Key points of this article
1.Generative AI's trough of disillusionment and workslop
With the payoff from AI investment still hard to see, generative AI is entering a "trough of disillusionment" as a reaction to overheated expectations. The spread of workslop — content that looks polished but is hollow inside — is a real problem, but as users get better at spotting the difference and autonomous AI agents keep improving, this issue should gradually fade.
2.PoC stagnation and the challenge of integrating outside knowledge
Many companies see their AI rollouts stall at the proof-of-concept stage because, fearing data leaks, they confine the technology to environments that only draw on internal company data. Making real use of AI means combining not just proprietary data but the vast body of knowledge held by society at large — the key is reconciling security with effective knowledge integration.
3.The human role, and the looming crisis in giving critical feedback
In the AI era, two roles remain for humans: giving instructions about purpose, and passing critical judgment on the results. But as work becomes automated, younger employees have fewer chances to learn the tacit know-how of the job, raising the risk that down the road there will be no one left who can judge quality correctly. Passing on that expertise — and safeguarding the quality of what AI learns from — is the challenge ahead.
——Lately "workslop" — a term for AI-generated material that looks polished but is hollow underneath — has been getting attention. 2026 is supposed to be the year AI agents move into the execution phase, yet many companies are still stuck at the testing stage. How should we make sense of that gap?
Sasaki: Talk of an "AI bubble bursting" stems from two things: the pullback after semiconductor stocks surged, and the fact that AI's contribution to economic growth and productivity still isn't clearly visible.
In the US, AI-related investment has surged, but for now it's mostly going into building data centers — how much wealth and productivity that eventually generates still isn't clear. Given that, I sense a kind of "disillusionment" with AI starting to set in.
In terms of Gartner's hype cycle, which maps technology maturity and adoption, a new technology first sees a rapid spike in expectations, then hits a trough of disillusionment, and finally settles into a stable plateau of adoption. I think generative AI is right at the point of entering that trough now.
In the end, it comes back to something we've heard endlessly: "the prompt matters." Ask a sloppy question and you get a sloppy answer, and that's what's producing the kind of output people are calling workslop.
In the past, you could tell at a glance if content was thin or the writing was clumsy. But now that AI lets anyone produce polished-looking prose, it's become a constant problem to tell from appearances alone whether there's any real substance behind it.
I think this connects to a deeper issue with computers in general. Back in 1988 when I joined a newspaper company, I once showed my editor a manuscript printed from a word processor, and he told me, "Convert it back to handwriting. When something's printed, it looks like it must be correct, and you stop noticing the mistakes." Today's workslop problem feels similar — once users get used to AI output and can instantly tell good from bad, the problem should naturally diminish.
And once AI agents can anticipate what users actually intend and autonomously produce genuinely good output on their own, this problem should move toward resolution as well.
