Group Actions
Engagement
  • Author Author: Catwell
  • Date Created: Date Created
  • Views 962 views
  • Likes 5 likes
  • Comments 2 comments
Related
Recommended

Is Using AI to Generate Code Actually Helpful?

Catwell

image

An AI generated image seemed apt here. (Image Credit: Igor Omilaev/Unsplash)

Generative AI is changing how people write software, but that depends on how businesses deploy it. If implemented wisely, it can boost development time and keep teams from performing repetitive work. However, if misused, the system may delay long-term growth, create fragile systems, and expose businesses to costly setbacks. MIT recently published an article explaining how generative AI introduces hidden costs when used for coding.

We're already seeing generative AI used more frequently with knowledge work, with coding at the forefront. For example, GPT-4.1 has implemented advanced programming features, inching closer to automated development. The tool promises significant productivity gains. GitHub reports that developers working with CoPilot complete tasks significantly faster, achieving a 55% efficiency boost. McKinsey's study discovered that generative AI can shorten completion times by 50% for certain development tasks.

It's important to note that these studies were conducted in controlled conditions. Developers tackled small and clearly defined tasks. Real-world software development involves implementing new code into sprawling legacy systems, managing dependencies, and ensuring long-term maintenance. Scaling AI-generative code across those environments may introduce new risks.

Technical debt, which is the hidden cost of quick fixes that make systems harder to maintain, poses a significant risk. It has already fueled high-profile cases, from the Y2K bug to the 2022 scheduling collapse at Southwest Airlines. According to the Consortium for Information & Software Quality, the United States' economy carries $2.4 trillion in projected technical debt annually, and many companies invest very little toward fixing it. AI-generated code can worsen this problem by introducing integration issues, instability, and duplication. Additionally, studies show that code quality has worsened with heavy AI use, including rising code churn and reduced delivery stability.

GitClear discovered an eightfold jump in duplicated code. Google's DevOps team reported that heavy AI use reduces delivery stability. Those risks can increase in brownfield systems as AI tools lack full awareness of existing code bases.

AI is more effective in greenfield projects, such as rapid prototyping. With rapid prototyping, early code is reworked, making debt less costly. In brownfield environments weighed down by legacy systems, or with inexperienced developers, AI can turn short-term speed into long-term fragility.

Organizations can balance opportunity with risk. This can be achieved by establishing clear policies for AI tool usage, prioritizing technical debt management as an engineering focus, and investing in training and mentorship for developers to recognize and mitigate AI-driven risks. 

Have a story tip? Message me at: http://twitter.com/Cabe_Atwell

  • Chatgpt and copilot AI have not improved my coding skills but it has improved my coding output.

    I use Arduino's for animation projects in model railroading. I recently used WS2812 neopixels in a project. A mini office tower needed lighting at different location. Typically I would run individual LED or an LED strip. I discovered fastled as a library to control individual LEDs in a string. Three wires only required.

    I'm a resurrectionist not a programmer. I find code parts and stitch them together to create a finished project. I had no experience with programmable LEDs and wanted to try some fancy light sequences. 

    After watching a friend of mine (true programmer) use chatgpt (paid service). I started partnering with AI. Learning how to communicate with AI was the biggest challenge. Never assume what it provide is correct. Test, wash-rinse-repeat cycle. I completed the task in record breaking time (for me). AI also provided suggestions on features that would improve what I was trying to accomplish.

    When I was employed I could ask professional programmers how to do something. They would provide a framework and I would go off an attempt a solution. AI is that programming partner.

    I understand what the code does but to write it myself would have taken me a very long time. Most of the time would be learning the details (i.e. what is the string to initialize the LED lights) to accomplish what I want. I am a novice with this tool. In the hands of a professional I can see advantages. 

    I confess I have mixed emotions with completed projects. AI limits my sweat equity in a project. Do I deserve all the credit when it is done. I don't recall hearing my carpenter friends says they were not happy with their finished project because they did the work on their new tool. 

  • It all comes down to testability.

    My beef with AI is that you cannot verify that what the code will do under all input combinations.

    That is a BIG risk especially for systems where people can be harmed.

    I will continue to watch developments, but for now I would be very cautious about adopting any AI generated code into a project.