Glossary

AI safety / seguridad de la IA: Preventing harm and maintaining appropriate reliability and control.
AI security / protección de la IA frente a ataques: Protecting systems and assets against unauthorized access or malicious interference.
Governance / gobernanza: Institutions, rules, authority, and accountability shaping AI decisions.
General-purpose AI / IA de uso general: Systems usable across many tasks.
AGI / IAG: Hypothetical broadly capable AI; definitions and thresholds vary.
Superintelligence / superinteligencia: Capability substantially above humans across many relevant domains.
Alignment / alineación: Making behavior serve intended goals and constraints.
Agent / agente: A system that can pursue tasks through sequences of actions, often using tools.
Training / entrenamiento: Updating model parameters using data and computation.
Inference / inferencia: Running a trained model to produce outputs.
Compute / cómputo: Computational resources used in training or operation.
Benchmark / prueba de referencia: A defined set of tasks and scoring rules.
Evaluation / evaluación: A structured assessment of capabilities, behavior, or risks.
Hallucination / alucinación: Generated content that is false or unsupported; not necessarily deliberate.
Proxy / indicador sustituto: A measurable stand-in for a goal, which can diverge from it.
Specification gaming / aprovechamiento de fallas en la especificación: Meeting a stated metric or rule while missing the intended goal.
Misuse / uso malicioso: A person's deliberate harmful use of a system.
Threat model / modelo de amenazas: An account of actors, capabilities, access, and pathways to harm.
Causal pathway / vía causal: Linked steps explaining how an outcome could arise.
Loss of control / pérdida de control: A situation where humans cannot effectively direct or regain control of a system.
Defense in depth / defensa en profundidad: Multiple safeguards intended to prevent one failure from causing harm.
Correlated failure / fallo correlacionado: Failures that share a cause or dependency.
Least privilege / mínimo privilegio: Giving only the permissions needed for a bounded task.
Monitoring / monitoreo: Observing signals that may require action.
Fallback / alternativa de respaldo: A usable process when the primary system is unavailable or unsuitable.
Appeal / apelación: A route to challenge a decision and obtain review.
Forecast / predicción: A claim about the future, with assumptions and uncertainty.
Calibration / calibración: How stated confidence relates to actual correctness across many judgments.

Technical words are tools for clarity. Explain the mechanism before naming the term. “Safety” and “security” can both translate as “seguridad”; specify which meaning you intend.