Responsibility can't be delegated
Send a text with a mistake and the mistake is yours, not "the neural net's." Make a decision based on an invented number and the consequences are yours. AI is a tool; whoever operates it is responsible for the result. This lesson is about how to verify quickly, without killing the time savings.
How much verification is needed
The depth of verification depends on two things: the cost of an error and how checkable it is.
- Low cost of error (a draft, ideas, an internal note): a quick skim. A mistake will embarrass you at most.
- Medium (a client email, a post, a report): read it in full + check facts and numbers. A mistake hits your reputation.
- High (a contract, a public promise, medicine, finance, legal questions): AI prepares the material, a human makes the decision, ideally a specialist in the field. A mistake costs money or health.
The rule from the hallucinations lesson works here too: the harder a claim is to check at a glance, the less it's trusted by default.
Quick verification techniques
- The links open. Click every link from the answer. A broken link instantly exposes fabrication.
- Numbers — to the source. Any figure that affects a decision, cross-check against the primary source or recompute (for tables — analysis with code from module 4).
- An independent repeat. The same question in a new chat with no hints. A divergence in answers is a marker of a risk zone.
- Cross-examination. "What source is this claim from? How confident are you? What in the answer should be double-checked?" — the model points out its own weak spots fairly well if you ask directly.
- A critic pass. "You are a picky reviewer. Find the errors and unfounded claims in the text below" — with a new chat, on the model's own answer.
Verification takes minutes if it's built into the process: while the model writes the next part, you open the links from the previous one. The net time saving with verification is still several times over.
Blind spots of trust
Three situations where vigilance switches off imperceptibly:
- Accumulated trust. The model gave ten correct answers — you accept the eleventh without checking. The frequency of hallucinations doesn't drop with your trust.
- An area of ignorance. In your own field you catch errors on the fly; in someone else's you don't see them at all. The less you understand a topic, the stricter the verification.
- Urgency. "No time to check" — that's how an invented quote ends up in a presentation. Checking a critical fact takes 2 minutes; cleaning up the consequences takes hours.
Key takeaways
- You're responsible for the result; "the neural net said so" is no excuse.
- Depth of verification = cost of error × difficulty of checking.
- Techniques: open the links, cross-check numbers, an independent repeat, a confidence interrogation, a critic pass.
- Blind spots: accumulated trust, someone else's field, urgency.