The Limits of Verifiable Work

By Computing

Anything that is easily verifiable will likely be solved by frontier models. AI does best in isolated environments where what counts as “correct,” and the signals that lead to it, are fast and obvious. Whether it is a math proof, code that compiles, a balanced spreadsheet, or short-term quantitative trading, if a task has a tight feedback loop, AI will master it.

Yet in the scope of human experience, these problems are only a fraction of what we set out to do. Most of what matters has no clear pass/fail test, and the data behind it is incomplete. Take long-term investing: between the day you buy a company and the day you sell years later, countless outside factors shift at once, and the final return rarely follows your original thesis neatly. When outcomes take years to play out and cannot be separated from luck and outside noise, there is no clean signal for a model to learn from. The same applies to human taste and preference. Even with emergent capabilities, models can only learn from what has been recorded. Yet much of human judgement is never written down, and as any market researcher knows, what people say they think is often not what they actually think or want. With so much of that inner context missing from the data, choosing the story worth telling and sensing what will actually resonate with people still comes down to human judgement.

Most jobs are a mix of both worlds. We often wish our work had clearer rules and simpler checklists, but as AI strips away the verifiable parts of our day, that remaining ambiguity is where our value lives. The moment you can easily prove your job is done right is the moment you know your job isn’t safe.

  • #ai
  • #careers
  • #investing
  • #judgement