Vibecoding7 min read

The strange double life of vibe coding

AI-assisted code gets dismissed as slop while major companies build with it. The gap opens onto a bigger question about work and what comes next.

In April 2026, Google CEO Sundar Pichai said 75% of new code at Google was generated by AI and approved by engineers. Anthropic reports that, as of May, more than 80% of the production-merged code lines measured at the company were attributed to Claude, with engineers directing and reviewing the work.

At the same time, the label “AI slop” has become part of the conversation about what gets submitted to open-source projects. The contrast is worth looking at closely, especially because the consequences reach beyond how a piece of code gets written.

The code and the label

“Vibe coding” does not always mean the same thing to everyone. In an early explanation of the term, developer Simon Willison distinguished accepting generated code without examining it from using AI while reviewing and understanding the result. The phrase has since been used more broadly for building with AI.

That distinction matters when reading the numbers. Google’s figure includes engineer approval. Anthropic describes engineers directing and reviewing its output. Neither claim establishes that those companies are shipping code without looking at it, and neither tells us what every engineering team does.

They do tell us that, at two companies building these models, AI is part of the actual development workflow. It is not just a demo or a weekend experiment. Both figures are company reports, not independent audits, and both companies have a commercial interest in AI. They are still useful things to put next to a blanket claim that anything built with AI is automatically worthless.

Linux Foundation Research reported in October 2026 that project maintainers estimated 25% of incoming pull requests were low-quality AI-generated submissions. That is a survey-based estimate, not an audit of every pull request. The foundation also described a rise in contribution volume and pressure on maintainers. An untested or poor-quality submission can leave someone else with real work to do.

That concern is different from dismissing a person or everything they build because an AI tool was involved. The useful questions are concrete: Does the application work? Can someone explain the parts that matter? Were the permissions, data and failure cases checked? Can the next person maintain it? Those are questions about the work, regardless of who or what typed each line.

What faster means

The productivity story is more complicated than either extreme. AI can produce a working draft quickly. It can also produce an answer that looks complete and needs significant correction. Lines of code alone cannot tell those apart. Anthropic explicitly cautions that its own increase in code output overstates the productivity gain because quantity does not measure quality.

In a METR study of early-2025 tools, 16 experienced open-source developers completed 246 tasks on familiar projects. The tasks where AI was allowed took 19% longer on average. That is a real result for that group and those tools; it is not proof that AI slows every developer or task.

The tools also keep changing. In a February 2026 update, METR said its newer data gave an unreliable estimate of the current productivity effect because of selection and measurement problems. Its researchers thought developers were likely more accelerated than in the earlier study, but could not confidently put a number on the change.

The 2026 Stack Overflow survey adds a useful detail: 48% of the 14,304 respondents to its trust question chose the answer that they trusted AI when they could easily verify its output. That is a survey of respondents, not a census of every developer. It suggests a way to make the discussion more useful: separate how quickly something appears from how much work it takes to verify and operate.

The cost of a model and a worker

A model does not need health insurance. A person does. The model also does not work for free: there are computing and model-service costs, as well as the time people spend specifying work, reviewing output, fixing mistakes and handling security. A meaningful comparison has to account for the whole job, not just the speed of the first generated answer.

Companies are already describing changes in workforce plans. In a message to Amazon employees, CEO Andy Jassy said he expected AI-related efficiency gains to reduce the company’s total corporate workforce over the next few years, while increasing the need for other kinds of work. That is a forecast from a CEO, not a count of jobs already replaced by AI. Amazon later announced approximately 16,000 role reductions in January 2026 as part of organizational changes. Its announcement did not say that each of those roles had been replaced by AI.

The wider employment data does not yet give a simple answer. Anthropic’s labor-market research found no systematic increase in unemployment among workers in highly AI-exposed professions since late 2022. It also found suggestive evidence that hiring of younger workers had slowed in exposed occupations. The researchers describe limits and alternative explanations; the study does not settle what caused an individual layoff.

Meanwhile, the U.S. Bureau of Labor Statistics projects software-developer employment to grow 15.8% from 2024 to 2034. A decade-long projection is not a guarantee, and it does not tell an individual worker what happens to a particular job next year. It belongs in the same picture as executives’ workforce forecasts, not in place of them.

The race toward something bigger

The capability question reaches beyond code. OpenAI’s charter defines artificial general intelligence in terms of highly autonomous systems outperforming humans at most economically valuable work, and says the timeline is uncertain. AGI and superintelligence are related ideas, not interchangeable labels. A system can exceed human performance on a particular task without establishing that it is broadly superintelligent.

Still, the progress can make the boundary feel close. Anthropic describes AI helping with both engineering and research toward future models. In one reported AI-safety research experiment, two human researchers recovered part of a defined evaluation gap in roughly a week. Anthropic says its agents recovered a larger share using about 800 cumulative agent-hours and $18,000 in compute; humans selected the problem and scoring rubric. The two efforts used different resources. Anthropic notes that the result did not transfer cleanly to production-scale models. Its account is evidence of a particular result, not proof that a system can independently build its own successor. The company says it is not there yet.

METR’s task-horizon work offers another way to watch the change: it measures task length by the time a human expert would take, then estimates where an AI agent succeeds at a specified rate. Its tasks are primarily well-specified software, machine-learning and cybersecurity work. METR warns that measurements above 16 hours are unreliable with its current task suite. A longer measured horizon is interesting; it is not a measure of every job or a declaration that superintelligence has arrived.

Some of what I see makes me wonder whether superintelligence is already here. These systems outperform people on some tasks, but I don't know whether the broader label has been earned. That depends on the definition and the evidence.

Both companies also describe risks as capabilities advance. OpenAI's charter warns that competition to develop AGI could leave too little time for safety precautions. Anthropic says fully recursive self-improvement could increase the risk of humans losing control of AI systems. Those are warnings about possible futures, not evidence that either outcome has already happened. They sit alongside the companies' reports of advancing capabilities and their plans to build more capable systems.

What to watch next

Over the next couple of years, I would watch three things together: the work agents complete with limited intervention, the hours and costs required to review and run that work, and what companies actually do with hiring and staffing. A model release can change the first. A production incident can change the second. A company announcement may illuminate the third, but it will not always explain every cause.

For a team trying this now, keep a simple record: choose a bounded task, save the starting requirements, record model and human time, include the cost of review, and check the result after someone unfamiliar with the build has to maintain it. Run that comparison again when the tools or the task change.

That record will tell you more about your own work than the label attached to the code.