Developers Can No Longer Explain Their Own AI-Generated Code as Critical Thinking Skills Atrophy

Engineers Struggle to Explain AI-Generated Code

Software engineers are facing an uncomfortable reality: they can no longer explain the code they submit for review. A developer with three years of experience recently admitted on Reddit that he relies on Claude to handle end-to-end tasks and struggles to defend his implementations when senior engineers ask questions during code reviews2

2

. "Half the time, I don't even fully understand the architecture or the code it just generated," he wrote, describing how his stomach drops when facing detailed questions about his merge requests.

This confession has sparked debate about whether AI assistants have crossed the line from productivity tool to cognitive crutch. The developer described his situation bluntly: "The crutch became the wheelchair." His critical thinking feels atrophied, his focus is shot, and he estimates that without AI, he would "code at the level of a first-semester freshman"2

2

.

OpenAI Engineers Cannot Decipher Their Own GPU Kernel Code

The problem extends beyond individual developers to frontier AI labs themselves. Jordan Nanos from research firm SemiAnalysis revealed that OpenAI engineers could not understand their own GPU kernel code while reviewing DeepSeek's multi-head latent attention implementation1

1

. "The guys who were scrolling through this code with us, it was kind of clear that they know a whole bunch about hardware, they know a whole bunch about the concepts in the system, and they just have no idea what this MLA kernel that they're showing us for DeepSeek actually does," Nanos told Fireside Alpha's Mansour Karam.

The AI understands it, tests it, and produces correct kernels that perform well, but human code comprehension has become optional. Jeremy Nixon, founder of chip optimization software company Infinity, told Business Insider that programmers are transitioning from creating products from scratch to babysitting large language models that cobble together code1

1

.

35% of AI Code Reaches Production Without Full Understanding

A recent analysis by software company Undo reveals the scale of the problem. Some 35% of AI-generated code "has not been fully comprehended before engineering teams push it to production"1

1

. Meanwhile, 29% of software engineer respondents witnessed productivity losses due to engineers needing to "unpick" AI-generated code multiple times per month, while 94% saw such incidents at least once in the past six months.

These statistics challenge the assumption that generative AI to write code automatically boosts efficiency. The sheer amount of time engineers spend reviewing AI outputs raises a fundamental question: is the technology more trouble than it's worth?

Amazon's 90-Day Code Safety Reset After Production Failures

Production failures are adding urgency to concerns about AI coding tools. Amazon implemented a 90-day "code safety reset" earlier this year following outages that disrupted customer orders1

1

. While the company disputed that AI was to blame, internal meeting notes revealed managers were concerned about the "high blast radius" of "gen-AI assisted changes."

Developers in discussion threads shared similar experiences. One software architect warned: "All fun and game till it crashes in production with permanent data loss and git blame and infosec sets up a meeting with you"2

2

. Another developer described being brought into four companies to fix troubled products, encountering what he called cognitive debt: "hundreds of bugs & regressions because nobody had any clue what they were writing" along with "redundant, duplicate and dead code everywhere."

One platform team used Claude for major Java and Spring upgrades, only to discover production bugs that engineers struggled to fix, ultimately requiring rollbacks of some microservices2

2

.

Paradigm Shift in Software Engineering or Skill Atrophy?

Researchers are finding that widespread use of AI can lead to cognitive decline among developers and skill atrophy, including among computer scientists1

1

. Developers fear they are drifting away from actually writing code and learning skills required to architect larger projects as employers push for more AI automation.

Some view this as a paradigm shift in software engineering rather than incompetence. AI researcher Sergey Cleftsow argued that "AI has advanced to the point where there's no need to worry about many things anymore. Developers simply need to define the architecture and correctness criteria; the neural network will handle the rest"1

1

.

Yet this optimism faces practical challenges. When AI also writes the tests, as one commenter noted, "the unit and coverage tests were written by AI and they don't really catch anything meaningful"2

2

. Without human understanding of the code, errors and hallucinations fall through the cracks.

The concern extends to the next generation of engineers growing up in an environment that actively undermines the value of understanding code itself. Whether software development can maintain quality and reliability as it becomes increasingly unrecognizable remains an open question for an increasingly digital world putting AI front and center.