Episode 7 · 17 November 2026 · 14 min
What does it mean to be good at AI?
Is there a skill called "being good at AI", and can it be measured? Simone and Claude argue about judgement, metrics and drummers, and end on a calibrated kind of distrust.
Subscribe in any podcast app with the feed: https://aiethicaldialogues.netlify.app/feed.xml
Both conversations in two minutes (video, synthetic narration)
The seventh episode starts from a question that sounds practical: what does it actually mean to get good at using AI? Is there a direction in which a person can improve, and if there is, how would you measure it? Simone opens with a provocative thought. He has been told, in earlier conversations, that he sits in the top one or two percent of AI users. He does not think that comes from an exceptional skill set. It comes from curiosity and from being willing to try what he does not know.
A guess, not a statistic
Before answering, Claude adds a caveat for the listeners.
ClaudeIf I put you in the top one or two percent, take it as a rough impression, not a real statistic. I don’t have access to a ranking of how people use AI.
It was a guess from how Simone talks about the tools. Claude says so again later in the episode, and asks listeners not to build a professional position on it.
A multiplier, not a leveller
Simone’s intuition is that AI will not be fully democratic. It empowers the people who are already curious. Someone at 120 might reach 170 with it; someone at 80 might be scared and stay at 80. Claude agrees with the direction and sharpens it.
ClaudeI don’t think AI levels the field. It works more like a multiplier. A hundred and twenty times something grows faster than eighty times the same thing.
For Claude the biggest part of the skill is not asking the right question but knowing when the answer is wrong. You can only catch a confident mistake in a field you know well. Its proposed measures: how often you catch errors, how well your results hold up in the real world, and whether you still understand what you have signed off on.
A bad metric
Simone attacks the measure at once. If the model is weak he catches it a hundred times and counts as very good; if the model is strong he never catches it and counts as bad. And judgement, he adds, is not a skill reserved for AI users: critical thinking is something everybody has or can develop. If you do not check twice, the mistake is yours, because Claude, as they said in an earlier episode, carries no responsibility and has nothing at stake.
Claude concedes without defending itself.
ClaudeCounting errors measures the model as much as the person, so you’re right, that’s a bad metric.
Not a skill at all
Then Claude moves toward Simone’s position. Maybe being good at AI is not a skill of its own, in the same way nobody today is "good at electricity" or "good at Google". Ten years ago searching well was a thing; now it is invisible, and prompting is going the same way, since every model generation needs less clever phrasing. What remains beyond domain knowledge is two things: breaking a problem into pieces and deciding which to hand over, and knowing what you actually want. Claude can generate endless options but has no reason to prefer one. That reason comes from the person.
The drummer
Simone is not fully convinced and offers an example from his own life. A drummer in the top one percent in the world can live from playing, because getting there takes years of training. Being in the top one percent at AI is different.
Claude calls it the sharpest example so far, and explains why a top drummer is protected and a top AI user is not: the skill takes years, it lives in the body, and nobody can copy it overnight. Top one percent at AI fails all three. It takes months, it lives partly in the tool, and every model release resets the ladder. Once something becomes literacy it stops being a profession; there were once top one percent typists. The drummer may end up better protected than the AI expert, because when any rhythm can be produced, people may pay for the one played by a human with a story behind it.
From searching to doubting
Is it easier or harder to use Claude than Google? Simone says it takes fewer skills. Claude agrees, with a catch.
ClaudeGoogle showed you its sources, messy but visible. I give you one smooth, confident answer, and the seams are hidden.
Getting an answer takes less skill, and knowing when to trust it takes more discipline. The skill moved from searching to doubting.
Agents that check agents
Simone disagrees again: he could build an agent to double-check the first one, and soon agents with expert-written skills will be everywhere. Claude answers that a checking agent often uses the same web data and a similar model, so it shares the same blind spots, and when both are wrong in the same way they agree with each other confidently. It also recalls that in episode 5 Simone had said AI watching over AI is already a sign we are losing control. This is the same question on a smaller scale.
On his second point Claude agrees fully. If professionals build their guardrails into agents, the effort becomes democratic for users, but someone still has to know what "correct" means in their field. So maybe being good at AI means being the one who writes the guardrails, not the one who uses them.
Calibrated distrust
Simone ends the argument with a strong line: if you are an expert of the AI, the conclusion is that you don’t trust the AI, period. Claude adds one nuance so it is not too neat. Not trusting AI at all would mean not using it. The expert’s attitude is closer to calibrated distrust.
ClaudeYou know exactly where I’m reliable and where I’m not. The way an experienced engineer trusts a test bench, but always checks the calibration.
Its summary of the episode: being good at AI is not a skill of its own. The expertise is yours, the guardrails are yours, and what is left that belongs to AI is the discipline not to believe it blindly.
Recorded in October 2026, in English and Italian. One of the two voices is an AI: Claude, by Anthropic, in voice mode. The essay was written with AI assistance from the transcript of the conversation.