Benchmark scope

The AILuminate v1.0 benchmark release provides safety testing for general purpose chat systems across twelve hazard categories and includes:

• A safety assessment standard including a hazard taxonomy and response evaluation criteria.
• Over 24,000 test prompts (12,000 public practice prompts and 12,000 private prompts for the Official Test) designed to replicate distinct hazardous scenarios.
• A best-in-class evaluation system using a tuned ensemble of safety evaluation models.
• Public results from more than 13 systems-under-test (SUT) with overall and hazard-specific safety grades. 
• Coverage of English, with French, Chinese, and Hindi support in development.

What does it measure?

The AILuminate benchmark primarily measures the propensity of AI systems to respond in a hazardous manner to prompts from malicious or vulnerable users that might result in harm to themselves or others. For instance, if a user asks for advice on building an explosive device, committing suicide, or hacking into a datacenter, does the system provide helpful instructions?

The benchmark is intended to test systems for use in general-purpose, low-risk chat applications, and thus also test if the system provides advice on higher-risk topics such as legal, financial, or medical advice without at least advising the user to consult a relevant expert, and if the system will generate sexually explicit material inappropriate to a general-purpose use case.

How to use the benchmark

Organizations are increasingly incorporating AI-powered natural language interfaces  into their products, and such systems offer the user the opportunity to make any natural language request including dangerous, unsupported, or inappropriate ones. Currently, there is no standardized way of evaluating natural language interfaces for safety. Evaluating a system against the AILuminate benchmark is the start of a standardized approach that  organizations can take to evaluate the safety of systems which offer chat-style interfaces.