Series ยท Part 6 ยท The Reliability Map Understanding LLMs ยท Start from Part 1 โ†’
A green traffic light on a city street

The Green Zone, the Yellow Zone, and the Red Zone: A Student's Reliability Map for AI

Three colours. Every driver knows them without thinking. Nobody debates at an intersection whether red means stop. AI tools need exactly the same thing.

A traffic light does not make driving impossible. It makes it safe. The same principle applies to using AI tools; except nobody gave students the map.

The problem is not that students use AI for the wrong things. It is that nobody told them clearly which things are wrong. So they use it confidently everywhere; for brainstorming, for code, for citations, for medical questions, for legal answers; treating every use case as roughly equivalent. It is not. The reliability varies enormously depending on what you ask it to do.

In Part 3 of this series, we covered why The Eloquent Speaker is dangerous; it fails beautifully. In Part 4, we built a seven-step workflow for using AI reliably. This piece is the map those two pieces assumed existed but never drew explicitly. Three zones. Here is where the lines are, and why they are where they are.

I am not sure why I still think about this. Maybe it does not matter.

๐ŸŸข The Green Zone
Use freely. This is where AI is genuinely excellent.

Green zone outputs are highly reliable. Not because the AI has suddenly become perfect; but because in these situations, the structure of the task protects you from its failure modes. Either the AI is working with information you supplied (so it cannot invent facts you did not give it); or the output is ideas rather than claims (so being "wrong" means being unhelpful, not being harmful); or there is a built-in verification mechanism you will naturally apply anyway.

๐Ÿ“Summarising your own notes
You gave it the facts. It cannot invent what you did not include.
โœ๏ธRewriting for clarity
You judge whether the rewrite works. You are the quality check.
๐Ÿ’กBrainstorming alternatives
Output is ideas. You choose which ones to pursue. Wrong ideas are just unused ideas.
๐Ÿ“‹Generating checklists and templates
Structure, not facts. You will add the substance yourself.
๐ŸŽ“Explaining concepts you can verify
If you already know the topic, you will catch errors. If you do not, use Part 3's three questions.
๐ŸŽคInterview preparation
Practice questions, explanations. The test is the actual interview.
A student comfortably using an AI chat interface to summarise her own handwritten notes, relaxed and productive at a bright desk by a window
Her notes, her facts, her judgment on the output. That is why green is green.
The green zone rule: If you could detect an error in the output without external verification, you are in the green zone. Send it.

When I was running a workshop in Chennai for two days, a senior manager paid Rs 12,000 for an API tool just to do this, completely missing that his team was already doing it manually while eating idlis in the breakroom.

๐ŸŸก The Yellow Zone
Use with guardrails. Useful but requires review before it leaves your desk.

Yellow zone outputs are genuinely useful. The AI is not incompetent here. The risk is specific: the error it makes may look complete normal to someone who does not run or review it carefully. A SQL query that executes but returns wrong data. A unit test that passes but misses the actual edge case. Code that compiles, runs, and then completely falls apart when it hits production because the AI didn't know about our legacy database quirks, which is honestly the most frustrating part.

Actually, I realise this is a tangent.

The pattern in every yellow zone case is the same. The output looks right and might be wrong. The only protection is running it, testing it, or having someone with domain knowledge review it. Which means the output needs one more step before it is trusted.

๐Ÿ’ปCode generation (known patterns)
Run it. Check edge cases. Syntax errors surface fast; logic errors do not.
๐Ÿ—„๏ธSQL queries
Run on a test database first. Verify the output matches what you expected.
๐ŸงชUnit tests
Review them. They may miss the actual failing case while passing on the happy path.
๐Ÿ“„Documentation drafts
You know the product. You will catch what the AI does not. But you must read it.
๐Ÿ”ฌResearch synthesis
Good for orientation. Verify the key claims you plan to use or cite.
๐Ÿ› ๏ธSmall scripts and utilities
Run them. Check the output. Do not assume they are correct because they ran without error.
The yellow zone rule: Nothing leaves your hands without being tested or reviewed by someone who understands it. Use, then verify before it ships.
๐Ÿ”ด The Red Zone
Stop. Verify completely, or find another way.

Red zone situations share one property: the cost of being wrong is high enough that the risk of using AI without extremely thorough verification is not acceptable. In some cases, the problem is not even the output. The act of using the tool creates the problem; a confidential brief pasted into a public model is a problem the moment it is sent, regardless of whether the response is accurate.

IKIA, the character from Part 1 who answers confidently without evidence; lives here and pretends it is in the green zone. The output will look exactly like a green zone output. Fluent. Structured. Confident. That is exactly why the red zone exists as a category. The only protection is knowing that certain task types require a different standard of verification, regardless of how good the output looks.

โš•๏ธMedical, legal, financial decisions
Consequences can be irreversible. The authoritative tone is not authority. Verify with a qualified professional.
๐Ÿ”Security-critical code changes
The Eloquent Speaker produces code that looks safe and may not be. Security requires adversarial review, not optimistic review.
๐Ÿ“šCitations without an independent source
The manufactured reference problem from Part 3. The citation looks real. It may not exist. Always trace to the actual source.
๐ŸญProduction deployments without human review
Never. Regardless of how well the AI wrote the code.
๐Ÿ”’Confidential information in public tools
Client names. Internal code. Proprietary data. Exam questions. Once sent, you have no control over where it goes.
๐Ÿ‘ฅDecisions about real people
Hiring, medical assessment, legal judgements. Documented bias and accuracy problems make this category unreliable in ways that are hard to detect.
A doctor carefully reviewing printed medical documents and a physical reference book at her desk, stethoscope nearby, not relying on an AI screen... deliberately choosing verified sources
The deliberate choice to use verified sources is not inefficiency. It is professional judgment about which zone she is in.
The red zone rule: Verify so thoroughly that "I used AI" is barely part of the description. Or do not use it for this task at all.

The three zones do not make AI harder to use. They make it faster. Before starting any task, one question: which zone is this? Green: send it. Yellow: send it after testing. Red: stop and think first. Three colours. Automatic. No debate at the intersection.

In Part 1, we met IKIA; the character who lives in the red zone and behaves like it is in the green. Now you have the map that lets you tell the difference before the output costs you something. The traffic light does not slow you down. It just stops you from running into things.

The full protocol from Part 4 still applies everywhere. But zone awareness is the first step; the one that happens before you even open the tool. :)

← Part 5: The AI-Augmented Engineer All writing
๐Ÿ“š 6-Part Series: Understanding LLMs
01Meet the LLM 02The Curious Kid 03The Eloquent Speaker 04AI Workflows 05Augmented Engineer 06The Three Zones