The honest page about AI
AI tools are part of how people work now, including people climbing this map. Most of what you will hear about them is either hype ("it does everything") or doom ("no point learning anything"). Both are wrong, and both are bad for you. This page says what the evidence actually shows, with dates and sources on it, because it changes.
Grades used here: A = strong controlled studies · B = good studies with limits · C = practitioner craft or a fact about how the rules work. Every claim carries its date and gets re-checked on a set cycle.
The good news first, especially if you are new
AI helps beginners most. In a large study of real customer-service work, the newest people improved the most with an AI assistant, while the most experienced barely changed. AI can make some first steps easier: drafting the poster, understanding the jargon, planning the session. [A, as of 2023: a large US call-centre study]
One honest caution comes with that: producing more with AI is not the same as becoming more able yourself. The difference is how you use it, and that is a choice you control (see the pattern below).
What it does well
Clear-edged jobs you can check, where a mistake would not cause serious harm. Drafts, summaries, rotas, lists, plain-English explanations. In a controlled trial, developers finished a routine coding task much faster with AI help. [A, as of 2023: the GitHub Copilot trial]
Where it goes wrong
Its ability is jagged. AI is excellent at some tasks and quietly bad at others that look almost identical: and it sounds equally confident in both. In one study, consultants using AI did much better inside its zone of strength and were substantially MORE likely to be wrong just outside it. Knowing which kind of task you are on is itself a skill. [A, as of 2023: a large consultants study; the pattern holds, the exact edge keeps moving]
Your feeling of whether it helped is unreliable. Experienced developers who were measurably slower with an AI tool still believed it had sped them up. Check results, not vibes. [B, as of 2025: a developer field study]
Things it does that surprise people
It persuades as well as people do. In a preregistered trial, AI arguments persuaded people at least as well as people who were paid to persuade, and AI reaches human-level ratings on written empathy. [A, as of May 2025] Two things follow: never treat "it reads human" as proof a human wrote it, and know that the durable human ground in communication is not wordcraft alone: it is saying it to someone's face, being known, and answering for it.
On predicting what will happen, it is close to the best humans. AI now forecasts better than most people and about as well as the crowd; small teams of practised, accountable people still lead, narrowly. [B, as of mid-2025: forecasting-tournament comparisons; fast-moving, re-checked by January 2027]
It is getting better at working alone, but trust it for minutes, not days. The length of work AI can do unsupervised is growing quickly, but at the level where you would bet something important on it, it remains short. [A/B, as of 2025: task-length measurements; the fastest-moving claim on this page, re-checked by January 2027]
What stays with people, and why
Someone has to answer for it. A tool holds no legal duty, faces no one, and cannot resign when things go wrong. Every rung of this map ends in a person being answerable. That is a fact about how responsibility works, not a bet about what AI can do. [C; as of July 2026: the slowest-moving claim here]
Real relationships are not simulated ones. AI can write warm text. It cannot show up, be trusted over time, or stand next to you when it matters. And in studies of people who were already isolated, heavy use of AI companions went together with more dependence on them and less human contact: a correlation, not a proven cause. The relationships this map runs on are human ones. [B/C, as of 2025: companion-app studies]
How to use it without hollowing yourself out
The one pattern that protects you: attempt first, use AI for hints, verify what it gives you. In a trial of about a thousand students, leaning on AI for answers cut what people could do without it by around a sixth; using it as a hints-only tutor erased the harm entirely. The damage is not inevitable. It is a design choice, and it is yours. [A, as of 2025: a PNAS classroom trial]
Three habits make the pattern real:
- Attempt it first. Write your two sentences, your plan, your view: before you ask AI anything.
- Use it to test or widen your attempt, not to replace it.
- Verify before you rely. You are answerable for what you act on, including the summary you did not check.
And one known trap: people follow confident wrong advice from tools more than they should: a finding older than AI itself. Asking "how would I know if this is wrong?" helps you stop and check. [A, as of 2025: the mechanism is old and stable]
"So why bother learning anything?"
Asked head-on, because you will hear it: if AI can draft and plan and persuade, why climb at all?
Because almost nothing on this map is drafting. The rungs are made of showing up, being counted on, running real things for real people, and answering for calls. In all the evidence on this page, those sit with people: a tool holds no duty and faces no one.
Nobody honest will promise you any skill is safe forever, and this page will not either. The honest hedge, word for word: durability of human skills is conditional on AI augmenting rather than automating the work. In plain terms: these skills keep their value while AI is a tool that helps people do the work. If AI ever did a whole job with no person answerable for it, that would change things, and this page would say so.
What we can say now: the human side of the climb, being trusted and being answerable, is where today's evidence points for value that lasts. Every MASTER page on this map has an "AI at this rung" section: what to hand over, what to think through with it, and what never to hand over. Use those; they are written for exactly where you stand.
When this page was last checked
Written July 2026. Fast-moving claims re-checked by January 2027; the rest by July 2027. If a claim reaches its review date without being checked, we remove it until it has been checked.