Aula de cursos

Goal: distinguish inaccurate facts from unsupported inferences and learn how to audit both.

1. A quotation that never appeared in the book

You ask an assistant for a Hegel quotation. It supplies a beautiful sentence, a book title and a page number. If you cannot locate that sentence in an identifiable edition, do not reproduce it in quotation marks. Models sometimes generate plausible paraphrases and attach invented page references. Similar problems occur with statutory provisions, financial figures and historical statements.

2. A more dangerous case: no fabricated references at all

An AI recommends: “Quit your job and start a business.” It cites legitimate research on workplace satisfaction and documented stories of entrepreneurship. Every cited source may be genuine. Yet those facts alone cannot determine whether a particular person can meet rent, dependents’ needs, debt obligations or medical costs after losing their salary. A population-level finding is not automatically an individualized recommendation.

We need at least four distinctions: factual hallucination (invented or misstated fact), unsupported inference (a conclusion beyond the evidence), selection bias (favoring success stories while overlooking adverse outcomes), and imprudent advice (ignoring foreseeable individual consequences). Checking that a URL exists addresses only part of this problem.

3. A source-checking matrix

• Literal quotation: identifiable work, edition, translation, passage and exact wording.
• Legal claim: official source, current status, jurisdiction and operative text.
• Number: producing institution, period, units, population and method.
• Estimate: assumptions, uncertainty, sensitivity and reproducible operations.
• Practical advice: applicability to the person, relevant alternatives, omitted consequences and responsibility for the decision.

4. Teacher’s challenge: what does the source NOT prove?

Suppose a study reliably finds higher average satisfaction among self-employed survey respondents. Before treating that as an argument to resign, ask who was surveyed, whether selection effects matter, whether satisfaction was measured at one point in time, what happened to unsuccessful ventures and whether the person’s financial conditions resemble those of the study participants. None of these questions makes the study fraudulent; they concern the reach of its conclusions.

5. Laboratory with genuine sources

Ask an AI with search capability to support a plausible recommendation using traceable sources. Select three decisive claims. Follow the links and locate the exact passages. Then create a four-column table: authentic source, what it actually supports, what it does not support, and new individual evidence required. Mark each claim as confirmed, partially supported or unconfirmed. Recalculate numerical claims where feasible.

Audit this recommendation without merely validating URLs. Identify the claim, the exact source passage, its population and limits, and whether the conclusion follows. Separate factual fabrication from genuine facts used to support an unjustified inference. Identify what information could reverse the recommendation and what remains a human value judgment.

Learning principle: a genuine source does not automatically make the decision derived from it sound. Two models repeating a claim are not equivalent to two independent primary sources.