AI Alignment
AI alignment is the work of making an AI system pursue the goals its makers and users intend, and follow their instructions and values, even in situations its builders never tested.
Trust, Risk & Safety
How AI goes wrong, how it is attacked, and how it is tested and kept safe.
a
- AI Alignment - AI alignment is the work of making an AI system pursue the goals its makers and users intend, and follow their instructions... More
- AI Benchmark - An AI benchmark is a standard public test used to compare AI models, such as a fixed set of exam questions or... More
- AI Bias - AI bias is when an AI system produces results that unfairly favour or harm certain groups of people, usually because its training... More
- AI Evals - AI evals are tests that check how well an AI system does a specific job, such as answering customer questions or writing... More
- AI Safety - AI safety is the work of stopping AI systems from causing harm, whether through mistakes, misuse or behaviour their makers did not... More
- AI Security - AI security is the protection of AI systems, and the data they use, from attack, theft and tampering. It covers threats such... More
d
- Data Poisoning - Data poisoning is an attack in which someone slips misleading or altered examples into the data used to train an AI model.... More
- Deepfake - A deepfake is a fake video, image or voice recording made with AI to show a real person saying or doing something... More
e
- Existential Risk - Existential risk from AI is the possibility that advanced AI could cause human extinction or a permanent, global collapse of civilisation, for... More
- Explainable AI - Explainable AI (XAI) is AI designed so people can understand why it reached a decision, such as which factors led to a... More
g
- Guardrails - Guardrails are the rules and filters placed around an AI system to keep its output safe, accurate and on topic. They block... More
j
m
- Model Collapse - Model collapse is the loss of quality when AI models are trained on content made by other AI models. Over several rounds,... More
p
- Prompt Injection - Prompt injection is an attack where someone hides instructions in text an AI tool reads, such as a web page, email or... More
r
- Red Teaming - Red teaming is deliberately attacking an AI system to find its weaknesses before others do. Testers try to make it produce harmful,... More
AI Benchmark
An AI benchmark is a standard public test used to compare AI models, such as a fixed set of exam questions or coding tasks. Each model gets a score, so results can be ranked side by side.
AI Bias
AI bias is when an AI system produces results that unfairly favour or harm certain groups of people, usually because its training data or design reflects past unfairness.
AI Evals
AI evals are tests that check how well an AI system does a specific job, such as answering customer questions or writing code. They score its outputs against agreed standards before and after launch.
AI Safety
AI safety is the work of stopping AI systems from causing harm, whether through mistakes, misuse or behaviour their makers did not intend. It covers today's risks and those from more powerful future systems.
AI Security
AI security is the protection of AI systems, and the data they use, from attack, theft and tampering. It covers threats such as prompt injection, poisoned training data and stolen models.
Data Poisoning
Data poisoning is an attack in which someone slips misleading or altered examples into the data used to train an AI model. The model then learns wrong patterns or hidden triggers, and its results get worse.
Deepfake
A deepfake is a fake video, image or voice recording made with AI to show a real person saying or doing something they never did. Modern tools can produce a convincing fake in minutes.
Existential Risk
Existential risk from AI is the possibility that advanced AI could cause human extinction or a permanent, global collapse of civilisation, for example if very capable systems pursue goals that conflict with human interests.
Explainable AI
Explainable AI (XAI) is AI designed so people can understand why it reached a decision, such as which factors led to a loan being refused. It lets staff, customers and regulators check and challenge results.
Guardrails
Guardrails are the rules and filters placed around an AI system to keep its output safe, accurate and on topic. They block harmful replies, protect private data and stop the AI acting outside agreed limits.
Jailbreak
A jailbreak is a prompt written to trick an AI model into ignoring its own rules. People use it to get banned content, or to make a business chatbot say things its owner never intended.
Model Collapse
Model collapse is the loss of quality when AI models are trained on content made by other AI models. Over several rounds, they forget rare details and produce narrower, more repetitive and less accurate output.
Prompt Injection
Prompt injection is an attack where someone hides instructions in text an AI tool reads, such as a web page, email or document. The AI then follows those hidden instructions as if they came from the user.