Il prossimoProssima guida
Autoriflessione nei cicli degli agenti
Tecnico
GUIDA alle applicazioni
Human-in-the-loop agents route selected decisions or actions to a person for review, correction, or approval.
Ascolta questa guidaNarrato dall'IA · ~2 min
The right checkpoint depends on impact, uncertainty, and the available human context.
Human oversight is a system-design choice, not a guarantee of accuracy. NIST AI RMF recommends defining roles and responsibilities for human-AI configurations. Article 14 of the EU AI Act specifies oversight requirements for high-risk AI systems; it does not require a human checkpoint for every AI use. Human reviewers need usable information, time, authority, and a way to intervene. Article 14 of the EU AI Act addresses human oversight for high-risk AI systems within its scope. Oversight measures should be proportionate to risk, autonomy, and context, and aim to prevent or minimize risks to health, safety, or fundamental rights. They should help overseers monitor operation, recognize automation bias, interpret output, disregard or override it, and interrupt the system when appropriate. A confidence threshold can route low-confidence cases to a reviewer, while checkpoints may suspend and preserve pending state until review and then resume. Active-learning labels may inform later updates, but improvement depends on data and retraining.
Pause a high-impact action for review before execution.
Escalate an uncertain case with the evidence that triggered escalation.
Record a human correction as labeled feedback when permitted.
Provide a stop or override control for covered workflows.
Oversight requirements depend on system use and applicable law. Test whether the human can understand, challenge, and stop the relevant action rather than counting a nominal approval click.
Newsletter gratuita
Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.
Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.
Mettiti alla prova
Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.
Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione
Human-in-the-loop agents route selected decisions or actions to a person for review, correction, or approval. The right checkpoint depends on impact, uncertainty, and the available human context.
La caratteristica distintiva è un punto di controllo umano in cui una persona rivede, modifica o approva l'azione proposta dall'agente prima che venga eseguita.
A confidence threshold is a configurable decision rule: cases below the chosen threshold can be escalated for review. The threshold requires task-specific calibration and monitoring; it does not ensure reliable decisions.
A checkpoint can suspend execution, save the pending state or requested action, and resume after a person reviews or responds. The exact mechanism depends on the agent framework.
Active learning can route selected cases to people for labels that may be used in later training. This can improve a model if the labels and retraining process are appropriate, but improvement is not automatic.
Article 14 applies to high-risk AI systems within the Act’s scope and requires human oversight designed for the system’s risks, autonomy, and context. It does not require the same approval gate for every human-in-the-loop design.
Continua a imparare
Altre guide selezionate per questo argomento
Il prossimoProssima guida
Autoriflessione nei cicli degli agenti
Tecnico