual prompts, assistants can form entire components (React or Angular). Some are already capable of creating GraphQL and integrating databases on their own. “Smart” assistants perform code review, and robotic refactoring sessions are able to identify up to 40% more suboptimal solutions than manual code reviews.
Transformer-based Code Models, such as StarCoder, are pre-trained on large amounts of data and can guide a developer on how to write a particular module. Thanks to machine learning, such systems predict code effectively by capturing the program’s context.
AI for test automation can automatically test code, giving a project 100% test coverage that human programmers would never achieve. Neural network testing is particularly effective in finding threats and can work 24/7 to find vulnerabilities in your architecture.
The history of development knows two radical upheavals that affected speed and quality. The first was Agile, CI, and DevOps. The second revolution is AI. For example, AI-assisted code generation helped API developers avoid engineering repetitive controllers. The AI understood the code’s context and was able to generate new modules for routing automatically. One result of its work without any human involvement is as follows:
from flask import Flask, request, jsonify
app = Flask(__name__)
@app.route(‘/users’, methods=[‘POST’])
def create_user():
data = request.json
if not data.get(‘name’):
return jsonify({‘error’: ‘Name is required’}), 400
# Logic to save user
return jsonify({‘message’: ‘User created successfully’}), 201
if __name__ == ‘__main__’:
app.run(debug=True)
AI can significantly cut down the hours spent on routine coding. It can produce a set of unit tests or a basic sorting function faster than a human could type the first line. For example, asking Copilot to generate a Python script to parse a JSON API response delivered a workable draft in under a minute, saving at least 20 minutes of manual work. That speed frees you to focus on architecture or solving a tough problem.
AI also functions like an on-demand mentor. When working with unfamiliar APIs, it can generate helpful suggestions with step-by-step explanations, simplifying the learning curve of complex libraries without the need to navigate dense documentation. As written by Science.org, the use of AI in a programming competition showed that the performance of developers who used it for initial code generation was 25 per cent higher.
Turing, a company that develops AI tools, shared a success story about a client that sells API access. They developed an AI assistant to find relevant information for the development team from API documentation and Stack Overflow. As a result, the company reduced development time by 40% and improved code quality by 30%.
Alongside the enthusiasm, there is fair skepticism. Developers note that while AI can write 70% of the code, you might spend just as much time on the remaining 30% because you need to understand the AI’s logic.
The quality can be unpredictable. One day, you get clean code; the next, you are unraveling a mess of redundant logic or bugs that only surface in edge cases. ChatGPT might provide a sorting algorithm that works fine until you feed it a list with duplicates, and it will choke. Fixing it can take longer than coding it yourself.
Security is another major concern. These models are trained on public code, so they might reproduce outdated practices or vulnerabilities like SQL injection risks. With cloud-based tools, there’s also a chance you’re exposing sensitive project details. For example, security flaws in the mobile app DeepSeek were partly tied to AI-generated code that left API keys exposed in plain text.
An anonymous ZenStack developer shared experiences with AI on his blog. While noting the increased productivity, he also listed significant problems, including losing control over development. If you are used to a RESTful API, you understand your legacy code. With a similar AI-based GraphQL API, making changes is not so easy, as you need to alter the schema first.
AI performs best on tasks with clearly defined boundaries. It excels at generating unit tests or well-defined algorithmic functions. However, its performance drops significantly in complex tasks, like integrating features into large or legacy codebases. A practical approach is to treat AI like a fast assistant: it delivers a quick first draft, and the developer steps in to refine it.
For code reviews, tools like CodeQL or DeepCode can scan code for style issues, bugs, or security risks before a human reviewer dives in. However, AI might fixate on minor details while missing the broader system context. The best practice is to let AI handle the initial pass, then bring in a human for the deeper look.
In one test published by frontend engineer Tom Smikowski, he gave the open-source tool CodeRabbit this tricky code: JavaScript export class ClickMeComponent { clickMessage = 5 as unknown as string; onClickMe() { this.clickMessage = ‘You are my hero!’; } }. The AI brilliantly passed the test, identifying the peculiar type casting and commenting on how this “numerical twist” changed the component’s initial state, proving it can understand not just syntax but also logical quirks.
Security remains a critical concern. AI-generated code can contain serious vulnerabilities, such as hardcoded credentials or patterns that open the door to cross-site scripting (XSS). To mitigate these risks, it’s important to perform security audits on all AI-generated code using static analysis tools like Bandit and following OWASP guidelines.
The use of open-source data also creates legal risks. In 2022, developers filed a class action lawsuit against Microsoft and OpenAI, claiming GitHub Copilot violates open-source licenses. The good news is that AI-powered testing algorithms can also find and fix many of these new vulnerabilities faster than a human cybersecurity professional.
AI agents are the next step, moving beyond suggestions to actively manage tasks. An agent in Cursor IDE was told to “build a simple REST API in Node.js,” and it set up routes, added error handling, and wrote a test suite in about 10 minutes—a task that would have taken an hour manually.
This leads to “vibe coding,” a term coined by Andrej Karpathy, where you describe the ‘vibe’ of the software you want in natural language and let the AI handle implementation. While tools like Replit’s AI agent show its power for rapid prototyping, this approach dramatically amplifies risks. Research from Stanford, for instance, shows AI-generated code is often riddled with security flaws, making rigorous human oversight absolutely necessary.
AI for coding is advancing fast. A 2024 report from Gartner predicts that by 2030, AI could write up to 50% of production code. Local models are also becoming popular due to privacy concerns.
The current consensus is that AI code generation delivers mixed results. It is highly effective for repetitive tasks like generating CRUD backends but requires careful oversight. The quality of the output must be verified, security implications cannot be overlooked, and developers must maintain their own skills.
The prevailing recommendation is to apply AI to routine elements—tests, boilerplate code, and simple functions—and then refine the output manually. When paired with code reviews, security audits, and a critical mindset, AI becomes a valuable tool. It is already reshaping how software is written, and while risks remain, careful use can make it a powerful asset for real-world projects.
