AI coding assistants promise more productivity, but do they deliver

AI-powered coding assistants like GitHub Copilot and Anthropic's Claude Code are becoming increasingly popular with developers. But are they genuinely boosting productivity or just reshuffling tasks and burdens?

AI coding assistants streamline workflows in various real-world scenarios. The most common usage is as an "AI pair programmer" within the IDE: tools like Copilot (AWS's CodeWhisperer or Replit's Ghostwriter) automatically suggest code completions or entire functions as you type. This helps with boilerplate, syntax, or generating repetitive code quickly.

AI-assisted debugging and learning new languages were highlighted as emerging parts of the emerging AI workflow​. For example, a developer might paste an error message or a function into ChatGPT and get an explanation or a suggested fix in seconds. Likewise, AI can generate unit tests or suggest edge cases to help improve code robustness.

GitHub copilot boosts productivity, but not for everyone.

A study with over 4,800 developers across Microsoft, Accenture, and a Fortune 100 firm revealed that those using GitHub Copilot were 26% more productive. Copilot users completed about a quarter more tasks and even made 13.5% more code commits per week, with a 38% increase in code compilation frequency, indicating faster iteration​.

Although the study found "no negative impact on code quality" for Copilot-assisted work​, the productivity boost was uneven. Less experienced developers saw the largest gains (often 20–40% increases), whereas some senior developers saw only single digits.

AI tools enhance coding practices for CNC development.

Boosts efficiency and fosters accurate programming.

Integrate AI tools effectively:
- Use GitHub Copilot
- Validate AI suggestions
- Optimize your coding process.

AI is a partner in boosting developer productivity.

undefined Isaiah Dupree (@IsaiahDupree7) March 2, 2025

Real-world adoption and workflow integration

Despite the mixed perceptions, the adoption of AI coding assistants has skyrocketed. According to the 2024 Stack Overflow developer survey, 62% of developers already use AI tools in their development process, and another 14% plan to start​.

Kenny Johnston, VP of Product at Instabug, shared with Cybernews why he thinks AI coding assistants are making developers more productive by taking the workload of toilsome work from their plates.

"Today, AI helps write boilerplate code, tests, commit messages, or understand code they didn't write, he said.

“In the future, entire classes of ‘Keep the Lights On’ developer work will be offloaded to AI agents that perform bug fixes, solve errors, adjust apps for increased scale, confirm to new frameworks, and update dependencies."

What excites Johnston is that these tasks will be completed with minimal developer intervention.

"This leaves developers to focus on the most valuable thing they can provide and craft cutting edge digital experiences for their users."

Recognizing AI's shortcomings in coding

Sevalla is part of Kinsta and hosts thousands of WordPress sites. When Cybernews spoke with Kristóf Dombi, Head of PaaS Development at Kinsta, he shared how SevallauisGitHub Copilot extensively in their daily work.

"We find it valuable in several areas, such as translations and copywriting. With the advancements in reasoning models, its coding capabilities have also become more reliable," he told us.

Dombi added that GitHub Copilot Edits allows for an iterative workflow on code changes, giving them the control to approve or reject suggestions as needed.

"LLMs in this context can be very beneficial for individuals who are newer to certain codebases and are looking to familiarize themselves more rapidly. It serves as a helpful aid when navigating codebases in programming languages that one might be less experienced with."

However, Dombi also warns that it's almost certain that everyone has encountered situations where these tools underperform, even with relatively simple tasks.

"Due to these experiences, we strive to maintain an objective and critical perspective on these new tools. We aim to apply them where they are most effective and evaluate case-by-case whether they are suitable for integration into our workflows or production systems. LLMs have undoubtedly made a significant impact on coding. But we believe the transformative "revolution" many anticipate is still yet to happen."

Does AI reduce workload or shift burnout elsewhere?

Research so far gives a mixed verdict on cognitive load. A recent NeuroIS (neuroscience information systems) study measured developers' mental effort via EEG when using ChatGPT for coding tasks versus traditional web searches. The pilot results found that AI assistance did not generally reduce cognitive load compared to manual methods – it sometimes even increased mental load.​

The likely reason is that interacting with an AI still requires formulating good prompts, interpreting the AI's answers, and integrating them into one's code – which can be mentally taxing. If the AI output is confusing or off-base, the developer must expend effort to understand or correct it, adding to the cognitive strain.

AI assistants can both alleviate and add to cognitive load. They excel at handling known-knowns (boilerplate, known patterns), relieving memory load, and can guide debugging to reduce problem-solving load. But they also introduce new mental tasks, crafting prompts and double-checking outputs. This means developers must remain mentally engaged. The net effect on cognitive effort likely depends on the task complexity and the accuracy of the AI's outputs.

Code quality, maintenance, and new challenges

While AI coding assistants undoubtedly speed up code generation, questions remain about the quality and maintainability of AI-written code. One concern is that they may encourage a "quick fix" mentality – pumping out code that works initially but isn't the cleanest or most idiomatic solution, thereby increasing technical debt.

The GitClear study revealed a worrying trend of AI-assisted commits showing a higher proportion of outright additions (often copy-pasted code) versus edits or deletions​.

Bugs and inaccuracies in AI-generated code are another major issue. Coding assistants powered by large language models can "hallucinate" and produce code that looks plausible but is incorrect or insecure. This is especially true in edge cases or when the prompt is ambiguous.

"LLMs in this context can be very beneficial for individuals who are newer to certain codebases and are looking to familiarize themselves more rapidly. It serves as a helpful aid when navigating codebases in programming languages that one might be less experienced with,"

Kristóf Dombi.

The Uplevel report's finding of 41% more bugs with Copilot use is extreme, but it underscores the risk. If developers trust AI suggestions too readily, they might introduce errors that wouldn't have occurred with slower, manual coding​.

There's also the challenge of maintaining and reading AI-written code. Often, AI-generated code is syntactically correct but not very readable or idiomatic to the project's style. It might use unusual variable names or a different architecture than the team expects, making it harder for others to understand later.

Best practices are emerging to mitigate these issues. Many teams treat AI assistants as junior developers, which is helpful for drafts, but code reviews and testing are mandatory for all AI-generated code. Some AI tools incorporate safeguards – for example, Amazon CodeWhisperer has a built-in security scan that checks the AI's suggestions for common vulnerabilities or secret keys before accepting them.​

GitHub is rolling out "responsible AI" features that cite references for Copilot's suggestions and warn if a snippet looks potentially insecure or biased. In open-source settings, maintainers request contributors to label AI-generated code or ensure it meets style guidelines.

Productivity boost or hidden technical debt?

All this indicates that while AI coding assistants can introduce new challenges, the developer community is actively working on processes to address them.

With proper oversight, teams report that the benefits of faster development, more thorough exploration of solutions, and less drudge work can outweigh the drawbacks. But without such guardrails, there's a risk of accumulating buggy, hard-to-maintain code that shifts the burden to a later stage.

Experts agree that when developers and organizations leverage AI assistants thoughtfully, they can reduce cognitive load on trivial matters, help debug and generate code more efficiently, and ultimately boost productivity. But if used naively or unchecked, they could create extra work by generating errors or poor-quality code that humans must fix.

As with any powerful tool, the outcome will always depend on the skill and judgment with which it's applied. As an eternal optimist, I'm hopeful the future will consist of improved tools (and practices) to further tip the balance toward genuine productivity gains for developers using AI coding assistants.​