Informed consent in digital design is already a complex field (GDPR, cookie banners, unreadable privacy policies). With AI it becomes even more critical: collected data feeds models, inferences create profiles the user did not explicitly provide, and automated decisions can exclude without explanation.
Opt-in vs opt-out is not a detail. Enabling AI features by default — invasive recommendations, behavioural analysis, automatic content generation — and hiding opt-out in settings is a classic dark pattern, made more powerful by algorithmic complexity the user cannot see.
Granular control means the user can choose what to enable: "Use AI to summarise, not to personalise prices", "Analyse my clicks but not my messages". Aggregating everything in one "AI on/off" toggle is convenient for the product team, but reduces real choice.
Human override must be designed as a first-class citizen: visible button "Talk to an operator", "Challenge this decision", "Correct the result". Defined response times. Confirmation that the request was received. No FAQ labyrinth that simulates support.
The right to correction is central: if an AI system misclassifies, misrecommends or misgenerates, the user must be able to report it and see the effect. Feedback that vanishes into a black hole destroys trust and prevents model improvement.
For designers in B2B contexts (admin dashboards, internal tools), the same rules apply to human operators using assistive AI: they must understand what the system does, be able to disable it and not be penalised for ignoring a wrong algorithmic suggestion.