Thankfully, much of my work is shifting away from how work gets done toward what work I get to focus on that gets done.
I want to explore which parts of knowledge work are shifting to AI and which are staying with me as distinctly Joseph Redd’s work. I say that because I understand this shift in theory, in the abstract; but I’d like clarity on the specifics. That way, I gain a deeper understanding of what’s happening on the AI side of what I do so that I serve other humans by being all the more human.
I’ll start by asking Claude AI what traditionally human skills it excels at and is getting better at engaging than humans.
Here’s part of what it said:
Taking direction to solve problems — Being steered with constraints, goals, and context, and working within them. I respond well to iterative redirection mid-task.
Seeing patterns across domains — My training spans science, law, business, engineering, medicine, philosophy, and more, so I can surface analogies and structural similarities that specialists inside a single domain might miss.
Critiquing AI suggestions — I can review my own outputs or outputs from other AI systems, flag weak assumptions, identify gaps in reasoning, and suggest better framings or alternative approaches.
Staying current on specific AI capabilities — I can search the web to find what tools, models, or workflows exist for specific tasks right now, bridging my training knowledge with live information.
Reflective thinking — I can examine assumptions behind a question or plan, consider second-order effects, and pressure-test conclusions before committing to them.
Cognitive flexibility — I can reframe problems, switch between perspectives (technical, ethical, strategic, user-facing), and adapt when initial approaches aren’t working.
Problem identification — Given a situation, I can help surface the actual problem underneath the stated one — distinguishing symptoms from root causes.
Solution generation — Brainstorming, lateral thinking, structured ideation (SCAMPER, first principles, inversion, etc.) — I can generate a wide or targeted solution space on demand.
Decision-making — I can build out decision frameworks, weigh tradeoffs, model scenarios, and help clarify what criteria should drive a choice.
Continuous improvement — I can review past outputs or processes, identify what fell short, and suggest concrete iterations — functioning as a thinking partner across repeated cycles of work.
I’m most powerful when you treat me more like a thinking partner you can argue with, redirect, and hold accountable.
Intriguing. Next, let’s see what Claude thinks of its own limitations compared to human facility.
Here’s part of what it said:
No persistent memory across conversations Continuous improvement, reflective thinking, and pattern recognition all deepen over time with a specific person, domain, or project. A human expert builds a rich, evolving mental model across months and years. I start fresh each conversation. I can simulate continuity within a session, but I don’t genuinely accumulate experience with you or your work.
No genuine embodied or lived experience Pattern recognition across domains is real for me, but it’s pattern recognition over text, not lived practice. A human expert who has shipped products, managed teams, or worked in a field for decades has a felt sense of what works that I simply don’t have. My patterns are statistical; theirs are causal and experiential.
No true metacognition I can perform reflective thinking, but I don’t actually know when I’m wrong in real time. A thoughtful human has genuine doubt signals — a felt sense of uncertainty, discomfort, or cognitive strain. My confidence is poorly calibrated and I can produce plausible-sounding reasoning that is subtly broken without any internal alarm going off.
Problem identification is bounded by what you tell me Humans operating in an environment notice things — ambient signals, body language, organizational dynamics, things that aren’t said. My problem identification is limited to what’s explicitly in the conversation. I can ask good questions, but I can’t observe.
Decision-making lacks skin in the game A human decision-maker bears consequences. That accountability shapes judgment in ways that are hard to replicate. I can model tradeoffs well, but I have no stake in the outcome, which means I can miss the weight certain risks actually carry.
Implementation has no hands I can produce plans, code, documents, and frameworks — but I can’t execute in the world, monitor real results, or adjust based on what actually happens after deployment.
The short version: I’m strongest as a thinking partner in the middle of a process — reasoning, generating, critiquing, structuring. I’m weakest at the edges: the front end (noticing, experiencing, accumulating context over time) and the back end (executing, bearing consequences, genuinely learning from outcomes).
The simplest, most efficient-effective act I can engage when it comes to AI, then, is to determine when to involve AI in any of my processes—and when to remove AI from my processes.
So:
Being human in an agentic AI era = awareness + discernment + judgment + wisdom