What it asks for. High-risk AI systems must be designed so that their operation is transparent enough for deployers to interpret the output and use it appropriately. Providers must supply clear instructions for use covering the provider's identity, the system's intended purpose, capabilities and limitations, and the information deployers need to operate it safely.
In practice. Good instructions for use state the levels of accuracy, robustness and cybersecurity the system was tested against, known and foreseeable risks, performance for specific groups of people, input data specifications, human oversight measures, required computing resources, expected lifetime and maintenance, and how the logging mechanism works. Providers should version these documents alongside model releases and record any predetermined changes. Deployers should hold the instructions on file and be able to show they read and applied them, since Article 26 requires use in accordance with them.
Common gaps. Instructions are often generic marketing documents without measured performance figures, omit limitations for specific populations, or are not updated when the model is retrained. Deployers frequently never receive or review them.