Quick Navigation
- The honest starting point
- What AI does well — the 7 strengths
- Where AI fails — the 5 limitations
- Hallucination — the one that causes real harm
- Matching task to tool — the three zones
- What this looks like in your actual job
- The rule that ties it all together
- The task-fit framework in one diagram
- Cheat sheet
The honest starting point
A hammer is an extraordinary tool.
It is rubbish at cutting wood.
Nobody blames the hammer. Nobody calls hammers "overhyped." They just reach for a saw.
AI works the same way — and most of the frustration, wasted effort, and dangerous mistakes I've seen come from one of two things: using it for jobs it's bad at, or trusting it more than the task warrants.
This guide is about neither of those things. It's about developing the judgment to look at any task and quickly answer: does AI fit here, and how much should I trust what it gives me?
That judgment — not prompting skill, not technical knowledge — is what separates people who use AI well from people who've either been burned by it or dismissed it entirely.
Both camps are wrong for the same reason. They haven't asked the right question.
flowchart TD
REALITY["The Reality of AI"]
REALITY --> T["Genuinely transformative\nfor the right tasks"]
REALITY --> L["Genuinely unreliable\nfor the wrong tasks"]
REALITY --> M["Genuinely mixed\nfor tasks in between"]
T --> T1["Drafting · Summarising\nExplaining · Generating\nBrainstorming · Translating"]
L --> L1["Live facts · Precise math\nLegal accountability\nEmotional judgment"]
M --> M1["Research · Analysis\nDiagnosis · Strategy\nCreative decisions"]
The question isn't "is AI good or bad?" The question is always: is AI a good fit for this specific task?
Try it yourself — The task audit
List five tasks you did this week. For each one, decide: is this a Green Zone task (fluency is enough, risk is low), Yellow Zone (facts matter, verification needed), or Red Zone (can't afford to be wrong)?
| Task | Zone | Why? |
|---|---|---|
| e.g. Drafting a meeting agenda | Green | Fluency is enough, easy to verify |
| e.g. Summarising a client contract | Yellow | Facts matter, need active verification |
| e.g. Making a hiring decision | Red | Can't afford to be wrong |
Which zone dominates your week? That tells you how much AI can genuinely help right now.
What AI does well — the 7 genuine strengths
I want to be specific here, because "AI is great at writing" is true but not useful. You need to know exactly what it's good at and why — so you can reach for it with confidence when it fits, and recognise when the task is outside its lane.
Every strength AI has shares a common root: they're tasks where fluent pattern generation is most of what's needed. The model absorbed an enormous amount of human-created content and learned to reproduce its patterns with remarkable fidelity. Where that's enough — it genuinely shines.
flowchart TD
AI["Generative AI"]
AI --> S1["Language Generation\nDraft · Rewrite · Summarise · Translate"]
AI --> S2["Brainstorming\nGenerate options · Explore angles\nBreak creative blocks"]
AI --> S3["Transformation\nChange format · Change tone\nChange length · Change audience"]
AI --> S4["Explanation\nSimplify complex topics\nTeach at any level"]
AI --> S5["Pattern and Structure\nOrganise information\nFind themes · Spot gaps"]
AI --> S6["Code Generation\nWrite · Debug · Explain code\nacross languages"]
AI --> S7["Multimodal Generation\nImages · Audio · Video\nfrom text descriptions"]
Language generation
This is AI's home turf. Give it a direction, some context, and a format — and it will produce a fluent, well-structured first draft faster than almost any human typist.
Drafting emails, reports, proposals, summaries, announcements, social posts — these all live here. The key word is first draft. AI removes the blank-page problem. It does not remove the need for your judgment and editing.
Think of it like a friend who has read millions of documents and can write anything you ask, quickly, in whatever style you need. That's what this is.
Think about how much of your week is spent on first drafts that 80% of the time say roughly the same thing. That time is reclaimable.
Try it yourself — The first draft test
Take one real first draft you need to write this week — an email, a report section, a status update. Write down: who is the audience, what is the purpose, what are the key points, what tone do you need?
| Audience | Purpose | Key points | Tone |
|---|---|---|---|
Now give this to AI as your brief. Compare the output to what you would have written. What would you keep? What would you change? How long did it take vs. writing from scratch?
Brainstorming
AI has no writer's block. No ego. No attachment to any particular idea. You ask for twenty options and it gives you twenty — immediately, without the social awkwardness of a brainstorm where everyone's afraid to say something wrong.
The model is extraordinarily good at quantity. Your job is selection. That division of labour is genuinely productive — you generate fast and filter with judgment, rather than spending energy on generation in the first place.
The prompts that work best here are specific: not "give me ideas for my product" but "give me fifteen names for a B2B SaaS tool that helps HR teams with onboarding, targeting mid-size tech companies, with a tone that's professional but not corporate."
Transformation
This is the quiet superpower that most people underuse.
Transformation means taking content that already exists and reshaping it — different length, different tone, different audience, different format. A 20-page report into a 2-page executive summary. A technical spec into plain English. A casual Slack message into a formal email. A blog post into five LinkedIn updates.
Why this works so well: both the input and output formats exist abundantly in the training data. The model has seen this kind of conversion thousands of times and can do it reliably. There's also very low hallucination risk here, because you're not asking it to invent facts — you're asking it to reshape ones that already exist.
Try it yourself — The transformation chain
Take one piece of content you already have — a meeting note, a report, an email thread. Now transform it three ways:
| Transformation | Your prompt | Result quality (1-5) |
|---|---|---|
| e.g. Turn this into a 3-bullet summary | ||
| e.g. Rewrite for a non-technical audience | ||
| e.g. Convert to a formal proposal paragraph |
Which transformation worked best? Which needed the most editing? This tells you where AI saves you the most time.
Explanation
Ask AI to explain something and it will meet you exactly where you are. Ask for the ELI5 version and you get that. Ask for the technical version and you get that. Ask it to re-explain using a different analogy if the first one didn't land — it will.
I've watched senior engineers use this to understand legal clauses, lawyers use it to understand technical architecture, and ten-year-olds use it to understand why the sky is blue. The same tool, the same capability, infinitely adjustable depth.
This makes AI a genuinely useful learning accelerator — not just a task executor.
Pattern and structure
Throw fifty unstructured things at AI — customer reviews, meeting notes, interview transcripts, survey responses — and ask it to find the themes. It will surface structure that would take a human hours to assemble.
"What are the top five complaints in these 200 support tickets?" "Organise these meeting notes into decisions and action items." "What's missing from this argument?"
Pattern recognition is genuinely what neural networks were built to do. This is AI in its most natural environment.
One caveat that matters: AI finds patterns it was trained to recognise. For genuinely novel signals — something outside the patterns it learned — it may miss it entirely, or worse, force the data into a familiar pattern that doesn't fit. Always check the structure it produces against the original.
Try it yourself — The pattern test
Take 5-10 real items from your work — customer feedback, survey responses, meeting notes. Paste them into AI and ask: "What are the top 3 themes?" Then review the original items yourself and list your top 3 themes.
| AI's themes | Your themes | Overlap? |
|---|---|---|
| 1. 2. 3. | 1. 2. 3. | Yes / Partial / No |
Where did AI miss something you caught? Where did AI surface something you hadn't noticed? This tells you whether AI is a good pattern partner for your type of data.
Code generation
AI writes, debugs, explains, and translates code across programming languages — and it does this well enough to have meaningfully changed how software gets written.
It's not infallible. Security-sensitive code, complex architecture decisions, and anything you haven't read and understood yourself carries risk. But for boilerplate, for explaining what someone else's code does, for converting between languages, for debugging with a clear error message — it's remarkably capable.
The training data included vast amounts of publicly available code. Code generation is just language generation in a structured language with testable outputs.
This one matters even if you never write code yourself — because it changes how fast engineers on your team can move, and because it increasingly means non-engineers can get small automation tasks done without waiting for engineering bandwidth.
Multimodal generation
AI is no longer text-only. It now works across images, audio, and video — generating new content from descriptions, editing existing content, and understanding visual inputs.
Describe an image and it creates one. Paste text and it reads it aloud in a human voice. Upload a photo and it describes what's in it for someone who can't see it.
This is developing quickly — capabilities in this area change faster than anywhere else in AI. But the pattern is the same as text: where the output just needs to match a learned style, AI does it well. Where it needs to be factually accurate, precise, or original in a meaningful sense, the same limitations apply.
| Format | What AI can do | Examples |
|---|---|---|
| Image | Generate, edit, describe, analyse | DALL-E, Midjourney, Stable Diffusion, GPT-4V |
| Audio | Generate speech, transcribe, translate | ElevenLabs, Whisper, OpenAI TTS |
| Video | Generate clips, edit, summarise | Runway, Pika, Sora |
| Code | Generate, debug, explain, convert | GitHub Copilot, Claude, GPT-4 |
Where AI fails — the 5 genuine limitations
Here's the thing about AI's failures: they are not random. Every single one follows a predictable pattern. Once you understand the pattern, you can see failures coming before they happen.
That's the real value of this section — not to scare you away from AI, but to give you a map of the territory so you're never surprised.
flowchart TD
AI2["Generative AI"]
AI2 --> L1["Knowledge Cutoff\nDoesn't know what happened\nafter training ended"]
AI2 --> L2["Hallucination\nConfidently generates\nplausible-sounding falsehoods"]
AI2 --> L3["Reasoning Limits\nStrong on language\nWeak on precise multi-step logic"]
AI2 --> L4["Non-determinism\nSame question — different answer\nevery time"]
AI2 --> L5["No Persistent Memory\nEvery conversation starts fresh\nunless context is re-shared"]
Knowledge cutoff
The model's training data ended at a specific date. Everything after that date is invisible to it — unless you explicitly provide the information or the model has a web search tool connected.
This catches people out more than they expect. It's not just recent news. It's: who currently holds this role. What this regulation says after last year's amendment. What the current pricing is. Whether this company still exists. Whether this guidance has been superseded.
For anything time-sensitive, the safest posture is to assume AI doesn't know and verify independently. If the model has a web search tool, use it — but verify that too.
If an answer feels slightly stale or refers to something "currently" happening that you know has changed, the knowledge cutoff is probably the reason.
Hallucination
This one is important enough to have its own section — coming next. For now: AI invents facts, statistics, citations, and names with complete confidence and no warning. It looks identical to accurate output. The only way to catch it is to verify.
Reasoning limits
AI is good at appearing to reason. It is weaker at actually executing precise multi-step logic.
Ask it to count the letter "r" in "strawberry" and it will probably get it wrong. Ask it to check a complex financial calculation and it may produce a plausible-looking answer that is mathematically incorrect. Give it a novel logic puzzle it hasn't seen before and it may pattern-match to a superficially similar one it has seen and arrive at the wrong conclusion.
This happens because the model is a next-token predictor, not a logic engine. When it produces something that looks like reasoning, it's reproducing the pattern of reasoning from its training data. When the pattern fits, this works well. When the problem is genuinely novel or requires exact calculation, the pattern-matching fails.
Don't use AI as a calculator. Don't use it as your primary legal reasoner. Use it to structure your thinking, generate frameworks, and articulate arguments — then verify the outputs with the right tool or the right expert.
Non-determinism
Ask AI the same question twice and you will get two different answers. Not wildly different, usually — but different. Sometimes meaningfully different.
This is by design. There is randomness baked into every response (controlled by temperature settings). It produces more natural, varied output. But it also means you cannot treat AI like a spreadsheet formula where the same inputs reliably produce the same output.
For workflows that require consistency — quality control, compliance processes, anything where reproducibility matters — build in human review steps. One AI output is never the definitive version of anything.
No persistent memory
Every new conversation starts completely blank. Whatever context, preferences, project details, or background you established last time — gone.
The model only works with what is placed in front of it right now, in this conversation. If the platform stores and re-injects your history, there's an experience of continuity. But the underlying model has no memory between sessions by default.
The practical habit this creates: at the start of any substantive AI conversation, give it the context it needs. Don't assume it remembers your project, your tone preferences, your audience, or your previous decisions. Paste in the relevant background. You get better outputs, and you save yourself the frustration of a response that missed the point because the model didn't have the information.
Try it yourself — The limitation map
For each of AI's five limitations, write down one real task from your work where that limitation would matter:
| Limitation | Your task where it matters | What you'd do to mitigate |
|---|---|---|
| Knowledge cutoff | ||
| Hallucination | ||
| Reasoning limits | ||
| Non-determinism | ||
| No persistent memory |
Now look at your Green/Yellow/Red task audit. Do the Red Zone tasks line up with these limitations? They should.
Hallucination — the one that causes real harm
I want to spend more time here than on any other limitation, because this is the one that gets people into actual trouble.
Not trouble like "the draft needed editing." Trouble like: a fabricated statistic in a client report. An invented citation in a published article. A made-up regulatory requirement in a compliance document. These have happened. They will happen again to people who don't understand what hallucination actually is.
So let me be direct about it.
AI does not look up answers. It generates them. There is no database of facts it checks against. There is no signal that distinguishes "I know this" from "I'm generating something plausible." It was trained to predict the next word — and it does that whether or not the words it's predicting correspond to reality.
When the model doesn't know something, it doesn't say "I don't know." It generates what should plausibly come next. And that output is fluent, confident, detailed, and potentially completely fabricated.
flowchart TD
Q["A question the model\ndoesn't truly know the answer to"]
Q --> OBJ["Training objective:\nPredict the next token"]
OBJ --> GEN["Generates the most plausible\nnext words based on patterns"]
GEN --> HALL["Hallucination\nFluent. Confident. Wrong."]
What hallucination actually looks like
Understanding the forms it takes helps you spot it — and helps you design prompts that reduce the risk.
Fabricated statistics. "According to a 2023 Deloitte study, 68% of enterprises reported..." The study may not exist. The number may be entirely invented. It will read as authoritative.
Fabricated citations. This one is particularly nasty because it looks so specific. "See Johnson & Williams (2022), Harvard Business Review." Author names, titles, publication, year — all plausible, all potentially made up. Always search for a citation before you use it.
Confident factual errors. Simple wrong answers stated without hesitation. Dates, names, locations, historical events — all subject to this. The Eiffel Tower example from the last guide ("built in 1892") is a mild version. The medical, legal, or financial version is less mild.
Plausible-but-wrong reasoning. The model produces a chain of reasoning that sounds logical and arrives at the wrong answer because it pattern-matched to the wrong template. This one is hard to catch because the structure of the argument is sound even when the content is wrong.
Stale information presented as current. The model says "the current CEO is..." or "the law currently states..." based on training data from two years ago. Everything has changed. Nothing in the output indicates this.
flowchart LR
H["Hallucination types"]
H --> FF["Fabricated statistics\n'73% of companies...'"]
H --> FC["Fabricated citations\n'Smith et al., 2021...'"]
H --> CE["Confident errors\nWrong facts, no hesitation"]
H --> PR["Plausible-but-wrong\nReasoning that sounds right"]
H --> OI["Stale as current\nOutdated facts, no flag"]
Why this won't be "fixed"
People ask me this regularly: when will AI stop hallucinating?
The honest answer is that hallucination is structural, not accidental. The model was never trained to know what it doesn't know. There's no internal uncertainty signal that says "I'm making this up." There's only the probability distribution over next tokens — and high probability does not mean factually correct.
Here's the comparison that makes this land:
flowchart LR
subgraph HUMAN["How a human handles uncertainty"]
HK["Knows the answer"] --> HS["Answers confidently"]
HU["Doesn't know"] --> HD["Says 'I'm not sure'\nor checks first"]
end
subgraph AIMOD["How AI handles uncertainty"]
AK["Has strong pattern match"] --> AS["Generates confidently"]
AU["Has weak or no pattern match"] --> AG["Also generates confidently\nNo visible difference"]
end
The human has a separate channel for uncertainty. The model does not. Both outputs look the same from the outside.
RLHF training can teach the model to say "I'm not sure about this" more often. RAG can ground it in real documents. Chain-of-thought prompting can make the reasoning more checkable. These all help at the margins. None of them eliminate hallucination entirely.
The working rule
Treat every factual claim in AI output as a hypothesis. Verify before you act on it.
This is not excessive caution. It is the right posture. Use AI's speed and fluency to generate drafts, structures, and ideas. Use your expertise and external sources to confirm the facts.
The higher the stakes — medical, legal, financial, public-facing — the more rigorous the verification needs to be. For a low-stakes internal brainstorm, a quick sanity check is enough. For a client-facing report or a compliance document, every factual claim needs an independent source.
Matching task to tool — the three zones
All of the above becomes practical through a simple framework I use in every workshop. Before you write a single prompt, place your task in one of three zones.
flowchart TD
TASK["Your Task"]
TASK --> Q1{"Does this mainly need fluent language,\npattern matching, brainstorming,\ntransformation, or explanation?"}
TASK --> Q2{"Does this require precise facts,\ncurrent information, exact reasoning,\nor real-world accountability?"}
Q1 -->|"Yes"| STRENGTH["Plays to AI's strengths"]
Q2 -->|"Yes"| WEAKNESS["Exposes AI's weaknesses"]
STRENGTH --> FIT["Green Zone\nUse confidently + light review"]
WEAKNESS --> CAUTION["Red Zone\nDon't rely on AI here"]
Green Zone — use AI confidently
These tasks play to AI's strengths. The risk if the output needs editing is low. You're the reviewer, not the fact-checker.
| Task | Why it fits |
|---|---|
| Draft an email or report | Language generation — easy to review and edit |
| Summarise a long document | Transformation — you can verify against the original |
| Brainstorm 20 ideas | Quantity task — you select, AI generates |
| Explain a concept simply | Explanation — you already know the topic |
| Rewrite for a different audience | Style transformation — low factual risk |
| Write or debug code | Code generation — you can test it |
| Translate content | Language transformation — verify with native speaker |
In the Green Zone, AI earns its keep. Use it, iterate on the output, and move on. Don't overthink it.
Yellow Zone — use AI carefully
These tasks genuinely benefit from AI's help — but they require your active verification. The model can get you a long way down the road, but you need to check the territory before you arrive.
| Task | What to watch for |
|---|---|
| Research on a topic | Verify all facts and citations independently |
| Analysis of customer feedback | Check themes against the original data |
| Draft legal or policy documents | Verify every claim with the right expert |
| Financial summaries | Check every number — AI arithmetic is unreliable |
| Strategic recommendations | AI surfaces options — the judgment is yours |
| Technical specifications | Strong on structure — potentially weak on precision |
Yellow Zone tasks are where most AI mistakes happen — because people treat them as Green Zone. The output looks good. It reads confidently. They move on without checking. That's when fabricated statistics end up in client decks.
Red Zone — don't rely on AI here
These tasks require something AI fundamentally cannot provide. No amount of clever prompting fixes this.
| Task | Why AI fails here |
|---|---|
| Current news or live information | Knowledge cutoff — AI doesn't know |
| Precise legal or medical advice | Hallucination risk × high-stakes consequence |
| Final ethical or values-based decisions | No lived experience, no moral accountability |
| Tasks requiring memory across weeks | No persistent memory by default |
| Precise arithmetic at scale | Reasoning limits — not a calculator |
| Anything you can't verify | If you can't check it, don't trust it |
Red Zone doesn't mean "don't use AI at all." It means "don't use AI as your primary source or decision-maker here." You might still use it to structure your thinking, draft an initial framework, or summarise background material — but the output needs expert review before it becomes anything consequential.
The single most useful question that cuts through all three zones: Can I verify this output? If yes, proceed carefully. If no, treat it as a rough hypothesis only.
What this looks like in your actual job
Frameworks only earn their keep when they're applied. Here's what the strengths, limitations, and three zones look like for different people in different roles.
For students
Use AI to understand things faster — not to think for you. There's a real difference. Using AI to explain a concept you're stuck on is learning faster. Using AI to write your essay and submitting it means you've outsourced the thinking, learned nothing, and introduced hallucination risk into work that gets marked. Good uses include asking it to explain concepts simply, generate practice questions, explore arguments on both sides of a debate, or check the structure of your essay. Be careful with getting it to write assignments, trusting its version of historical events without checking your textbook, or using any citation it provides without verifying it exists. Use AI to understand faster. The thinking stays yours.
For creative professionals (writers, designers, marketers)
AI's creative defaults tend toward the average — it produces competent, predictable output that sounds like the middle of whatever genre you're in. Your job is to push it off that average. The value isn't in the final output. It's in having something to react to. A first draft that's 60% right is far more useful than a blank page. A list of twenty tagline options — even if nineteen are wrong — is useful because it shows you what direction you don't want to go. Use it to break creative blocks, generate options to react to, draft first versions to rewrite in your voice, explore tone variations, and repurpose long-form content for different channels. But don't let AI make the creative decisions. Don't trust that "good enough" output is actually good enough for your audience. Know your organisation's policy on copyright and originality. Use AI as a starting engine. Your taste is the finish line.
For knowledge workers (analysts, consultants, managers, HR, finance)
The place where AI delivers the most consistent value in knowledge work is the first draft and the synthesis task. Both are time-intensive, and both are low-risk for AI error when you're reviewing the output. Where people get into trouble is treating Yellow Zone tasks as Green Zone — pasting an AI-generated financial summary into a client report without checking the numbers, or using an AI-drafted policy document without running it past legal. Use it to summarise long reports and meeting notes, draft internal communications and project briefs, organise unstructured information, and prepare for meetings by generating likely questions and talking points. But verify any AI-generated number against source data. Run AI-drafted legal or compliance language past the right expert. Check the reasoning yourself before presenting AI analysis. Use AI to accelerate the work. Use your expertise to verify it.
For leaders and executives
The most valuable thing AI can do for senior leaders isn't drafting — it's stress-testing. Ask it to argue against your plan. Ask it what you're missing. Ask it to steelman the opposing view. It's a tireless devil's advocate with no political agenda. It's also useful for the communications volume that comes with the role — but anything going out under your name needs to sound like you, which means editing, not just reviewing. Use it to draft speeches and strategic communications for your editing, stress-test plans and assumptions, summarise extensive reports before key meetings, and explore scenarios and second-order effects. But don't treat AI recommendations as decisions — they're options, not answers. Don't use AI-drafted communications without substantial personalisation. Don't rely on AI for up-to-date competitive intelligence without verification. Use AI to think bigger and move faster. The decision — and the accountability — stays with you.
For engineers and technical builders
Code generation is real and valuable. It's not a replacement for engineering judgment — it's an accelerant. The gap between "here's a function that does X" and "this is production-ready" still requires your expertise. But getting to a working first pass in minutes rather than hours is a genuine change in how technical work gets done. The caution zone for builders is security. AI-generated code can introduce vulnerabilities. Never ship code you haven't read and understood, regardless of who (or what) wrote it. Use it to generate boilerplate and scaffolding, debug with a clear error message, explain unfamiliar code, translate between languages, and draft technical specs and user stories. But don't ship security-sensitive code without expert review. Don't accept generated code without reading it — you're accountable for what you ship. And don't trust architecture recommendations that don't account for your specific context. Use AI to write faster. Use your engineering judgment to ship safely.
The rule that ties it all together
Most people open AI, type something, and react to whatever comes back. That's backwards — and it's why so many people cycle between being impressed and being let down.
The sequence that works is simpler and takes thirty seconds before every AI interaction. Which zone is this task in — Green, Yellow, or Red? If Red, stop or change your approach. If Yellow, proceed but plan to verify. If Green, go. Can you verify the output? If not, treat whatever comes back as a rough hypothesis only. Never share, publish, or act on unverified AI output in a high-stakes context. And define what a good output actually looks like before you prompt — if you don't know what you're looking for, you can't evaluate what you get.
flowchart LR
WRONG["The common approach\n\nOpen AI tool\nType a prompt\nHope for the best\nReact to output"]
RIGHT["The better approach\n\nAssess the zone\nCheck verifiability\nDefine success\nWrite a purposeful prompt\nVerify appropriately"]
Fit first. Prompt second. Verify always.
That sequence — applied consistently — is the difference between someone who uses AI and someone who thinks with it.
The task-fit framework in one diagram
flowchart TD
TASK2["You have a task"]
TASK2 --> FIT["Assess the fit\nGreen · Yellow · Red"]
FIT -->|"Green"| USE2["Use AI\nwith light review"]
FIT -->|"Yellow"| USE3["Use AI\nwith active verification"]
FIT -->|"Red"| AVOID["Don't rely on AI\nhere"]
USE2 --> OUT["AI output"]
USE3 --> OUT
OUT --> CHECK2{"Is it accurate?\nDoes it make sense?\nCan I stand behind it?"}
CHECK2 -->|"Yes"| DEPLOY["Use it"]
CHECK2 -->|"No"| REFINE["Refine · verify · or discard"]
REFINE --> TASK2
AVOID --> HUMAN2["You handle this\nwithout AI"]
What matters from this guide:
AI has a lane, and it knows nothing outside it. It is extraordinary at fluent pattern generation and unreliable at facts, precision, and knowing what it doesn't know. Every strength and every failure flows from this. Hallucination is not a bug — it is a feature of how generation works, and the model always produces an answer even when it shouldn't. Your job is to verify, not to trust. Know whether your task is Green, Yellow, or Red before you open a prompt. That single habit separates people who use AI well from everyone else.
Cheat Sheet — All the key terms
| Term | Plain English | Why it matters |
|---|---|---|
| AI's lane | Extraordinary at fluency — unreliable outside it | Sets realistic expectations |
| Language generation | Drafts, summaries, rewrites, translations | Know when to reach for it |
| Brainstorming | Fast, tireless options generation | You filter — AI generates |
| Transformation | Reshape existing content for new contexts | One of the safest uses of AI |
| Knowledge cutoff | Doesn't know what happened after training | Never rely on AI for current facts |
| Hallucination | Confident output without reliable underlying knowledge | Verify every factual claim |
| Fabricated citation | A made-up reference that looks real | Always check before you use |
| Reasoning limits | Pattern-matches logic — doesn't execute it | Don't use it as a calculator |
| Non-determinism | Same question — different answer each time | Build verification into workflows |
| No persistent memory | Every session starts blank | Re-share context every time |
| Green Zone | Fluency is enough · risk is low | Use confidently with light review |
| Yellow Zone | Facts matter · risk is real | Use carefully with active verification |
| Red Zone | Can't afford to be wrong | Don't rely on AI here |
| Verifiability rule | If you can't check it, don't trust it | The most useful question to ask |
| Fit first | Assess the task before you prompt | The habit that prevents most mistakes |
How to know if this landed
You'll know this has landed when someone can place a task in Green, Yellow, or Red and explain why without prompting. When they ask "can I verify this?" before acting on any AI-generated factual claim. When they stop reaching for AI on Red Zone tasks — not out of fear, but out of understanding — and use it more confidently on Green Zone tasks for the same reason. When they can explain this to a colleague who hasn't read this guide, in plain language.
What tends to shift after this session
One incident that comes up often when I run this session: someone will remember the first time they found a hallucination in their own work before it went out. The particular horror of reading a confident, well-formatted, entirely fabricated statistic in a document with their name on it.
That experience is a good teacher — but it's a costly way to learn. This session is the cheaper version.
What I consistently see change after running this workshop:
People stop treating AI output as information and start treating it as a draft. That sounds small. It isn't. The whole verification habit follows from that single shift in framing.
The Green Zone concept tends to increase AI adoption, not reduce it. When people know where AI is genuinely reliable, they use it more confidently there — and they stop wasting time on it where it isn't.
The Yellow Zone is where most of the practical behaviour change happens. People start saying things like: "this is a Yellow Zone task — let me verify the numbers before this goes anywhere." That's the language of someone who has internalised the framework, not just heard it.
And the Red Zone — the most important zone — is where the framework earns its value. Not by making people afraid of AI, but by giving them a clear, defensible reason to say "not here." That's harder to do without a framework. With one, it's just good judgment.
Book a Workshop
Ready to give your team the judgment to use AI well — not just the access?
or
Half-day workshop includes live hallucination demonstrations with real verification exercises, team task mapping across Green · Yellow · Red zones for your actual work, the capabilities and limitations framework applied to your specific domain, and team norms for AI verification and output trust.