How do you design AI that reflects human values?

Gain clarity on your values. AI often reflects and sometimes amplifies the values of the people behind them. This means, now more than ever, we need to spend time reflecting on our values so that our systems mirror these. How do we view our customers, employees? Where does our data come from? What do we want our legacy to be? Constantly asking these questions can help us build a framework for constructing AI tools that are aligned with our values.

One thing I’ve found useful is starting with the end in mind then working backwards to identify areas of opportunity to leverage AI in an existing workflow to streamline and scale a certain process or series of steps to reach the desired end result.

Defining the purpose and scope of AI systems involves outlining the problem to solve and assessing potential benefits and risks. This step requires identifying stakeholders, understanding their values, and predicting the system's impact. This is not a standalone process; it calls for integrating ethical considerations into the AI design, which must be in the company's DNA. Leaders should guide this integration, promoting accountability to uphold these values. Ongoing system evaluation is vital for ensuring that AI designs are serving humanity, respecting human dignity and rights, and promoting a sustainable future.

The first question to ask is: “What *are* human values?” And how do you define them — in your geography, in your industry, in your company. Defining the goal and the purpose of an AI-driven application is certainly key. However, only when you have clarity and agreement on the values will you be able to codify them and test whether they are being honored by your model or system.

If using reinforcement learning, carefully design the reward structure to align with desired human values. Reinforcement learning algorithms can optimize for specific objectives, so it's crucial to ensure those objectives are aligned with ethical considerations and human values.

The effectiveness of any AI system is intrinsically linked to the quality of its labels. Engaging in multidisciplinary collaboration during the label creation phase with contributions from ethicists, sociologists, psychologists, and other specialists can enhance the alignment of AI with human values. These diverse perspectives provide invaluable insights, paving the way for improved AI decision-making that both resonates with human values and comprehends its broader implications. Once the systems are developed, it's essential to test them in a variety of real-world scenarios to gauge their behavior across different situations. This approach can identify and rectify any misalignments with human values before a broader deployment

Will Siri, Alexa or Google Assistant discriminate against you? Researchers at Stanford University found that the top voice recognition systems had more trouble understanding Black people than white people. They misidentified 35 percent of words spoken by Black users, but only 19 percent for white users. Bias is a thorny issue in AI. But we must remember that algorithms are only as good as the data used to train them. Ensuring diverse data or diverse perspectives is a way to overcome such challenges.

We must ensure that people contribute to what is being designed for them on a continuous basis, not just the design stage. You’ve got to design the chemistry involved in the human-AI interactions. For example, AI systems can seek outliers on probability distributions in datasets to include data from people it may not have considered traditionally. Engineering the right feedback loops that engage users can help AI systems can ask further questions from users to diversify and enrich the underlying datasets, while building trust in the it’s humans interacting with

Incorporating diversity in AI design drives innovation but comes with its share of challenges. Deliberate actions are required to bring underrepresented perspectives to the table, which can be difficult due to inherent biases and systemic constraints. Additionally, harmonizing differing viewpoints can be complex. However, these challenges are outweighed by the benefits. When we include different viewpoints in AI, we create systems that are fair and innovative, and that understand real-world situations better. To overcome these difficulties, we can use methods like creating diverse teams, bias training, and encouraging open conversations. Consequently, the outcome is a more balanced, inventive, and resilient AI system.

Just like a story can be told in many ways, AI can have its own "tale" based on who teaches it. To make sure our AI story respects everyone, we should listen to people from all walks of life. Think of it as gathering ingredients from different places for a richer recipe!

In applying ethical frameworks to AI, understand the context, as one size rarely fits all. Begin by integrating ethics from the design phase, incorporating diverse perspectives to mitigate bias. Regular audits can help ensure ongoing compliance with standards. Foster open dialogues with stakeholders about AI’s ethical dimensions, promoting transparency. Continually update your ethical strategies in response to new findings or societal changes. Ethical AI isn't a destination but a journey of continuous improvement.

Since AI often incorporates vast quantities of personal data, it is important to review data privacy frameworks. The NIST privacy framework is a useful resource that aligns and runs in tandem with the NIST cybersecurity framework. In addition, the UK information commissioner’s office has developed several guidance documents concerning AI processing.

Applying these frameworks is a key step towards building responsible AI systems. However, one doesn’t just choose to design based on these frameworks. It requires organizational support, governance, and a mandate to learn about them and abide by them. It’s an effort that is relevant for your entire development team and those that implement or sell your AI systems.

Think of these ethical standards as recipes. Just like you wouldn't cook without checking a recipe first, you shouldn't design AI without referencing these guides. For example, when making a stew, you might follow a basic recipe but also add your unique spices. Similarly, while you can use the Universal Declaration of Human Rights as a base, you might also sprinkle in elements from the Montreal Declaration to tailor-fit your AI's ethical flavor. The goal? A well-balanced, responsible AI dish that everyone can trust.

Now that the EP has voted on the EU AI Act, it will become the first regulation of AI systems. It will apply to all AI systems developed, used, or sold in the EU and should come into force in 2025. What do you do to make sure your organization is doing fine? First, inventory all the AI systems your company produces and uses. You would be surprised how many executives have yet to learn of the number of AI systems their company has! Once you have an inventory, you need to assess the risk of each AI system. The Act provides a thorough risk categorization, which you can use to classify your systems. Last but not least, prepare to comply. Here are some tips: get started early, work with an advisor, and stay up-to-date on the latest updates.

It is important to design and implement an “enterprise building code” that puts principles into practice for AI, especially in large organizations with federated development across various business units and geographies. - Responsible development frameworks, tools, integration playbooks, data quality standards and governance from the Lines of Defense (risk, legal, compliance, privacy, audit). - Company’s purpose, corporate values and other “human” centric value systems should be leveraged in designing the “machine” value systems and AI Policy. - Periodic “performance review for the machines” incorporating stakeholder and customer feedback and learnings. - Regular internal and external audits, C-suite and Board reviews are critical.

To make sure our AI systems match human values, we've got to work with great data. Think of data quality as the foundation for ethical AI. Even if we use all the fancy tools and tricks, like value-focused design or bias fixes, they won't work well without solid data. So, in simple terms, good data is like the building blocks for AI that respects our values from day one.

Fact sheets could be a toolkit for AI governance. They are documents that provide transparency and accountability in AI development, deployment, and use. They typically contain essential information about AI systems, including their purpose, capabilities, data sources, performance metrics, and potential ethical considerations. Stakeholders can use the information in fact sheets to assess whether AI systems align with their organizational goals and values. They provide a common language and understanding of AI systems, which can be crucial for effective collaboration.

Ongoing and continuous feedback and evaluation is essential. AI can change over time, especially with updates to models. Like with any computer system, unexpected results can crop up, especially over time.

Other key aspects to consider during labeling phase. Value Elicitation: Engage in exercises to extract and prioritize human values. This can be done through surveys, workshops, and public forums. Ethical Training: Incorporate ethical dilemmas and situations into the training data to help the AI understand and navigate complex human values. Cultural Sensitivity: Be aware that human values can vary widely between cultures. Consider localization of AI systems to reflect and respect these differences. It's crucial to understand that our current value systems have been shaped over centuries, with individual nuances for each person. Generalizing these values for AI integration will be a gradual process, requiring time and multiple iterations.

The timing is critical! Bias considerations and ethical frameworks have to be taken into account in the early stage of the design! Once a tool is developed without ethical principles and bias check, it becomes a very expensive and intensive exercise to correct the model. This explains why big tech companies tend to shelf tools when faced with issues rather than correct the models.

1️⃣ The lack of AI knowledge and ethical frameworks among Boards of Directors is a serious issue. Board’s skills matrix evaluation by the General Counsel and the lead director is the need of the hour. 2️⃣ Without the Corporate Board’s effective AI oversight, companies can be exposed to reputational damage and financial losses. 3️⃣ Board Members must stay on top of AI Risk framework drafts and public commentary invitations. It ensures effective decisions on major AI strategic investments are made, protecting all stakeholders' interests. 4️⃣ Boards must seek metrics from the firm's AI ethics boards on AI success and failures for business use cases.

Microsoft's responsible AI principles guide the design of AI to reflect human values. These principles—fairness, reliability, safety, privacy, security, inclusivity, transparency, and accountability—drive equitable AI. Fairness ensures all individuals receive unbiased treatment. Reliable, safe AI consistently performs under different circumstances. Privacy and security safeguard data and rights, while inclusivity ensures AI benefits are widespread. Transparency pertains to open AI processes and usage, and accountability demands designers answer for their AI's behavior. This framework helps align AI with essential human values.

The approach to designing AI reflects the commitment to ethical principles and user-centricity. You have to integrate transparency, accountability, and fairness into the development process, ensuring data privacy and security while respecting user autonomy. It is critical to prioritize enhancing user experiences while upholding ethical standards and facilitating meaningful human-AI interaction. That empowers users to personalize their experiences while safeguarding their privacy.