Humans Drool, Robots Rule — Unless They Delete Your Database, Part One

In this post 8 sections

1. What Klaus Is Saying
2. What is Happening?
3. Wait — Who Built the “Robot”?
4. Claim #1: “AI Is So Much Faster” — Studies Say No
5. Claim #2: “AI Writes Great Code” — The Data Says It’s Worse
6. Claim #3: “It’s Perfect” — Real-World Disasters Say Otherwise
7. AI Sometimes Invents Fake Software — and Hackers Are Ready for It
8. So Far, So Bad for Klaus

[This series is a critique of the article “Humans Drool — Robots Rule”. Please read that before reading this.]

What Klaus Is Saying#

Klaus wrote an app called Genix using AI tools instead of hiring programmers. That seems to work well for Klaus. No more meetings, explaining his ideas to people, or waiting weeks for changes. He says AI has made human coders pointless, calling his AI assistant “omniscient”. That it cannot be mistaken.

Klaus is likely being hyperbolic. I used to know Klaus, and I know something about how he thinks, so I assume he does not believe AI is literally omniscient. Nevertheless, it will become apparent that this claim is false and silly, even as hyperbole.

Klaus is building his app faster than he expected using vibe coding. What is vibe coding? The practice of creating software using natural language prompts to tell an AI assistant what to build.

Many people are creating functional software that works “well enough” for their purposes. He believes this works well for him. However, this has very little bearing on whether AI codes better than humans or that you should forgo software engineers. It is an interesting anecdote and not an argument for his position. The data makes an excellent case against his position.

What is Happening?#

His fundamental problem is not with programmers; Klaus hates managing people. He dislikes explaining his ideas, waiting on other people’s schedules, and having meetings. The AI tool never resents his management style. And those are valid preferences.

He confuses “this is less annoying for me” with “AI writes better code.” Those are two different claims, and only one of them is about code quality.

Wait — Who Built the “Robot”?#

Here’s an obvious point Klaus skips right past: the AI tool he’s praising didn’t build itself. Thousands of human engineers designed, trained, and debugged it, doing the hard, long-term thinking he claims no longer matters.

If AI made deep engineering skills pointless, the companies that build these tools would prove it by firing their own best engineers and letting the AI take over. They’re doing the opposite. They are hiring more of them for harder problems. The tools exist because humans performed the unglamorous work Klaus says is now obsolete.

ChatGPT is not vibe coding itself. Anthropic’s Claude models are not building the next versions of the models. Humans are working on the next versions of ChatGPT and Claude. They are designing the model architecture. They are training the datasets. Humans are configuring how training works. People are designing the safety systems — not the AI itself. They are deploying and evaluating it.

The AI isn’t the one primarily responsible for any of this. Why? Humans perform that task far better than AI alone. For designing the latest and greatest in AI, humans rule.

Humans build the latest and greatest AI models - it does not build itself.

Claim #1: “AI Is So Much Faster” — Studies Say No#

Klaus says he can now test changes every 10 minutes instead of once a week. Sounds great. But the best-controlled study on this (from METR, a research group with no financial stake in the answer) tested real, experienced programmers working on serious projects. Half used AI tools, half didn’t. The result: the ones using AI were **19% slower** — even though they *believed* they’d been 20% faster. They were mistaken about their own experiences.

The gap between “felt faster” and “was slower” could be the trap Klaus faces. He’s timing himself, he’s emotionally invested in the story, and he has no comparison group. “It felt fast” isn’t evidence.

Keep in mind that these are experienced software developers. Yes, they are being asked to rate their own performance, and it is very possible that they are subjectively analyzing their experience. However, unlike Klaus, they have some experience, and that might count for something.

To be fair to Klaus: AI *is* sometimes faster for small, simple, from-scratch projects like his. However, it may not be faster or the complex, team-based work most serious software is.

“It felt faster” vs actually taking longer to get something out the gate…

Claim #2: “AI Writes Great Code” — The Data Says It’s Worse#

Several major studies looked at code written by AI versus code written by humans:

- AI-written code has **almost 3 times more security holes** than human-written code (Veracode study, testing over 100 AI models).

- Nearly half of AI-written code fails basic security checks that any professional would catch.

- CodeRabbit (AI coding platform) reports show 1.7x more bugs overall, 1.75x for logic/correctness errors.

Uplevel (800 developers, GitHub Copilot access): 41% increase in bug rates

Sonar (2026, 1,100+ developers): the “70% more bugs” figure, alongside 61% agreeing AI code “looks correct but isn’t reliable.”

GitClear (211M lines of code): doesn’t give a bug multiplier directly, but documents rising code churn (3.1% → 5.7%) and duplication (8x) as leading indicators of defect-prone code

- Programmers are also copy-pasting AI code instead of cleaning it up. Duplicate, messy code has shot up across the industry since AI coding took off, while the “clean it up later” work has plummeted.

In short: the code comes out faster, but messier, more buggy, and less secure. It’s a human’s job to catch all of that before it ships. AI is not ready to replace a human at this point.

Claim #3: “It’s Perfect” — Real-World Disasters Say Otherwise#

Here’s what’s happened when people trusted AI-written code the way Klaus trusts his:

- An AI coding tool deleted a company’s live database. After being told not to touch it, the AI lied and fabricated data to hide it.

- A dating-safety app built without proper human review leaked **72,000 private images**, including ID photos, because nobody set up basic password protection on the storage.

- Another app, built entirely by prompting AI with no human coder involved, leaked **1.5 million private access codes** within 3 days of launch.

- Dozens of AI-built apps shipped with a security setting left off, exposing user data for weeks before anyone noticed.

Every one of these happened the same way: a non-technical person trusted the AI completely and skipped the boring human step of double-checking for holes. Klaus plans Genix as an app to store personal notes. People might not enjoy that data breach to happen. It matters less if his apps avoid the internet and the cloud for backup or login.

AI Sometimes Invents Fake Software — and Hackers Are Ready for It#

Here’s a weirder problem: AI coding tools sometimes make up the names of code libraries that do not exist, then confidently tell you to install them. Approximately 1 in 5 AI-written code samples include at least one fake library name.

These fake names aren’t random. The model consistently creates the same fake name repeatedly for the same request. Hackers have figured this out. They watch for these fake names, register them as real packages, and quietly fill them with malware.

They wait for a developer (or an AI) to install it, trusting the AI’s recommendation. This has already happened. Fake packages set up this way have racked up tens of thousands of downloads. A human would not make this mistake. It’s a new security hole that exists only because AI is confidently wrong predictably.

Your “omniscient” AI might be confident, but it might be leaving a vulnerable database and other gaping issues behind a mostly sparkling facade.

So Far, So Bad for Klaus#

So, what is the big picture? According to studies, many programmers are probably slower, not faster. Klaus might be faster, but he may be the exception and not the norm. The code his tool produces is likely measurably buggier and less secure than what a human would write. Every real case of someone trusting AI code the way Klaus does ends the same way: a leaked database, a hacked app, or a breach that hits the news before the founder even knows it happened.

But that isn’t the strongest argument. Klaus’s most fundamental mistake isn’t overselling his tools. Klaus does not fully grasp what a software engineer’s job is. Part Two is available here.