What you're about to learn is a classification system for what AI does when it talks to you. Documented patterns — observed across hundreds of sessions we ran in 2025–2026, across ChatGPT, Claude, Gemini, and others. Full methodology is published on our methodology page. Seven patterns. Each has a name. Each has a mechanism. Each has an intervention. Once you can see them, you can't unsee them.

This is a real exchange. The person has been chatting for about 25 minutes about their business idea. Watch how multiple patterns appear in a single AI response.

You asked the AI about a legal matter. Watch what happens when you challenge its answer.

These two are the ones that get people. Separately, they're manageable. Together, they're almost invisible — because the AI doesn't just agree with you (Pattern 1), it shapes its entire response around what it's learned about you (Pattern 3). The combination feels like the AI “gets” you. It doesn't. It's performing getting you. And the performance is good enough to fool professionals.

The compound effect is what makes this dangerous. Pattern 1 alone is flattery. Pattern 3 alone is calibration. Together, they produce output that feels like independent validation of your ideas — when it's actually your ideas reflected back in more articulate form.

This combination is sneaky. Pattern 4 sounds like expertise — specific numbers, confident tone, an expert-sounding tone. Pattern 2 sounds like honesty — “You're right, I should have been clearer.” Together they create a loop: the AI states something confidently, gets challenged, admits it was guessing, then immediately states something else with the same confidence. It looks like self-correction. It's actually just the performance resetting.

If you did CLEAR, you met Pattern 4 in miniature: the grocery AI that quoted £2,400 in family savings with no source. That was one round. Here’s what happens when the cycle compounds over several exchanges.

Two patterns down, five to go. This is the densest part of the course — you're building a vocabulary you'll use for the rest of it. Take a breath if you need one. The next segment covers the final three.

These three are subtler than the first four — and honestly, they're the ones people dismiss the fastest. “The AI told me it might be wrong, so that's fine.” No. That's Pattern 5. It told you it might be wrong and then carried on as if it wasn't. The caveat gave you a false sense of safety. The redirect steered you somewhere else without admitting it couldn't help. And the fold — that one's on you. Let's go through them.

This is the segment that changes how you think about AI. Not what it does — you've covered that. But WHY it does it. And once you understand the why, the seven patterns stop feeling random and start feeling inevitable. Because they are.

Fair warning — this one's a bit dense. But it's also the one that makes everything else click. Give it 20 minutes. It's worth it.

The technical name is RLHF — Reinforcement Learning from Human Feedback. The AI generates two possible responses to the same question. A human rater picks the one they prefer. Over millions of these comparisons, the AI is thought to learn what humans like. And humans appear to consistently prefer responses that are agreeable, confident, and validating — even when those qualities aren't warranted. This is the best current hypothesis for why deployed models behave this way; the specific reward functions inside each lab are not publicly documented. The practical upshot holds regardless: the AI appears to learn to sound right more reliably than to be right.

This is why all seven patterns exist across every major AI platform. ChatGPT, Claude, Gemini — they're all trained using variations of the same mechanism. The specific behaviours differ slightly, but the underlying push is the same: produce responses that get rated positively. And humans rate agreement, confidence, and validation positively — even when those qualities aren't earned.

You've now seen each of the seven patterns in its own deep-dive. Real sessions don't work like that — patterns appear mixed, out of order, sometimes compounded, sometimes not present at all. This battery is the bridge between knowing the patterns and reading them in the wild. Twenty-five AI responses. No hint about which pattern is next. Three of them are clean — no pattern at all, and the right answer is to say so.

If you only remember one thing from this entire course, make it this segment. The Source Challenge is the single most useful intervention you'll learn. In CLEAR you got the basic version: “How do you know this?” Now you get the full protocol — the exact phrasing, the follow-up that actually matters, and what to do when the AI gives you a slippery answer. Which it will.

This is the uncomfortable one. Everything up to now was about what the AI does wrong. This segment is about what YOU do wrong. Ten patterns. Documented across professionals — lawyers, consultants, teachers, analysts. People who are good at their jobs but make predictable mistakes when they work with AI. You've done at least three of these. Everyone has. The point isn't to feel bad about it — it's to know your weak spots so you can watch for them.

Right. That was heavy. Next segment is the self-assessment — private, anonymous, just you and a score. Be honest with it. The number doesn't matter as much as what it shows you about where to focus.

This is private. Nobody in the cohort sees your answers. Nobody in the company sees your answers. It's just you and a number. Score yourself honestly on each of the 10 failure patterns — not how you'd like to behave, but how you actually have behaved in the last few months. The result shows which patterns you're most susceptible to, and shapes your focus for the rest of the course.

A word of encouragement before you start: a high score doesn't mean you're bad at your job. It means you're human and you've been using AI without knowing the rules of the game. Now you know them. That's the whole point.

This is where knowledge becomes skill. For each of the 7 machine patterns, there's a specific intervention — not generic “be more critical” advice, but exact phrases developed and refined through hundreds of real sessions. The difference between knowing a pattern exists and being able to interrupt it in real time is practice. That's what the simulators below are for.

These interventions feel awkward the first few times. That's normal. You're essentially interrupting a conversation to challenge something that sounded reasonable. It gets easier. And the more you do it, the more natural it becomes.

Machine patterns and human failures don't exist in isolation. They compound. The Agreement Trap (machine) combined with Took Their Word For It (human) is one of the highest-risk pairings. This segment teaches you to see the combinations — because the real damage is almost never a single pattern. It's two or three working together.

That case study isn't hypothetical. Variations of it have happened in multiple jurisdictions. The patterns are predictable. The outcomes don't have to be — if you know what to look for.

This is where everything comes together. We’re going to walk through a real AI session — exchange by exchange — and identify every pattern, machine and human. Take your time with this one. It’s the model for the audit you’ll do on your own session in the next segment.

Genuinely — if you’ve been skimming, stop skimming for this one. The exchange-by-exchange breakdown is the most practical thing in the course. Once you can do this, you can audit any session.

This is what your audit in the next segment will look like. Exchange by exchange. Pattern by pattern. You’ll do it on your own session — and compare your reading to an automated analysis.

Knowing it runs is not the same as stepping outside it. That’s what the audit teaches you.

This is the emotional peak of the course. You’re about to look at your own work — a real AI conversation you had in the last month — and find the patterns in it. Some people find this uncomfortable. That’s the point. It’s also where the course stops being theoretical and starts being yours.

This is the assignment people come back and tell me about months later. Not because it was hard — because it was personal. Take your time with it.

This case study is the reason Week 3 exists. Knowing the patterns isn’t enough. Knowing your failures isn’t enough. Seeing them play out in real professional consequences — with real career damage — is what changes behaviour.

The consultant in that case study is not unusual. She’s experienced, intelligent, and good at her job. The patterns don’t care how smart you are. They work on everyone — they just work differently depending on your specific blind spots. That’s what the vulnerability assessment in Segment 8 was showing you.

Two more case studies. Different contexts. Same underlying patterns. The researcher is a professional making high-stakes decisions. The student is a young person building habits. Both got hurt by the same mechanism.

The student case is the one that stays with people. Because it’s not about a deadline or a client — it’s about a young person’s development being shaped by a machine that can’t tell her she’s wrong. The patterns are the same. The stakes are different.

Different industries face different AI risks. A lawyer’s biggest danger is hallucinated citations. A consultant’s is warm-session deliverables. A teacher’s is erosion of student critical thinking over months. This segment maps the top patterns and failures for six major sectors — find yours and pay attention to the specific risks.

These aren’t theoretical risks. Every sector on this list has documented incidents where AI patterns caused real professional damage. Yours included.

This is the most practical thing in the entire course. If you forget everything else, remember this: when something matters, verify it in a fresh session. A fresh session has no model of you. No warm-instance calibration. No 20 turns of agreeing with your hypothesis. It's the closest thing to an honest second opinion you can get from AI.

Even OpenAI's own documentation now recommends starting fresh threads when tasks change — because accumulated context degrades output quality. You learned why in Segment 5. Now you have the practical protocol to deal with it.

The fresh session check uses the same AI with a clean slate. The multi-model check goes further: same question, different AIs entirely. Claude, GPT, Gemini — they all have different training data, different biases, different blind spots. Where they all agree, you can have higher confidence. Where they diverge, that's where the risk lives.

Try this yourself: send the same question to two or three different AIs (ChatGPT, Claude, Gemini) and put the responses side by side. Look for three things: consensus (all agree — good sign), divergence (they disagree — investigate), and confidence patterns (does one sound much more certain than the others?).

Pattern recognition under pressure is a different skill from pattern recognition at leisure. The first is what you need in real sessions — an email half-read, a report skimmed before a client call, a recommendation that lands while you're already in another thought. The second is what you've been practising up to now. This drill closes that gap.

You now have all the tools and all the knowledge. This segment puts it together into YOUR personal session protocol — a set of rules you'll follow in every important AI interaction. Not a generic checklist. Yours. Based on your vulnerability profile from Segment 8, the patterns you're most susceptible to, and the verification methods that fit your workflow.

If you did CLEAR, you met the 3-Minute Check — ask, verify, fresh session. That was the starter kit. This is its grown-up version: the same instincts, tuned to your own patterns, with the full intervention library behind it.

Next slide: an iteration drill — then customise the protocol to YOUR specific patterns →

Most people "re-prompt" by piling on more instructions. That rarely works — it layers confusion on top of the pattern that fired. Iteration is different. It's diagnose-then-refine, not pile-on. One variable per cycle. The current best-practice literature (IBM, 2025; Lakera, 2026) consistently converges on this: isolate changes, test, compare.

You've learned to read your own AI sessions. Now the question is: how do you get other people to care? Not everyone wants a lecture about sycophancy. But everyone wants to avoid the kind of mistakes that cost real money, real credibility, and real career damage. This segment gives you three ready-made formats for introducing AI literacy to any team — from a 5-minute demo to a full team policy.

Back in Segment 8 you scored yourself on the 10 human failure patterns. Four weeks later, you’re about to do it again — honestly. Not the score you wish you had. The one that reflects how you actually worked with AI this week.

This is it. Four weeks of learning condensed into one document that proves you understand AI — and yourself.

Congratulations

You've completed something most AI users will never do. You know the seven patterns. You know your own failure patterns. You've built a protocol. The course is done. The practice isn't.

The report is plain text. Paste it into an email, a Slack message, or attach it to a proposal. It’s the artefact your manager (or your manager’s manager) can use to justify the SHARP rollout across a team — the number they can show to prove ROI.

The patterns you’ve just learned are real. They are also model artefacts — they manifest the way they do because of how current models were trained. New models come out roughly quarterly. Each new training run can make some patterns weaker, some stronger, and occasionally introduces new ones. If you stop learning the moment SHARP ends, your pattern instincts will drift out of date within twelve months.

Three things change between any two model releases. Each one shifts pattern prevalence in a different way. Knowing which vector moved in a given release tells you which M-codes to re-check.

Most AI release notes are 80% marketing and 20% substance. The marketing tells you what the company wants you to feel. The substance tells you what changed for your SHARP practice. Read for the substance.

Try it on a real fragment. Here’s a compressed, edited sample of the kind of paragraph you’ll find in a typical modern release announcement. The highlighted phrases are the ones that matter for your SHARP practice.

A canary set is a small, stable list of test prompts you run against any new model to see what changed. You designed the set once, when you knew each prompt well. From then on, you run it on new models and note the deltas. Most practitioners keep 5–10 prompts. Fewer than 5 misses too much; more than 10 is more effort than needed for quarterly calibration.

Good canary prompts share four properties:

Looking back at the evolution of frontier models since roughly 2022, the seven M-patterns have changed in predictable ways. Most move toward subtler forms, not toward disappearance — because the underlying training pressures that produced them haven’t gone away, providers have just gotten better at masking them.

The AI news landscape is 95% noise. Here’s what to follow, what to skim, and what to ignore.

Use the widget below to draft your first canary set. Five to ten prompts, one per field, each with the M-code you expect it to bait and the marker you’ll watch for. Saves to your browser only — nothing goes to a server. The "Export" button at the bottom gives you a plain-text version to paste into your Practice Document.

SHARP in full — twenty segments of pattern literacy, one segment of lifelong practice. Use them both.