## Defensive UX The core of designing AI systems that interact with humans is all about **defensive UX**: - Set the right **expectations** about the system's capabilities, especially, don't inflate its capabilities to avoid frustration. - Provide **transparency** by explaining the generated results, making it obvious which content is AI-generated, and referencing sources where possible. - AI should be **optional**: Allow for efficient dismissal of the AI if undesired or sub-par. - Collect user **feedback** to improve your models; This should be specific to every inference and in a format that makes it actionable for the model developer. Feedback tracking should include usage analytics. MSFT, GOOG, and AAPL all have published their views on how to achieve that UX, clearly biased by the culture of each company. Beyond these guidelines, you should consider how to [deploy Machine Learning into production](Taking%20machine%20learning%20into%20production.md). ### Microsoft **Microsoft** [has reviewed 20 years of work](https://www.microsoft.com/en-us/research/project/guidelines-for-human-ai-interaction/overview/) covering 168 publications on the interaction of humans with AI. They published their findings as their own [Guidelines for Human-AI Interaction](https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/) paper. In it, they describe the following 18 fundamental guidelines:  Ref.: https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/ ### Google **Google** has a [People + AI Guidebook](https://pair.withgoogle.com/guidebook/). Of particular interest are the [23 patterns](https://pair.withgoogle.com/guidebook/patterns) they recommend. While the Microsoft guidelines focus on the user's mental model, Google's patterns are organized to support data and model development. ### Apple **Apple** has its famous [Human Interface Guidelines](https://developer.apple.com/design/human-interface-guidelines) which contains a section on [Machine Learning](https://developer.apple.com/design/human-interface-guidelines/machine-learning). Apple's recommendations, therefore, are centered on the UI and UX directly. The article contains advice on how to implement [feedback mechanisms](https://developer.apple.com/design/human-interface-guidelines/machine-learning#Explicit-feedback), hedging for [inevitable mistakes](https://developer.apple.com/design/human-interface-guidelines/machine-learning#Mistakes), and [handling uncertainty](https://developer.apple.com/design/human-interface-guidelines/machine-learning#Confidence), for example by providing [multiple options](https://developer.apple.com/design/human-interface-guidelines/machine-learning#Multiple-options).