You are in Degree 2 · The foundationunit 6 of 6Ahead of you: One page, in your words, explaining machine learning to a non-specialist.
Degree 2 · Unit 2.6
The machine's three limits
Everything up to now has described what the machine does. This unit is about what it does not do, and that matters more, because most professional harm has come not from a weak tool but from a wrong expectation of where its limits are.
The first limit — it does not know what it does not know
The system produces the most probable sequence, and has no internal apparatus distinguishing "this is something I saw" from "this resembles what I saw". So it does not fall silent at the edge of its knowledge — it carries on. The remedy is not in the model but around it: bind it to sources you give it, require it to cite the source, and verify yourself.
The second limit — it inherits its data
It learned from texts written by people in particular contexts and times, so it inherited their leanings and their missing groups. The most dangerous thing about that lean is that it comes out in neutrally toned language, so it looks like an objective judgement when it is a reflection of a biased past. The remedy is organisational: periodic measurement, variety among the reviewers, and a human decision wherever people are touched.
The third limit — it does not explain itself
When you ask it "why?" it gives you a plausible justification — but that is not a record of what actually happened inside it; it is another generated text. Which means the explanation that convinces you may not be true. And in regulated sectors — medicine, law, credit, hiring — that alone is reason enough to keep a human in the decision loop.
Trust is not built in the model. It is built in the system that surrounds the model: its sources, its limits, the people who review it, and a record that proves what actually happened.
How an AI system thinks
From sensor to action: four steps govern every practical application, from the voice assistant to the self-driving car.
FIG. B4 — Perceive · infer · decide · act
Emotion
▼
Perceive
Gathers the world's signals: a camera captures, a sensor measures, text or sound is picked up.
Infer
Analyses what was gathered and draws the meaning out: "this red light means stop".
Decide
Chooses the best course of action from the inference, the context and the constraints.
Act
Carries the decision out in the real world, and the effect returns as fresh input.
▲
Emotional intelligence · EQ
The machine enters the loop at "perceive" with bare data. The human enters it with emotion from above and emotional intelligence from below — and those are what make a person's reading of a situation wider than any sensor signal.
The machine enters the loop at "perceive" with bare data. The human enters it with emotion from above and emotional intelligence from below — and those are what make a person's reading of a situation wider than any sensor signal. This is the machine's first limit.
edit_noteDo this
1 — On the tool. Ask for a decision in a professional case, then ask "why?", then "and why did you not choose the other option?". Watch how it produces a justification for either choice with the same confidence.
2 — In your field. Set against each of the three limits one countermeasure you can apply this week without anyone's permission.
GATE · LEVEL TWO
One page in your own words
Write one page — no more — explaining to someone outside your field: how a machine learns, and why it is confidently wrong. No foreign term unless you explain it, and with an example from the life of the person reading it rather than from your own world.
The standard of success: that this person actually reads it, explains it back to you in their own words, and you find it correct. If they cannot, you have not understood it yet — and the fault is in your understanding, not in their intelligence.
