LegközelebbKövetkező útmutató
Miért rontja el az AI-videó fizikáját?
Vizuális AI
Nyelvi AI ÚTMUTATÓ
Generative AI can produce polished, decisive language without having verified that its statements are true.
Hallgassa meg ezt az útmutatótAI-narrált · ~3 min
NIST calls this kind of confidently presented false or erroneous output confabulation; readers should treat tone as a writing feature and evaluate evidence, uncertainty and sources separately.
Conversational systems are trained to produce useful-sounding text, and language that reads smoothly can feel authoritative. That feeling is not evidence that the answer was checked against reality. The U.S. National Institute of Standards and Technology uses “confabulation” for generative AI outputs that confidently present erroneous or false content. It notes that these outputs can include fabricated citations or explanations that make a wrong answer seem justified. Why can that happen? A language model generates likely continuations from learned patterns and the current input. It may complete a familiar-looking answer even when a needed fact is missing, the prompt is ambiguous or its training information is outdated. The model can state a guess in the same polished tone it uses for a correct fact. Some products add retrieval, calculators or other tools, but those tools may not be enabled for every turn, and retrieved material can also be misread. Do not infer confidence from phrases like “certainly,” detailed explanations, exact numbers or formal citations. Ask what evidence supports the claim, then inspect it. Open citations, check author and date, confirm that the cited passage says what the chatbot claims, and compare with a reliable source. For calculations, run the calculation independently; for code, execute tests and review security implications; for policy, use the current official document. You can ask a chatbot to distinguish what it knows from what it is inferring, state what information is missing or list sources. These prompts may make uncertainty more visible, but a self-reported confidence score is not a guarantee of calibration. For high-stakes decisions, use accountable human expertise and authoritative records. When evidence is absent or conflicting, preserve the uncertainty instead of turning a fluent answer into a fact.
A chatbot supplies a precise but nonexistent book citation, so the student searches a library catalog before citing it.
A model explains an incorrect calculation in a fluent step-by-step answer, prompting the user to check the arithmetic independently.
A customer-support assistant states an outdated return rule confidently, so an agent opens the current policy page.
A writer asks for uncertainty and sources but still verifies each cited document rather than relying on the response’s tone.
Systems may improve at expressing uncertainty, citing sources and abstaining when evidence is weak, but those behaviors need evaluation in the relevant setting. Interfaces that show which source supports each claim can make review easier, yet users still need to open and assess the source. Better models will not make tone a reliable truth test. Education and product design should reward calibrated uncertainty and make verification straightforward, especially when an answer could affect health, finances, safety or someone’s rights. Keep reassessing performance as systems and uses change.
Ingyenes hírlevél
Minden hétköznap reggel három ellenőrzött AI történet, egyszerű angol nyelven. Örökké ingyen, hirdetések nélkül.
Minden hétköznap egy e-mail. Egy kattintással lemondhat. Soha nem értékesítjük vagy osztjuk meg az Ön címét.
Teszteld magad
Azonnali visszajelzés minden válaszra, és megosztható tanúsítvány egy ellenőrizhető azonosítóval, miután sikeresen teljesít egy kurzust.
Támogassa az ingyenes AI oktatást. AI Understanding egy 501(c)(3) nonprofit – nincs hirdetés, nincs fizetőfal, soha. Tegyen adományt
Generative AI can produce polished, decisive language without having verified that its statements are true. NIST calls this kind of confidently presented false or erroneous output confabulation; readers should treat tone as a writing feature and evaluate evidence, uncertainty and sources separately.
Generative models can produce plausible but false citations and explanations.
NIST defines the risk as confidently presented erroneous or false content.
Current policy should be checked against the source of record.
Language models generate likely continuations and can fill gaps with plausible but false material.
A verbalized confidence score is not automatically calibrated or evidentiary.
Tanulj tovább
További útmutatók készültek ehhez a témához
LegközelebbKövetkező útmutató
Miért rontja el az AI-videó fizikáját?
Vizuális AI