AI workloads and their underlying infrastructure, models, and use of data must be reliable and safe in any scenario into which they are deployed. This principle emphasizes the importance of building AI systems that are dependable and secure, capable of functioning correctly under diverse and unforeseen circumstances. It also involves rigorous testing and continuous monitoring to prevent failures and mitigate risks.

For example, AI is increasingly used in healthcare for diagnostic purposes. To ensure reliability and safety, a hospital might implement an AI-based diagnostic tool and conduct extensive testing in controlled environments before full deployment. Continuous monitoring and updates ensure that the tool performs accurately and safely in real-world medical scenarios.

Software was nearly entirely deterministic prior to the advent of generative AI, which is to say that its programming provided for a specific number of defined outcomes from any input it was provided. When new lead is created, check if contact exists. If contact does not exist, create contact.

Deterministic programs can be tested for every possible outcome, because the outcomes can be quantified and defined.

Modern AI is largely non-deterministic, meaning that the program chooses its own path, its own adventure if you will, each time that it is run. Responses, even to an identical prompt, vary each time the prompt is given and the response is generated.

Let’s illustrate this non-deterministic phenomenon with an innocent example.

In writing this chapter, we provided Microsoft Copilot with the following simple prompt:

Please paint me a picture of a lighthouse.

Copilot returned a response several seconds later, generated by Microsoft Designer using DALL-E 3.

We then repeated the same prompt, only to have Copilot return a different generative image.