AI Coding Assistants Don't Boost Productivity Or Prevent Burnout: Study
In the realm of technology, generative artificial intelligence (GenAI) has seen rapid advancements over the past two years, particularly since OpenAI launched ChatGPT on November 30, 2022. However, the integration of artificial intelligence (AI) in various sectors, including software development, is not a novel concept; it has been around for some time, aiming to enhance productivity and alleviate workloads through the automation of routine tasks.
AI's application in software development was designed to assist developers in creating and delivering software applications, which is particularly relevant in today's landscape dominated by the Software as a Service (SaaS) model.
How Have AI Tools Impacted Developers?
A study published in 2022, based on data collected in 2021, revealed that a significant 88% of software developers felt more productive when using GitHub's Copilot. Additionally, 59% reported lower frustration levels, 60% expressed greater job satisfaction, and 74% felt empowered to concentrate on more fulfilling work.
A senior software engineer reflected, “I find I have to think less, and when I do think, it’s about the enjoyable aspects. It sparks creativity, making coding both fun and efficient.”
In terms of efficiency metrics, developers reported faster completion of tasks (88%), greater speed in handling repetitive work (96%), heightened engagement (73%), reduced time spent searching for information (77%), and decreased mental effort on repetitive tasks (87%).
Another investigation involving 95 developers divided into two groups—45 utilizing GitHub Copilot and 50 not—found that 78% of Copilot users completed the task of creating a web server in JavaScript, averaging 1 hour and 11 minutes. In comparison, the group without Copilot managed a 70% task completion rate, averaging 2 hours and 41 minutes, representing a significant 55% slowdown.
As we approach 2024, GitHub’s Copilot has integrated OpenAI’s GPT-4, enhancing its capabilities to provide intelligent fix suggestions. A survey of 165 software engineers indicated that 50.9% found it extremely useful, while 29.7% regarded it as "quite a bit" useful.
Contradictory findings emerged from a study conducted by Uplevel, which analyzed responses from 800 software developers.
Insights from Uplevel
Uplevel specializes in evaluating coding metrics and discovered that the hype surrounding GenAI and tools like Copilot has not yet translated into significant productivity gains. Their analysis focused on performance metrics between teams using and not using Copilot, assessing factors such as pull request cycle time, throughput, bug rates, and extended working hours.
Rather than delivering on the promise of improved coding speed and quality, the study found that Copilot-generated code had 41% more bugs. Additionally, there was little effect on pull request cycle times, throughput, or coding speed, with only a minor 1.7-minute reduction in cycle time.
Moreover, the AI assistant did not mitigate burnout risks. Uplevel’s “Sustained Always On” metric—tracking extended hours beyond regular work—showed that while both groups experienced declines, those with Copilot saw a 17% reduction, compared to a 28% reduction for those without it.
Ivan Gekht, the CEO of Gehtsoft, noted the challenges of relying on LLMs for productivity: “To enhance productivity, both the LLM must be competitive with human capabilities, and the user must know how to use the LLM effectively.” He continued, “Software development is largely cognitive—grasping requirements, designing systems, and navigating limitations. The coding itself is the easier part.”
Gekht added, “Understanding and debugging AI-generated code is becoming increasingly challenging, making it often more practical to rewrite code than to troubleshoot.”
So, has generative AI played a role in improving your productivity and operational effectiveness? If so, in what ways?

Comments
Post a Comment