Understanding the Language of AI Safety

AI safety comes with its own vocabulary. Here are the terms you’ll encounter most often on this site, explained in plain language — followed by a short list of established frameworks doing related work around the world.

Glossary

  • AI Alignment — The effort to make sure an AI system’s goals and behavior actually match human intentions and values, rather than pursuing a literal or narrow interpretation of its instructions that produces harmful side effects.
  • Explainable AI (XAI) — Techniques and design principles that make an AI system’s decisions understandable to humans, instead of the system operating as an unreadable “black box.”
  • Black Box Problem — The difficulty of understanding why a complex AI model (especially deep neural networks) produced a particular output, even for the engineers who built it.
  • Autonomous Weapon Systems (AWS) — Weapons capable of selecting and engaging targets with little or no human control over the final decision to use force.
  • Human-in-the-Loop — A design requirement that a human must review, approve, or be able to override an AI system’s decision before it takes effect, especially for high-stakes actions.
  • Kill Switch — A hardcoded, reliable mechanism to immediately halt or shut down an AI system, which cannot be overridden by the system itself.
  • Model / Data Poisoning — An attack in which corrupted or malicious data is introduced into an AI system’s training process to manipulate its future behavior.
  • Frontier AI / Frontier Model — The most advanced, largest-scale AI models at any given time, whose capabilities are not yet fully understood or tested at deployment.
  • Deepfake / Synthetic Media — Realistic but fabricated audio, video, or images generated by AI, which can be used to impersonate real people or spread disinformation.
  • Guardrails — Built-in rules, filters, and constraints designed to stop an AI system from producing harmful, unsafe, or unauthorized outputs.
  • Jailbreaking (AI) — Techniques used to bypass an AI system’s built-in safety restrictions, tricking it into producing content or actions it was designed to refuse.
  • Agentic AI — AI systems that can plan and carry out multi-step tasks autonomously — for example, browsing the web, executing code, or managing workflows — rather than simply responding to a single prompt.
  • Existential Risk (X-risk) — The category of risk in which an advanced AI system could cause catastrophic, irreversible harm to humanity as a whole, not just to individuals.
  • Algorithmic Bias — Systematic and unfair discrimination produced by an AI system, often reflecting biases present in its training data or design.
  • Artificial General Intelligence (AGI) — A hypothetical future AI system capable of understanding and performing any intellectual task a human can, across domains, rather than excelling narrowly at one task.
  • Red Teaming — The practice of deliberately probing an AI system for weaknesses, safety failures, or exploitable behavior before it is released, in order to fix them in advance.

Related Frameworks & Organizations

Educational reference only — GlobalAIRules is not affiliated with the organizations below.

  • EU Artificial Intelligence Act — The European Union’s binding legal framework regulating AI systems by risk category.
  • OECD AI Principles — Intergovernmental principles for trustworthy, human-centered AI, adopted by dozens of countries.
  • NIST AI Risk Management Framework — A voluntary framework from the U.S. National Institute of Standards and Technology for managing risks in AI systems.
  • Asilomar AI Principles — A widely cited set of AI-safety and ethics principles developed by the Future of Life Institute.
  • UN High-Level Advisory Body on AI — A United Nations body established to recommend international AI governance approaches.
  • IEEE Global Initiative on Ethics of Autonomous Systems — An engineering-standards effort focused on embedding ethics into autonomous system design.