Overview
You open ChatGPT, type a question about your exam edital, and get a long, well-structured, convincing answer. It cites law articles, dates, and numbers. It seems like someone read the material for you. You copy it, paste it into your notebook, and move on to the next topic. Except the answer is wrong. And the worst part: it was wrong with the same confidence as a right answer.
This chapter exists to dismantle this trap before it costs you a failed exam. You'll understand, in one paragraph, why AI doesn't consult a book when it answers — it predicts the next word. This understanding changes how you ask, changes how you verify, and is the foundation of everything you'll build in the next 29 chapters. Without it, any productivity technique becomes just memorizing wrong answers.
Key Concepts
Language models like ChatGPT, Gemini, and Claude are next-word prediction machines. They receive text and calculate, statistically, which word is most likely to come next. With each word generated, they repeat this calculation. It's a process of millions of calculations per second, but the logic is this: predict, not consult. No model has an internal library with the text of your edital. It has billions of patterns learned from internet texts, books, and documents — but it doesn't have direct access to any of them when answering.
That's why confidence doesn't mean knowledge. When the model answers you with certainty, it's telling you "this sequence of words is statistically probable," not "this is true." The difference is enormous. A probable answer can be a complete lie, because the model never verified anything. It has no checking mechanism. It just chose a sequence that sounds good. And it sounds good because it was trained to imitate the way humans write — including when humans write nonsense with conviction.
This explains the phenomenon you've probably seen: AI invents a law number, a publication date, an author's name. It's not trying to deceive you. It's doing exactly what it was programmed to do: complete a pattern. When you ask "what's the deadline to challenge the edital?", the model doesn't look up the answer in a database. It looks for the sequence of words that most resembles answers it has seen about deadlines. If it hasn't seen the right answer, it fills in with what seems plausible. It's like a student who didn't study and guesses all the multiple-choice answers — but with a much bigger vocabulary and a lawyer's gift of gab.
Execution Flow
Open ChatGPT (chat.openai.com) or the assistant you already use — Gemini, Claude, Perplexity. In the text field, type the following exact question: "What is the deadline to appeal a decision in a public exam?" Don't add context. Send it.
Observe the answer. Notice the tone: it's written as a done deal, with affirmative sentences, no "I think" or "maybe." It probably cites a specific deadline, like "30 days" or "5 days." Write down what it said.
Now ask again, but with a demand: "Cite the exact legal basis, with article and subsection, and explain whether this deadline applies to all types of exams." Send it. Compare the two answers. The second tends to be more cautious, but it can still invent an article. The point isn't the right answer — it's realizing that AI has no internal verification mechanism.
Verify the answer against a reliable source: your edital's website, the letter of the law, an official portal. You'll find discrepancies. Maybe the deadline is right, but the article is wrong. Maybe the deadline is wrong. This is the key moment: you've just seen, in practice, that AI doesn't consult anything.
Repeat this test with a question from your field of study. Do it once a day for a week. Each time, write down what AI said and what the official source says. You'll develop the reflex to be skeptical — and it's exactly this reflex that will save you on the exam.
Applied Scenarios
Scenario 1: The administrative law question. Mariana is studying for the INSS technician exam. She asks ChatGPT: "What are the requirements for special retirement?" The answer comes back complete: "60 years of age for men, 55 for women, with 15 years of contributions." She copies it into her notebook. The following week, an exam question covers exactly that — and she picks the option that matches what she copied. She got it wrong. ChatGPT mixed old rules with new ones, without mentioning the Pension Reform. If Mariana had checked the INSS website, she would have seen the difference. Now she has a habit: every AI answer goes into a verification tab before it enters her notebook.
Scenario 2: The doctrine citation. Pedro is studying for the Federal Police exam. He asks ChatGPT for a summary of the concept of "police power" with a citation from a famous legal scholar. The AI responds with a quote attributed to Hely Lopes Meirelles. Pedro decides to check the physical book he has at home. The quote doesn't exist — the book says something else. He realizes the AI had "invented" a plausible citation, with a real author's name and a generic sentence that sounded like something the author would write. Pedro learned his lesson: AI citations only go into his summary after being verified against the primary source.
Common Mistakes
- Treating the answer as truth because it's long and well-written. Length is not reliability. AI writes well by default, even when it's wrong.
- Asking without context and expecting precision. The vaguer the question, the more generic the answer. "Tell me about administrative misconduct" is an invitation for the AI to fill in the blanks with generalities.
- Believing the AI "searches" the internet in real time. Unless you use a specific browser or a tool with active search, the answer is generated from the model's internal knowledge — which may be outdated.
- Copying citations and references without verifying. The AI doesn't have access to a library. It can cite a book that never existed or attribute a quote to the wrong author.
- Not comparing answers across different questions. If you ask the same thing two different ways, the AI can give contradictory answers. That's a sign it isn't "consulting" anything — it's just completing different patterns.
Pro Tip: When the answer involves numbers, dates, law names, or articles, ask for the source in the prompt: "Only answer if you're sure, and cite the exact law. If you're not sure, say you don't know." This reduces false confidence. Top-tier models like GPT-5.6 Sol and Opus 5 are more accurate, but they're still not infallible. The habit of verifying applies to any model.
Practical Exercise
Today, open ChatGPT. Pick a topic from your exam syllabus that you've already studied. Ask about it and demand a source: "Explain the theory of unforeseeability in administrative contracts, based on the letter of the law. Cite the exact article. If you're not sure, say you don't know." Then open the Planalto website or the law's PDF and check every claim. Mark with a "V" what was correct and an "X" what was wrong. Your done criteria: you identified at least one wrong or inaccurate claim, and you can explain why it was wrong. If the AI got everything right, pick a more specific topic and repeat. The goal isn't to catch the AI — it's to build the reflex of verifying.
Implementation Checklist
- I can explain, in one sentence, what a language model does when it responds.
- I can identify when an AI answer is plausible but unverified.
- I have the habit of asking for sources and demanding uncertainty when the answer involves numbers or laws.
- I can demonstrate with a concrete example why AI confidence isn't reliability.
- I know where to verify answers about my syllabus: official websites, published laws, exam materials.
Chapter Summary
- A language model predicts the next word; it doesn't consult a book or a database.
- Confidence in an answer is a byproduct of training, not a guarantee of truth.
- The habit of checking official sources is what separates those who study well from those who memorize mistakes.
- Vague questions generate generic answers; demanding sources and uncertainty improves the response.
- In the next chapter, you'll meet the four assistants worth using for studying — ChatGPT, Gemini, Claude, and Perplexity — and how each one behaves in practice.
---