Vibe Coding, Al-Generated Code, and the Rise of the Digital Employee

Quick question. If a piece of software breaks tomorrow and nobody on your team actually wrote the code by hand, whose fault is it?

That is not a hypothetical anymore. It is the exact question keeping engineering leaders at JPMorgan, Goldman Sachs, and thousands of smaller companies up at night, because right now, an enormous and fast growing chunk of the world's code is not being typed by a human at all. It is being generated by AI, based on a plain English description of what someone wants, in a practice everyone is now calling vibe coding.

By April 2026, GitHub said that nearly half of all new code on its platform was AI-generated. Not assisted by AI here and there. Generated. Some surveys put developer adoption of AI coding tools above eighty percent. This is not a niche hobby for indie hackers anymore, it is standard practice inside some of the most risk averse, heavily regulated institutions on the planet.

So is this the biggest leap forward software engineering has seen in a generation, or are we quietly racking up a bill that someone, somewhere, is going to have to pay? I went looking for what the people actually running this technology inside major banks, consulting firms, and tech companies are saying, on the record, in 2026. Some of it will surprise you. Some of it confirms what a lot of developers have suspected for a while: this is genuinely useful, and it is also genuinely risky, often at the exact same time.

What Is Vibe Coding?

Let's clear something up first, because the name makes it sound a lot more casual than it is. Vibe coding does not mean developers are just vibing, throwing prompts at a chatbot and hoping for the best. At least, it is not supposed to mean that.

The term comes from Andrej Karpathy, the AI researcher and OpenAI co-founder, who used it in early 2025 to describe a new style of building software. Instead of writing every line yourself, you describe the outcome you want in plain language, and an AI model handles the implementation. You review it, nudge it, sometimes argue with it, and eventually ship it. It caught on so fast that Collins English Dictionary named it Word of the Year within about twelve months, which tells you something about how quickly this idea took over the conversation.

Karpathy himself has pushed back on people who think this means engineers stopped thinking. Writing about the leap he saw in coding agents around late 2025, he described a real turning point, where, in his words, agents “didn't work before December and basically work since,” with enough long term coherence to genuinely disrupt how people build software day to day. That is a very different claim from “AI now writes all our code and nobody checks it,” and the distinction matters a lot for everything that follows in this article.

What actually changed between 2023 and 2026 is the jump from simple autocomplete, the kind GitHub Copilot offered back in 2021, to agentic tools that can plan a multi-step task, touch several files, run their own tests, and fix their own mistakes with very little hand holding. Tools like Devin, Claude Code, Cursor, and Lovable now sit somewhere on a spectrum between a very fast assistant and something closer to an independent digital coworker.

Industries on Vibe Coding and Digital Employees

Forget the hype cycle for a second. The more interesting story is what large, conservative institutions are doing with their own money, because banks do not bet billions on a fad lightly. JPMorgan has gone further than most banks on AI generated code but it has done it carefully. Internally, the firm built its own large language model platform called LLM Suite, which staff use for everything from code review to plain old brainstorming, and its payments division built a tool called PRBuddy that writes pull request descriptions and flags what reviewers should look at first. On its own developer blog, JPMorgan Payments said folding AI into its software lifecycle has helped streamline workflows and enhance the quality and reliability of its software, citing a McKinsey study showing developers complete tasks roughly twice as fast with generative AI tools in the mix. That said, JPMorgan has been noticeably slower to put generative AI in front of actual customers compared to fast moving fintech rivals, because regulation, governance, and data protection are not optional extras for a bank this size. Katie Hainsey, who heads AI, ML, and Data and Analytics for the firm's Digital, Marketing, and Operations group, has pointed to a real but modest productivity bump of ten to twenty percent from code creation and conversion tools internally, a solid number, but nowhere near the breathless three to five times productivity claims you sometimes see in vendor marketing.

Goldman Sachs took a more dramatic step. In 2025, it began piloting Devin, an autonomous AI software engineer built by the startup Cognition, across its roughly twelve thousand person engineering organization. Marco Argenti, Goldman's Chief Information Officer, did not mince words about what this meant.

Argenti calls this a hybrid workforce, where AI agents take on the unglamorous jobs human engineers tend to dread most, like dragging old legacy code into modern frameworks, while humans focus on framing the problem clearly and checking the output. He has also floated the idea of a new kind of employee entirely, what he calls an AI native, someone whose actual job is to manage a small team of AI agents, the same way a manager runs a team of junior staff. Here is the detail that matters most though, Goldman kept hiring human software engineers the entire time it was piloting Devin, which is a pretty strong signal that, at least for now, this is addition rather than replacement.

McKinsey has run some of the most detailed research on this anywhere, based on a lab of more than forty of its own developers plus a survey of nearly three hundred publicly traded companies, and the findings genuinely cut both ways, which is refreshing in a space full of confident hot takes. Developers in McKinsey's lab wrote new code in nearly half the time and documented existing code in roughly half the time with generative AI tools helping. The top fifth of companies surveyed saw sixteen to thirty percent gains in productivity and time to market, alongside thirty one to forty five percent improvements in software quality, at the same time, not as a trade off. But on unfamiliar or genuinely complex tasks, those gains shrank to under ten percent, and junior developers sometimes took seven to ten percent longer with AI tools than without them.

McKinsey's own researchers are blunt about why the gap exists. As the firm put it plainly, “simply giving developers AI tools does not meaningfully move the needle,” and what separates the winners from everyone else is whether a company actually rebuilds its workflows, training, and accountability around the tool, instead of just switching it on and hoping for the best.

Microsoft CEO Satya Nadella has taken a more upbeat tone heading into 2026, writing that the industry is “beginning to distinguish between spectacle and substance.” He has even pushed back publicly on the term “AI slop,” arguing the conversation needs to move past arguing over labels and toward actually building good systems. But academia is less convinced things are fine. Edward Anderson, a professor at the University of Texas at Austin's McCombs School of Business, co-authored a 2025 study in MIT Sloan Management Review specifically on what happens when AI touches old, messy legacy systems, the kind every large company secretly has somewhere.

That warning carries real financial weight. The Consortium for Information and Software Quality estimates technical debt already costs United States companies around one and a half trillion dollars a year in lost productivity and cybercrime exposure. Careless AI use, Anderson argues, can quietly speed that number up rather than slow it down.

The View From the Ground: "It's Still on the Developer to Check the Work"

Strip away the big institutional names for a second and talk to the people who actually run AI and platform engineering teams day to day, the ones who get paged at 2am when something AI generated breaks in production, and you get a much more grounded take.

That one line cuts through most of the noise around this topic. It does not matter whether the code came from a junior engineer, a senior architect or autonomous AI agent with its own browser and command line. Somebody, somewhere, still has to actually check whether it works. The job has not disappeared, it has just changed shape. A developer's role in an agentic workflow starts to look less like typing every line and more like being an editor in chief, someone who reviews, tests, questions, and signs off before anything reaches production.

Put another way: the tools are new, but the rule is exactly as old as software itself. Nobody gets to ship code, human written or machine written, without someone taking responsibility for whether it actually does what it is supposed to do.

Productive Breakthrough or Expensive Waste of Energy? What the Data Says

This is the real question underneath everything else in this article, and the honest answer is that the evidence genuinely pulls in two directions at once. So let's look at both sides without flinching from either.

The case that this is genuinely productive

• GitHub reports Copilot users are roughly fifty five percent more productive on average and daily AI tool users save close to three and a half hours a week.
• Pull request turnaround dropped from an average of 9.6 days to 2.4 days for teams using AI coding tools, a seventy five percent reduction, mostly thanks to AI assisted descriptions, review suggestions, and test generation.
• McKinsey's top performing companies are seeing speed and quality improve together, not at each other's expense, because they paired the tools with retrained workflows and real accountability.
• Developers report a genuine wellbeing bump too. McKinsey found AI tool users are roughly twice as likely to say they feel fulfilled and regularly hit a state of flow, since the tedious boilerplate work gets handled for them.

The case that we should all be a little worried

Now for the other side. A December 2025 analysis by CodeRabbit of 470 open source pull requests found that AI co-authored code contained roughly 1.7 times more major issues than code written entirely by humans, with misconfigurations seventy five percent more common and security vulnerabilities running 2.74 times higher.

• Security firm Veracode separately reports AI generated code carries 2.74 times more vulnerabilities than human written code, and new AI related security flaws jumped from six in January 2026 to thirty five by March 2026.
• A randomized controlled trial from METR, a group that evaluates frontier AI models, found experienced open source developers were actually nineteen percent slower using AI coding tools, despite predicting beforehand they would be twenty four percent faster, and still believing afterward that they had been.
• GitClear's analysis of over 150 million lines of code found an eightfold jump in duplicated code blocks since AI coding tools went mainstream, a pretty clear sign that long term maintainability is quietly getting worse even while short term output looks great.
• Stack Overflow's 2025 developer survey found forty five percent of developers say debugging AI generated code is itself time consuming, and trust in AI accuracy among professional developers is at an all time low.

So which is it? Honestly, both. The tools genuinely save time on simple, well understood, repetitive work. They lose a lot of that advantage, or actively cost time, the moment a task gets complex or unfamiliar. The benefit is real, but it is wildly uneven, and it depends almost entirely on how disciplined a team is about actually reviewing what comes out the other end, rather than just shipping it because it compiled.

What Happens If We Let the Bad Code Pile Up?

Here is the uncomfortable part. If low quality, barely reviewed AI output keeps stacking up at the current pace, the consequences will not stay invisible for long. Mario Zechner and Armin Ronacher, two engineers behind the Pi coding harness inside the OpenClaw AI agent system, warned about exactly this in a widely shared Wall Street Journal report in May 2026, describing what they called a looming vibe slop crisis.

They are not the only ones saying it. A January 2026 academic paper from researchers at several universities, bluntly titled Vibe Coding Kills Open Source, argued that the rise of AI generated contributions is reducing genuine engagement with open source maintainers, creating real hidden costs for the volunteers who keep huge chunks of the internet's infrastructure running for free. A messy, very public incident involving the rsync file synchronization tool, where AI assisted commits sparked enough backlash that a protest GitHub issue went viral across tech social media, is a small but telling preview of how fast trust can evaporate once people suspect AI generated changes are landing in software millions of systems quietly depend on.

If this keeps going unchecked, a few things seem likely. Maintenance costs climb, because more developer time goes into untangling duplicated, poorly structured code instead of building anything new. Security incidents tied to AI generated vulnerabilities, hardcoded secrets, and weak authentication get more common, not less, especially as more of this code reaches production without a careful human pass. Public trust in software, already shaky after years of breaches and outages, takes another hit, particularly in industries like banking and healthcare where a single serious bug is not just an inconvenience. And companies that treat AI output as a finished product, rather than a rough draft, end up paying for it twice: once for the speed they gained, and again for the cleanup they could not avoid.

Will Vibe Coding End Up Making AI Mediocre?

There is a slower, deeper risk hiding underneath all the immediate bug reports and security warnings, and it is worth sitting with for a moment. AI coding models learn from the code that already exists out in the world, including, increasingly, code that other AI models wrote.

Think about what that means if it keeps compounding. If a huge share of the new code landing in public repositories is itself AI generated, full of duplicated patterns, inherited bugs, or shortcuts nobody properly reviewed, then the next generation of models gets trained partly on its own mistakes instead of on careful, well reasoned human work. Researchers studying this more broadly call the related risk model collapse, where a model trained repeatedly on its own output gradually loses nuance, diversity, and accuracy across successive generations.

So here is the genuinely uncomfortable question nobody wants to say out loud at a product demo. If today's AI coding tools produce work that is faster but, on average, a bit worse than careful human code, and tomorrow's models get partly trained on that same slightly worse code, does the whole ecosystem quietly drift toward mediocrity instead of improving? It is far too early to say this is happening at scale, and the major AI labs are actively working on data curation methods specifically to head it off. But serious researchers are flagging it as a real risk, not a hypothetical scare story, and it is one more reason human review cannot be treated as a step you eventually phase out once the tools get good enough.

Will vibe coding make AI mediocre? Not necessarily. But it almost certainly will, if the industry keeps treating AI generated code as finished work instead of a draft that still needs a human, or several humans, to actually check it before it ships.

The Rise of the Digital Employee: What This Means for Your Job

The phrase digital employee, the one Goldman Sachs uses for Devin, is turning into a real organizational category rather than just a clever bit of marketing copy. Bank of New York Mellon has reportedly brought on digital employees from AI Hub that handle coding tasks and payment instruction validation, complete with their own company logins, reporting to a human manager the same way any other employee would.

That points toward a workplace that looks meaningfully different from today's. More and more, engineers describe their job shifting away from typing every line themselves and toward structuring tasks for AI agents, setting clear boundaries, reviewing what comes back, and staying accountable for the final result, a bit like managing a team of very capable but occasionally unreliable junior staff who never sleep and never ask for a raise.

This shift carries real consequences that go well beyond software teams specifically. Anthropic CEO Dario Amodei has predicted AI could wipe out as much as half of all entry level white collar jobs within five years, while Ford CEO Jim Farley has warned more broadly that white collar work in general faces serious disruption, pointing out that entry level hiring at tech companies has already fallen by half since 2019. Goldman kept hiring junior engineers throughout its Devin pilot, which suggests the bank does not see digital employees as a full replacement yet, but the long run trajectory is genuinely uncertain. The people most exposed appear to be the ones just starting out, the same people who have always learned this craft by writing a lot of code themselves before anyone trusted them to review someone else's.

That last point might end up being the most important one in this entire piece. If junior engineers spend their early years reviewing AI output instead of writing code from scratch, the pipeline that has always produced senior engineers capable of catching AI's mistakes could be at risk too. Training the next generation of reviewers may turn out to matter just as much as training the next generation of models.

Conclusion

Vibe coding and agentic development are not a passing trend, no matter how the term itself ages. The money behind it is too large, the productivity gains in well scoped, well reviewed situations are too real, and the adoption by careful institutions like JPMorgan and Goldman Sachs is too deliberate to wave away as a fad.

But it is just as clear that this technology is not yet a substitute for solid engineering judgment. The CodeRabbit, Veracode, GitClear, and METR research all point the same direction: speed without review quietly creates debt that someone eventually has to pay down, usually at the worst possible time. The companies actually getting value out of this shift, according to McKinsey's own numbers, are the ones that rebuilt their processes, training, and accountability around AI, rather than just flipping the switch on and hoping for the best.

Maybe the simplest way to frame all of this is the truest one. AI coding tools and digital employees are not replacing the need for skilled engineering judgment, they are raising the price of not having it. As Mr. Brahma Mutyala put it, the responsibility for checking whether the work is actually correct, human or digital, never goes away. It just moves to a different spot in the workflow. The developers and companies who take that responsibility seriously stand to gain enormously from this technology. The ones who do not are likely to learn the hard way that fast and wrong is still wrong. Just faster.

Aadarsh Senapati

AI enthusiast · Writer · Developer
Bhubaneswar, Odisha, India

Aadarsh is a backend developer and data analyst, currently finishing his B.Tech in CSE at SRM University AP. Outside coursework, he spends a lot of his time building GenAI projects: RAG pipelines, document Q&A tools, and a few compliance-focused AI apps, mostly using LangChain, FAISS, and FastAPI. You can find his work on GitHub and Hugging Face.

He's also worked on the research side, as lead author on two papers on graph neural networks for recommender systems: one on dynamic similarity-aware attention, up on arXiv, and another accepted at the COMSYS conference in 2026. Between building applied tools and digging into the research, he tends to come at AI topics from both ends.

He writes about AI, machine learning, and web tech, mainly to make sense of fast-moving topics for himself and for anyone else trying to keep up.

This article is based on his current understanding of the subject. The space changes fast, so take it as a snapshot rather than a final word, and he's learning right alongside everyone reading it. If something doesn't add up, or you just want to talk AI and tech, feel free to reach out.

Website aadarshsenapati.in

LinkedIn aadarsh-senapati

GitHub aadarshsenapati

Thanks for reading, and I hope this deep dive into the hidden world of AI was both insightful and enjoyable.