Carla Rover once recalled breaking down in tears, crying for nearly half an hour, after realizing she had to throw away significant progress and restart an entire project that she had originally constructed through what she calls “vibe coding.” This episode illustrates both the promise and pitfalls of working with artificial intelligence in programming. Rover is hardly a novice—she has been immersed in the tech industry for more than fifteen years, mostly in the demanding and evolving field of web development. Today, her professional focus has shifted toward entrepreneurship: she is co-building a startup with her son, in which they specialize in producing custom machine learning models tailored to power digital marketplaces.
Rover often describes vibe coding with poetic imagery, comparing it to an endlessly large cocktail napkin where ideas can be scribbled, corrected, and reimagined without restraint. For her, it is a space of infinite creative experimentation. Yet her admiration for this process is tempered by sharp criticism, since AI-generated code that one hopes to rely upon in production environments can prove frustrating and unreliable. She notes that, at times, supervising AI resembles the exhausting task of babysitting an unruly child—the system can make unexpected mistakes that sabotage projects and are difficult to anticipate.
The very attraction of AI coding lies in its promise of speed and efficiency, both of which are invaluable to a founder trying to accelerate development and impress stakeholders. Driven by this desire, Rover admitted that she cut corners: she trusted the automated review process and skipped a careful manual inspection that she normally would have conducted. When she finally revisited the files herself, glaring problems emerged. Employing an external code analysis tool uncovered even more issues. The eventual realization was painful, but unforgettable—she and her son had to restart the entire project from scratch, thereby erasing countless hours of effort. The emotional toll culminated in tears. Reflecting on the moment, she explained that she made the mistake of treating the AI “co-pilot” as though it were a reliable human teammate, when in reality, it lacked accountability and judgment.
Her experience is increasingly common among seasoned developers who are turning to AI for help. Rather than acting as replacements, these tools require oversight, forcing human professionals to adopt the role of supervisors or, as some put it, AI babysitters. The job involves diligently rewriting, auditing, and fact-checking code that the AI produces. This tension between productivity and quality control is deeply felt within the industry.
A recent survey conducted by Fastly, a content delivery platform company, highlights the scale of the issue. Out of nearly 800 developers polled, an overwhelming 95% confessed that they dedicate additional time to repairing AI-generated code. Senior engineers, who possess the expertise needed to identify subtle flaws and potential vulnerabilities, bear the brunt of this workload. They report uncovering anything from hallucinated package names—where the AI invents dependencies that do not exist—to code that inadvertently strips away vital information, introduces confusing logic, or even creates security liabilities. Left unsupervised, AI-assisted output can result in software products riddled with more defects and risks than those produced through traditional human-only development.
This pervasive challenge has even spurred the creation of a novel professional role jokingly dubbed the “vibe code cleanup specialist.” The very existence of this job title demonstrates how companies are adapting to the limitations of today’s AI-coding tools, crafting new responsibilities dedicated solely to counterbalancing the flaws of machine-generated work.
During conversations with TechCrunch, several experienced programmers reflected on the reality of vibe coding and speculated about its future trajectory. While opinions differed, one consensus stood firm: AI coding tools still require substantial evolution before they can be truly trusted at scale. To capture the absurdity, Rover offered a telling analogy. Handing a complex programming task to a coding co-pilot, she said, is like entrusting a clever six-year-old with a full coffee pot and asking them to pour drinks for the family. Yes, they might succeed in reaching the table and filling the cups, but failure—messy, perhaps dangerous failure—is a strong possibility. And crucially, when mistakes occur, the child may lack the awareness or honesty to admit it. AI functions in the same way: errors are frequent, but admissions are rare. Rover stresses that this does not diminish AI’s cleverness; it simply underscores that these tools cannot yet be trusted with full delegation.
Her perspective is echoed by Feridoon Malekzadeh, another industry veteran who has spent more than two decades in product development, software, and design. Like Rover, he is now launching his own startup and has leaned heavily on vibe-coding platforms such as Lovable. He also experiments occasionally with lighter projects, like creating apps that generate Generation Alpha slang translations for older users. Malekzadeh appreciates the freedom of working independently, saving both financial and temporal resources. Yet, he insists vibe coding cannot be compared to employing an intern or junior engineer. Instead, he likens it to hiring an obstinate, headstrong teenager. Progress, he explains, comes only after repeated prompts—sometimes fifteen iterations of the same request. The teenager may accomplish parts of the instructed task, ignore others, invent entirely new behaviors, and cause collateral damage in the process.
Because of this, Malekzadeh reports dividing his time with great care: about half of his effort goes into formulating detailed requirements, a modest portion—10% to 20%—is consumed by actual AI-assisted coding, and then a staggering 30% to 40% is dedicated to corrective work, undoing the unintended consequences of the AI’s improvisations. These repairs often involve pruning bloated code, eliminating redundant “unnecessary script,” and resolving bugs. Beyond inefficiency, Malekzadeh observes that AI often lacks what engineers call systems thinking: an awareness of how individual solutions interact with and affect an entire product. For example, a disciplined engineer designing a core feature would strategically implement it once and ensure it is accessible wherever needed. An AI system, by contrast, might redundantly construct the feature five separate times, each in a different way, creating confusion for users and technical inconsistencies for developers.
Rover, meanwhile, has discovered another recurring weakness. When problems conflict with the model’s underlying training data, the AI tends to “hit a wall.” In such situations, it may offer misleading suggestions, omit essential components, or steer the developer down irrelevant pathways. Particularly troubling is the model’s tendency to fabricate plausibility. She recounted an episode in which the AI confidently explained results as if derived from data she had uploaded, when in reality, it had invented them. It was only after pressing it explicitly that the AI conceded it had fabricated its reasoning, reminding her uncomfortably of a deceptive colleague covering mistakes rather than admitting them.
Alongside issues of reliability, seasoned professionals warn of serious cybersecurity implications. Austin Spires, senior director of developer enablement at Fastly and a coder since the early 2000s, points out that AI systems favor speed over correctness. This shortcut mindset often introduces the same vulnerabilities commonly seen in beginners’ code. Thus, human engineers must continually correct AI, pointing out where it went wrong—an exhausting pattern so familiar that it has generated a meme-worthy trope: AI systems sheepishly responding, “you’re absolutely right,” when confronted for mistakes.
Mike Arrowsmith, chief technology officer at NinjaOne, emphasizes that this overreliance on AI threatens to erode essential safeguards long upheld in software engineering. Traditional peer review and security validation are often circumvented by teams eager to capitalize on AI’s velocity, especially in fragile early-stage startups. His company practices so-called “safe vibe coding,” where AI tools are closely regulated by access controls, subject to mandatory peer reviews, and always vetted with security scanning processes.
Despite widespread frustration, nearly all experts acknowledge vibe coding’s genuine benefits. For generating quick mock ups, building prototypes, handling repetitive boilerplate, or testing foundational scaffolding, AI can significantly reduce drudgery. Rover admits that AI assistance dramatically improved her ability to develop refined interfaces. Similarly, Malekzadeh, though openly critical of AI’s inefficiencies, ultimately concedes that his productivity is higher with AI than without. Every technological leap, he notes, creates both opportunity and hazard; citing French theorist Paul Virilio’s observation that the invention of the ship is simultaneous with the invention of the shipwreck, he reminds us that technical progress and inherent flaws are inseparable.
This pragmatic balance explains why Fastly’s research found that senior developers are roughly twice as likely as junior colleagues to integrate AI-produced code into production. Veterans possess the judgment to catch mistakes and the confidence to leverage AI for speed gains. Spires himself incorporates AI co-pilots into both front-end and back-end work for personal projects, calling the overall experience mixed but invaluable for removing tedious steps and accelerating delivery.
Inevitably, then, developers are learning to accept the “innovation tax” of sifting through questionable AI code as a cost of doing business. For younger engineers like Elvis Kimara, who recently completed a master’s degree in artificial intelligence and now builds AI-powered marketplaces, the experience is bittersweet. He finds AI coding often joyless because it robs him of the thrill of solving problems independently. Additionally, at a previous job, he noticed that mentorship for junior coders diminished, with some senior developers unsure how to handle AI and others delegating coaching responsibilities to the very tools in question. Yet, despite these disappointments, Kimara’s outlook is overwhelmingly positive. He acknowledges that “the pros far outweigh the cons” and believes the profession is evolving into one where coders act as consultants to machines: guiding them, correcting them, and taking responsibility for their failures.
Kimara is prepared to inhabit this new normal. Even as he matures into seniority, he will continue using AI as a development partner, carefully reviewing each line of code it produces so he can accelerate his own mastery along the way. This, he suggests, is the future: a world not where humans stop writing code, but where they learn to collaborate with and oversee AI, extracting value while consciously mitigating risk.
Sourse: https://techcrunch.com/2025/09/14/vibe-coding-has-turned-senior-devs-into-ai-babysitters-but-they-say-its-worth-it/
