The initiative seeks to address a problem that has become far more visible with the spread of generative AI: the machines we interact with no longer behave like traditional software. They can make mistakes, interpret the same request differently, learn from our behavior, remember certain interactions and change the way they respond over time.
This raises a question that goes far beyond technology: what should our relationship with an artificial intelligence be like?
Microsoft has been researching this question for more than two decades. The result is a set of 18 principles that seek to establish basic rules for designing that relationship and, above all, to ensure that people can understand, correct and control the AI systems they use.
The so-called Guidelines for Human-AI Interaction condense more than twenty years of research on human-computer interaction and were originally presented in an academic paper published in 2019 at the CHI conference.
Although the guidelines predate the explosion of ChatGPT and today's generative artificial intelligence, many of the issues they raise have now acquired far greater relevance: what an AI should remember, how it should learn from us, how it should explain its mistakes or to what extent we should be able to control its behavior.
G1. Make clear what the system can do
The user must understand what the artificial intelligence system's real capabilities are. It seems an elementary recommendation, but it is essential to prevent a person from attributing capabilities to an AI that it does not actually possess.
G2. Make clear how well it can do it
It is not enough to explain what an artificial intelligence does. It is also necessary to communicate how well it can do it. The user must be able to understand that the system can make mistakes and roughly what level of reliability to expect.
G3. Time services based on context
Artificial intelligence should decide when to act or when to interrupt by taking into account the task the person is performing and the context they are in. Not every recommendation is useful at any given moment.
G4. Show contextually relevant information
The system should offer information related to what the user is doing at that moment and to their environment. Relevance depends not only on the content but also on the context.
G5. Match relevant social norms
An artificial intelligence does not interact in a social vacuum. Its behavior should align with the user's expectations, taking their social and cultural context into account.
G6. Mitigate social biases
The system's language and behavior should avoid reinforcing stereotypes or reproducing unfair biases. This is one of the issues that would later become one of the major debates around the development of artificial intelligence.
G7. Support efficient invocation
Asking artificial intelligence for help should be simple. The user must be able to request or activate its services easily when they need them.
G8. Support efficient dismissal
Being able to ignore a feature is just as important as being able to use it. The system should make it easy to dismiss recommendations, interventions or services the user does not want to use.
G9. Support efficient correction
What happens when the artificial intelligence makes a mistake? Microsoft believes the user should have simple mechanisms to edit, refine or correct the result. A mistake should not become a dead end.
G10. Scope services when in doubt
When the system does not fully understand what a person wants, it should try to clarify their intent or reduce its services in a controlled way. In other words, an artificial intelligence should also know how to manage its own uncertainty.
G11. Make clear why the system did what it did
This rule goes straight to one of the major problems of artificial intelligence: explainability. The user should be able to access an explanation of the system's behavior and understand, at least to some extent, why it produced a given result.
G12. Remember recent interactions
Microsoft recommends that systems keep a short-term memory that allows them to refer efficiently to previous interactions. The relationship with an artificial intelligence should not necessarily start from scratch in every exchange.
G13. Learn from user behavior
The system can use a person's actions to progressively personalize their experience. In this way, the AI learns from the user's preferences and behavior over time.
G14. Update and adapt cautiously
But learning does not mean constantly changing. Microsoft warns that changes to the system's behavior should be made cautiously to avoid disruptive changes that confuse the user. A tool that works in a certain way today should not behave completely differently tomorrow without an understandable reason.
G15. Encourage granular feedback
The user should be able to communicate their preferences during ordinary interaction. This is not just about rating whether something is good or bad, but about providing more specific signals that let the system better understand what the person expects.
G16. Convey the consequences of user actions
If a user's decision is going to change the system's future behavior, that consequence should be communicated immediately. The person needs to know what effects their interaction with the artificial intelligence will have.
G17. Provide global controls
Personalization cannot be left exclusively in the hands of the algorithm.
The user should have general controls that let them decide what the system can monitor and how they want it to behave.
G18. Notify users about changes
The last recommendation is especially relevant in systems that constantly evolve. When an artificial intelligence adds new capabilities or modifies existing ones, it should inform the user.
The 18 rules can also be read as a kind of journey through the relationship between a person and an artificial intelligence. The first two, G1 and G2, correspond to the beginning of that relationship. Before using a system, we need to know what it can do and how far we can trust its results.
Rules G3 to G10 focus mainly on everyday interaction. Artificial intelligence should provide relevant information, respect certain social norms and allow the user to easily use, ignore or correct its services.
When the system fails, rules G9 and G11 take on particular importance: the user must be allowed to correct the mistake and be given explanations of what happened.
Finally, rules G12 to G18 address an increasingly important question: what happens when our relationship with an artificial intelligence extends over time.
That is when memory, learning, personalization and adaptation come into play, but also the need to establish controls and communicate changes.
The spread of generative artificial intelligence has radically transformed the way millions of people interact with software. It is no longer just about pressing buttons, navigating menus or typing words into a search engine. Now we converse with machines.
We explain what we need, add context, correct their answers, ask for alternatives and expect them to understand what we have just said. In some cases we also expect them to remember our preferences and adapt their behavior.
That transformation makes many of the questions Microsoft Research raised years before the popularization of large language models especially timely.
What should a user know about an artificial intelligence's limitations? What happens when it makes a mistake? How can they correct it? Why did it produce a given answer? What should it remember? How does it learn from our behavior? To what extent can we control that learning? And what happens when the system changes?
These are questions of design, but also of trust. A system that creates unrealistic expectations about its capabilities can end up disappointing the user. An artificial intelligence that fails without allowing a simple correction generates frustration. A system that reproduces certain biases can amplify them. And a tool that constantly changes without explaining those changes can end up eroding the trust of those who use it.
The Guidelines for Human-AI Interaction seek to provide a common language for anticipating these problems before the product reaches the user. Their main value lies precisely in translating years of academic research into concrete design decisions.
The examples, reusable patterns, work tools and processes included in the HAX Toolkit make it possible to use these recommendations both to build a first product with artificial intelligence and to review systems that are already running.
The goal is not to prevent artificial intelligence from evolving, but to ensure that this evolution is understandable to the people who will use it. Because the more capable these systems become, the more important it will be to establish clearly what they can do, when they can make mistakes and what ability the user retains to intervene.
The HAX Toolkit can ultimately be understood as more than a design manual. Its 18 principles sketch out basic rules for a relationship that is rapidly becoming everyday: the relationship between human beings and intelligent systems.
Microsoft Research began systematizing these recommendations before generative artificial intelligence became a mass-market product. However, the current spread of these systems makes many of those questions even more relevant.
Artificial intelligence can learn, remember, recommend, anticipate and adapt. But it can also make mistakes. And precisely for that reason, designing ever more intelligent machines is not enough. We also need to design how we want to relate to them.