AI & human judgment
When AI Is No Longer Scarce: Walter Benjamin on Knowledge, Expertise, and Power
4 min readGenerative AI, Walter Benjamin, Professional Judgment, Talent Development, AI Governance
Key takeaways
- Generative AI lets newcomers produce work that looks professional, so a polished output no longer proves that real expertise went into it.
- Expertise is not disappearing, but its authority now has to rest on reliable data, sound reasoning, a clear scope of use, and clear accountability when things go wrong.
- Judgment is built by gathering data, making mistakes, and correcting them; organizations that hand all of that to AI while asking people to review AI will run short of capable reviewers.
- Being able to use AI is different from having a say in how it works, and whoever controls models, data, and platforms shapes the conditions of work.
- These questions reach into education, labor markets, and public institutions, so individual productivity gains cannot settle them.
Benjamin's question, asked again
When Walter Benjamin wrote about mechanical reproduction, his concern went beyond how many copies of a work could be made. Photography and film changed the conditions under which art was produced and viewed, and with them, how people understood its value.
I want to carry that question into the age of AI. A legal opinion, a people decision, or a strategy analysis used to earn trust for reasons that had little to do with how hard the writing was. People trusted it because they knew what training the author had, which cases they had seen, and what consequences they would answer for. That experience and those institutions were bound up with the final document.
AI separates the content from the person who produced it. A complete, well-organized report might reflect years of research, or a few minutes of prompting and editing. Both may be useful, but readers can no longer infer the professional process from the surface of the output. Expertise remains; the way we used to recognize it has stopped working.
Why does handing work to AI leave fewer people able to review AI?
Judgment is built by gathering data, making mistakes, and correcting them, so organizations that hand all of that to AI while asking people to review AI will run short of capable reviewers. Mechanical reproduction changed how existing works were made and distributed. Generative AI goes further and takes part in forming the content itself. Consider an HR analytics report: a manager poses the question, a system pulls the data, a model helps categorize and draft, and an HR professional checks definitions, adds context, and makes the recommendation. One name may appear on the final report, but it was produced by people, models, data, and process together.
This leads to the tension I think HR leaders should pay closest attention to. Experienced professionals can spot what is wrong with a report because they have made bad assumptions, handled exceptions, and lived with the consequences of their decisions. If AI takes over the drafting, comparing, and revising, junior staff deliver faster but miss the experience that builds judgment. Organizations end up asking people to review AI while removing the very work that develops the ability to review. When a reviewer of a performance analysis or a financial forecast sees only a well-written finished product, "human in the loop" becomes little more than the last box on a workflow.
My conclusion is direct: professional development cannot shrink to teaching people how to write prompts. Organizations need to preserve opportunities for independent analysis, tracing sources, handling exceptions, and owning a judgment.
Does being able to use AI mean having a say in how it works?
More people can use AI, but users do not decide the rules it runs on. What data a model is trained on, what it may generate, which information it ranks first, and whether an organization can switch vendors are not choices an individual user makes. Authority that rested on information gaps is being challenged, while new dependencies may concentrate around models, compute, data, and platforms. Doing more is not the same as having more control over one's working environment.
So I argue that "democratizing knowledge" should be measured by more than access. It also depends on whether people can understand a system's limits, choose among tools, question its results, and take part in deciding where it is used. Once AI reaches education, labor markets, public services, and national security, those choices ultimately call for public institutions, not just better individual skills.
Three things to take away
- When assessing AI-assisted work, look past how complete it seems and ask where the data came from, whether the reasoning holds, where it applies, and who is accountable if it is wrong.
- When designing early-career development, deliberately keep some work in which people gather data, analyze independently, and handle exceptions themselves, rather than having AI draft everything.
- When adopting an AI tool, assess whether your team understands its limits and whether you could switch vendors later, not only whether people know how to use it.