EU AI Act Self-Assessment

Decision process fully documented and accessible. Users can trace how any output was produced. Training data provenance disclosed.

Core decision process documented. Implementation details proprietary but decision factors disclosed. Training data methodology described.

Key decision factors undisclosed. Users see outputs but cannot understand what drives them. Training data composition unknown.

Decision process entirely inaccessible. Black-box outputs with no reasoning chain. No training data disclosure. Trade secret protection on methodology.

Outputs determined entirely by task inputs. Same query produces same result regardless of user behaviour. No engagement optimisation.

Minor personalisation based on explicit user preferences. Task fidelity maintained. No engagement-metric optimisation.

System optimises toward engagement metrics as an explicit design goal. Outputs diverge from pure task fidelity toward what keeps users interacting.

Output production is engagement-first. System actively learns user engagement patterns and optimises for sustained interaction over task accuracy.

System presents outputs users evaluate with their own judgment. No attention retention mechanisms. Easy to stop using.

System presents alternatives alongside its outputs. Friction built into continued use. Override is easy and psychologically costless.

Active attention-recapture (autoplay, streaks, push notifications on by default). Override is physically possible but psychologically costly.

System is the primary reference point for user decisions. No meaningful alternative. Override is structurally unavailable or prohibitively costly.