Microsoft, Google, and IBM: The Battle for Ethical AI Guidelines

The Guide to Designing AI Experiences

Three of the largest tech companies in the world, namely Microsoft, Google, and IBM, have recently published guidelines aimed at framing human interaction with artificial intelligence in a responsible manner. These initiatives seek to establish standards for ethical and effective design of AI systems. However, each company has its own interpretation of what responsible AI design means, which is reflected in the diversity of the proposed frameworks.

When ten people working in the industry are asked what responsible AI design entails, they generally provide ten different answers. Terms such as ethical frameworks, trustworthy AI principles, and checklists for responsible innovation are frequently mentioned. However, the fundamental question remains: what makes an AI experience successful when someone interacts with an AI-powered product? And who has succeeded in formalizing this in a useful way?

It turns out that some organizations have indeed taken this initiative. Over the past few years, Microsoft, Google, and IBM have invested in developing public frameworks for human-AI interaction (HAI) design. These frameworks go beyond general ethical advice, offering concrete resources based on models intended for teams creating real products. While their approaches differ, and none cover all aspects, they collectively form a valuable set of resources.

Microsoft: 18 Guidelines and a Library to Support Them

Microsoft has contributed to this effort with its HAX Toolkit, which is based on 18 guidelines for human-AI interaction. This framework was first introduced in a 2019 paper by Saleema Amershi and her colleagues from Microsoft Research. Since then, it has become a go-to reference in the field of AI design.

The HAX Design Library is where these principles come to life. Each guideline is paired with design templates and concrete examples, allowing designers to apply them directly in their projects. The library is also filterable by product category, type of AI, and design goal, making it easier to find relevant resources. Options include transparency, personalization, reliability, fairness, and appropriate reliance, enabling quick access to what is most pertinent for a specific context.

The guidelines cover the entire lifecycle of an AI interaction, from defining initial expectations to managing errors and the ongoing adaptation of the system. A crucial aspect is the notion of appropriate reliance, which aims to calibrate user trust according to the system's actual capabilities. This is particularly important to avoid potential dangers associated with over-reliance on AI systems, such as in the case of a medical triage tool.

Google: Starting with the Right Questions

Google's People + AI Guidebook, produced by their PAIR (People + AI Research) team, takes a different approach. Rather than numbered guidelines, it is structured around six thematic chapters:

• User Needs + Defining Success
• Mental Models
• Feedback + Control
• Explainability + Trust
• Data Collection + Evaluation
• Errors + Graceful Failure

Each chapter is accompanied by worksheets designed to transform the advice into something actionable within a team framework. The second edition, published in 2021, added a set of standalone design templates organized around the questions teams tend to ask in practice: How should I use AI in my product? How do I involve users with new AI features? How do I explain my AI system to users?

This shift from chapters to questions reflects a more honest understanding of how professionals actually work. It makes the resource easier to consult at a specific moment rather than reading it from start to finish.

One template stands out for anyone thinking about chatbots or conversational interfaces: explain to understand, not to be exhaustive. The principle is that when presenting the AI's reasoning to users, you should focus on what they need to move forward, rather than exposing everything that happens behind the scenes. People do not want to be overwhelmed by technical justifications while performing a task; they want enough information to feel informed and in control. This seems obvious when stated like that. It is more challenging to execute correctly.

Traffic to the Guidebook increased by 560% between February and August 2023, as generative AI products flooded the market and teams suddenly needed something more concrete than philosophy. The PAIR team has since begun updating the resource specifically for generative AI, a project that involves even more questions than answers.

IBM: Ethics as Infrastructure, and a Challenge for the Field

IBM approaches this territory from two angles. Their design practice is grounded in a set of ethical principles for AI design, intended to serve as a common foundation for all IBM products. These address areas such as accountability, explainability, value alignment, fairness, and user data rights.

The accompanying AI Essentials Framework is a tool for teams built around five pillars: intention, data, understanding, reasoning, and knowledge.

Their more provocative contribution came at CHI 2024, where a team from IBM Research presented six design principles for generative AI applications. The paper made a specific and well-argued point. Most existing HAI frameworks, including those from Microsoft and Google, were initially designed for AI that makes decisions. Think classification, ranking, prediction. Generative AI works differently. Instead of reaching a conclusion, it produces something: a draft, an image, a block of code. The design challenge shifts from helping users evaluate an outcome to helping them shape one.

The six principles are divided into two groups. Three revisit concerns that will be familiar to most designers but are specifically reformulated for generative contexts.

• The first asks teams to design responsibly, acknowledging that generative outcomes carry new risks related to misinformation and bias.
• The second focuses on designing for mental models, helping users build an accurate understanding of what the system can and cannot do.
• The third returns to the idea of trust and appropriate reliance, which takes on added complexity when a system can produce responses that seem confident but are simply not true.

The other three address characteristics that are genuinely new for these systems:

• Design for Generative Variability recognizes that the same prompt can yield significantly different results each time, which goes against the consistency and predictability that traditional UX design has long prioritized.

• Design for Co-Creation focuses on giving users the controls they need to actively shape and guide the generative process, rather than simply receiving what the model decides to produce.

• Design for Imperfection asks designers to be transparent about results that may be plausible but inaccurate, incomplete, or biased, and to integrate mechanisms that allow people to identify and correct these flaws.

Together, the principles resemble less a checklist and more a shift in how designers are invited to think about the human aspect of these systems. They are still relatively new and not yet as field-tested as the resources from Microsoft or Google, but as a provocation, they are well-argued and timely.

What They Share, and Where the Gaps Are

The three frameworks converge on a set of concerns:

• Transparency about the system's capabilities and limitations
• Support for user control and correction
• Feedback loops allowing both the user and the system to improve
• A version of graceful failure when things do not go as planned

These are, in a sense, the non-negotiables of HAI design, the things that continue to appear regardless of the framework you read.

The alignment itself is significant. They were developed independently, by different organizations, using different methodologies. The fact that they arrive at such similar foundations suggests that these foundations are quite robust. It also means that if your team uses one of these resources as a starting point, you are unlikely to go astray.