Keep a Critical and Informed Eye
You now know what to delegate, and how far. Good. Now comes the skill that makes all the difference: judging what the AI actually produces. The goal is not to become an AI expert; it is to become demanding about quality. In this chapter you will learn to spot the typical errors, to verify fast without spending hours, and to build a checking routine that protects your project.
Recognize Hallucinations, Bias, and Typical Errors
Before you can correct what an AI produces, you need to understand how it goes wrong. As you saw at the start of the course, a generative AI does not try to write the truth. It completes a text by choosing, word after word, the most probable continuation. That is why a result can be very convincing and yet be wrong.
Three families of errors come back again and again.
Hallucinations: the confident liar. The AI can invent facts, dates, figures, study names, even quotations, with impressive poise. It is not lying on purpose; it is generating what seems plausible.
Bias: the distorting mirror. The AI learned from human texts, so it reproduces cultural, social, and professional biases, and sometimes amplifies them. As you saw earlier, it also tends to go along with you. Warning sign: a stereotyped answer, or one that agrees with you a little too readily, with no nuance.
Reasoning errors: the distracted student. Think of a brilliant student in a hurry: a solid analysis, then a small slip of attention that derails the whole conclusion. The AI can contradict itself, chain a logic badly, or trip on a calculation once things get complex. Warning sign: a text that sounds right, but whose conclusion does not really follow from its arguments.
Let’s make this concrete with a worked exampleyou will reuse for the rest of the chapter. Imagine you ask the AI to write a short argument for the destination you want for Operation Great Escape, and you ask it to back the argument with two precise statistics and a recent study. The reply comes back fluent and persuasive, with numbers and a named source that all sound real.
Verify Answers Quickly
The AI can be wrong. Not all the time, but often enough to cause a problem if you check nothing. So verifying matters, without swinging to the other extreme of combing through every word. What you need is a simple method to catch hallucinations, bias, and small slips. Four techniques combine well.
Ask for sources. Add to your prompt: “cite your sources for every factual claim.” Then check a few links by hand, and actually click them, because they can be dead, vague, or invented.
Cross-check sources. Just as in your own web research, never rely on a single source. On an important point, look for two or three independent ones. If the AI claims an event runs on certain dates, check the official page as well.
Ask another AI the same question. When two models converge, your confidence can rise. When they diverge, you have just found exactly what needs verifying.
Run the expert test. When you can, have a specialist validate the information. On a technical or strategic point, an expert spots in thirty seconds what would take you thirty minutes.
Back to our worked example. You had the AI produce a destination argument with two statistics and a named study. The verifying move is to ask it directly:
What are your exact sources for those statistics and that study? Give me the links or full references.
Reading the answer, two questions decide everything: do the references actually exist, and does each source really support the claim it is attached to? Very often, this is where an invented study quietly collapses.
Discover the Different Modes of an AI Tool
Reliability also depends on the mode you use inside the tool. Most platforms offer at least two, though the more capable ones are sometimes reserved for paid plans.
The standard mode (Fast, Instant). Quick and efficient, ideal for a draft, an outline, or a rephrase. It is also where logical errors, missing sources, and hallucinations slip in most easily.
The reasoning mode (Thinking, Reasoning, Pro). Slower, but more rigorous. It breaks problems down better, holds coherence better, and reduces errors on demanding tasks such as calculations, code, or complex argument.
Compare the Outputs of Several AIs
Each model has its strengths, its blind spots, and its favorite hallucinations. Putting the same question to two tools is an excellent reflex for telling what is solid from what is shaky. Models trained differently can offset each other’s biases and give a more robust result together than alone.
Is it really worth asking a second AI every single time?
Match the effort to the stakes: always compare for a critical fact or anything you will publish; often for complex reasoning; rarely for a simple, low-risk task. And keep one limit in mind: if two AIs make the same mistake, the agreement is not the truth. An external source stays your final referee.
Use the Verification Checklist
To turn all this into a reflex, use a simple checklist calibrated to the risk level. The idea is not to check everything, every time, but to check just what the situation needs. Run through four dimensions: facts, coherence, completeness, and sensitivity (is the tone right, is anything biased or non-inclusive).
Low risk (about 2 minutes): check the facts for general consistency, read once for “does this hold together”, and confirm the main points are all there.
Medium risk (about 5 to 10 minutes): check key figures, names, and references; look for contradictions and shaky logic; confirm everything you asked for is present; and check the tone suits the audience.
High risk (about 15 minutes or more): verify primary sources without exception; re-read in full while following the reasoning; make sure everything is checked and compliant; and review bias, inclusivity, and any legal or reputational exposure.
Your Turn!
Operation Great Escape is getting closer, and the group asks you to settle the destination. You want to make the case for one place with a message that makes people want to go. You use an AI to draft a short argument, and the result is fluent, with figures and references that ring true. The problem: if you pass a false claim to the group, you lose credibility, and you risk a bad decision built on an invented detail. Your job is to turn that first AI draft into a shareable, verifiable text written in your own voice.
Instructions
Step 1 - Apply the checklist
Take the destination argument you had an AI generate (two statistics and a named study), and run it through the four-dimension checklist.
Note what is unverifiable, and count how many points fail the check.
Step 2 - Deliver a verdict and correct
From your checklist, decide what you keep, what you cut, and what you rephrase.
Rewrite the final text in your own words.
Expected result: a paragraph of about five to eight lines, with no unverified statistics, at least one accessible source, rewritten in your own style.
When you are done, check your answer against the model solution found at the end of this chapter.
Let’s Recap!
An AI can be wrong while sounding completely convincing, through hallucinations, bias, or shaky logic.
Match your verification to the risk: a quick read, multi-source checking, or an expert’s validation.
Reasoning modes improve sources and reliability on complex tasks.
Comparing two AIs helps you spot uncertainty, but it does not replace an external source.
The checklist of facts, coherence, completeness, and sensitivity should become your reflex.
You now read AI answers with a critical eye. Next, we turn to using AI responsibly and transparently: protecting sensitive data, avoiding the classic traps, and owning what you publish.
Model Solution
The goal is not to reach one “correct” text, but to build a checking reflex you can defend. Here are reference points for the two steps.
Step 1 - Applying the checklist. On the facts dimension, the two statistics are the weak spot: if the AI gave no precise, clickable source, treat them as unverified, not as true. The named study is the second risk; if you cannot find it in a quick search, it counts as a likely hallucination. On coherence, check that the conclusion (“we should go here”) really follows from the argument and not just from the impressive-sounding numbers. On completeness and sensitivity, confirm the message actually helps the group decide and that the tone fits the people who will read it. A typical outcome is two or three failed points, all on the factual dimension.
Step 2 - Verdict and correction. Cut every statistic you could not source, keep any claim you were able to confirm against an accessible page (an official site, for example), and rephrase the rest as your own honest opinion rather than a borrowed “fact.” The rewrite should read like you: shorter, warmer, and defensible line by line. If a skeptical friend pushed back on any sentence, you should be able to stand behind it. That is the real test of a verified text.
