What's the point?

Something strange is happening in the world of technology. We built AI to help humans do things faster and better. Increasingly, though, AI isn't helping humans at all. It's helping other AI. And when you stop to look at what that actually means, one question keeps surfacing: what's the point?

The loop

Here are things that are happening right now, in production, at scale:

Each of these, taken individually, sounds like a reasonable use of technology. Zoom out, though, and the picture starts to look circular. AI generates something, then AI processes what was generated. Humans stand off to the side, vaguely supervising, hoping the loop produces something useful.

The email that ate itself

The email case is especially revealing. It demonstrates something researchers have started calling the AI productivity paradox: the tools designed to save us time end up creating more work, not less. A UK government trial of Microsoft 365 Copilot across 1,000 licences found that while users completed emails faster, there was no evidence of improved overall productivity. Users averaged just 1.14 Copilot actions per day. The AI hallucinated throughout the trial, requiring manual fact-checking that consumed whatever time was saved. Excel analysis done with Copilot was actually slower and lower quality than without it. Harvard Business Review put it bluntly in a February 2026 headline: "AI Doesn't Reduce Work, It Intensifies It." A separate HBR piece coined the term "workslop" for the torrent of AI-generated content that floods inboxes and Slack channels, noting that 95% of organizations see no measurable return on their AI investments despite doubling their usage since 2023. When AI expands your draft into a polished email, and the recipient's AI compresses it back into bullet points, the net information transferred is roughly the same as if you'd just sent the bullet points yourself. The two AIs did work. The humans gained nothing.

The ouroboros problem

There's a deeper issue lurking beneath the loop: model collapse. A landmark 2024 paper published in Nature, led by Ilia Shumailov at the University of Oxford, demonstrated that when AI models are trained on data generated by previous AI models, their quality degrades progressively. Each generation inherits and amplifies the errors and blind spots of the one before it. Shumailov compares it to photocopying a photocopy, over and over, until you're left with a dark, unreadable square. The researchers call this the AI ouroboros, after the ancient symbol of a serpent eating its own tail. As AI-generated content saturates the internet, and as new models scrape that content for training data, the feedback loop tightens. The diversity and richness of human-generated knowledge gets diluted, replaced by a kind of statistical average that grows blander and more error-prone with each cycle. This isn't theoretical anymore. Practitioners have reported observable quality degradation in everyday AI tools compared to outputs from just two years ago, with the same inputs producing noticeably worse results.

Almost right is worse than wrong

In software development, the loop creates a particularly insidious failure mode. Research shows that AI-generated code creates 1.7 times more issues than human-written code. AI pull requests have a 32.7% acceptance rate, compared to 84.4% for human-written ones. Bug rates have risen 9% per developer since AI coding tools became widespread. The most dangerous category isn't code that's obviously broken. It's code that's almost right. As one analysis put it: "Wrong code fails tests immediately, ten minutes to fix. Almost-right code passes tests, looks clean in review, ships to production, then detonates at 3am under edge cases nobody tested." When an AI review tool examines AI-generated code, it's one statistical model checking the plausibility of another statistical model's output. Neither has true understanding of the system's intent, its edge cases, or the business context it operates in. The review catches surface-level issues while the subtle, almost-right bugs slip through, because both models share similar blind spots.

What gets lost

The thing that gets lost in the loop isn't efficiency. It's learning. When a student uses AI to write an essay and a teacher uses AI to grade it, neither the student nor the teacher engages with the material. The student doesn't struggle with articulating a thought, which is where understanding actually forms. The teacher doesn't read closely enough to notice a genuine spark of insight or a fundamental misunderstanding that needs addressing. When a developer accepts AI-generated code and delegates review to another AI, the developer doesn't build the mental model of the codebase that makes them effective at debugging, extending, and reasoning about the system over time. When an email is expanded by AI and summarized by AI, neither the sender nor the receiver has to think carefully about what they actually mean. The loop doesn't just waste computational resources. It atrophies the human capacities that make the work meaningful in the first place. We're not saving time. We're hollowing out the process.

So, what's the point?

The honest answer is that sometimes there isn't one. Sometimes AI-on-AI is pure friction disguised as progress, two machines doing busywork while humans watch from the sidelines. But that doesn't mean every use of AI is pointless. It means we need to be more deliberate about where in the loop a human actually needs to be present, and where AI genuinely removes toil rather than just displacing it onto another AI. A few principles that might help:

The technology is powerful. The question is whether we're pointed at the right problems, or whether we've gotten so excited about the capability that we forgot to ask what it's for. Maybe that's the real point.

References