Description
Audits whether an AI feature's objective, training signals and behaviour match what users and affected parties actually value, surfacing proxy gaming and harms.
A value alignment audit examines the gap between what an AI system is optimised to do and what the people it serves actually value. Drawing on the alignment literature synthesised by Brian Christian, it inspects the objective function and proxy metrics (clicks, dwell time, thumbs-up), the training and feedback data, and observed behaviour, looking for specification gaming, reward hacking, distributional shift and disparate effects on groups. UX research contributes by eliciting users' and affected stakeholders' values, collecting reflective judgments (regret, endorsement) that differ from momentary behaviour, and testing failure cases. Rigour comes from explicit mapping of objectives to values, documented red-team scenarios, measures that capture considered preferences, and continued monitoring after launch.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Unmigrated
- Stage
- Evaluation, Post-launch
- Data
- Mixed
- Effort
- Weeks
Steps
- For any AI-optimized product feature:
- Identify the objective function being maximized.
- (2) Survey users: 'Is maximizing [objective] what you actually want from this product?'
- Identify potential divergences: what behaviors does maximizing the objective incentivize that users would disapprove of?
- (4) Measure: compare metric-optimized outcomes to user-welfare outcomes (e.g., maximize engagement vs. maximize reported satisfaction).
- (5) Document divergences as ethical design debts.
Sources
- The Alignment Problem — Brian Christian
Also known as
- AI values audit
- Alignment review for AI UX
- Specification gaming review
Categories
- Analytics & Behaviour
- Computational & Machine Learning