Responsible AI provides a framework of principles to guide our work and to ensure that we deploy AI safely, ethically and in compliance with the growing number of laws regulating AI.
However, discussions about responsible AI tend to simmer under the surface of the broader conversation around artificial intelligence until something bad happens, such as when facial recognition misclassifies people based on their race, or a self-driving car causes an accident, or, more recently, generative AI produces photorealistic deepfake images of public figures.
What Is Responsible AI?
Responsible AI is a framework of principles for ethically developing and deploying AI safely, ethically and in compliance with growing AI regulations. It’s composed of five core principles: fairness, transparency, accountability, privacy and safety.
These misfires often drive loud, “hot take” debates about the benefits and risks of artificial intelligence, with references to out-of-control robots and hand-wringing about how to keep AI from misbehaving. Then the controversy predictably subsides as public attention turns elsewhere.
But for those of us working in the technology sector, responsible AI should always be front and center as we consider how to apply artificial intelligence in new products and systems.
What Are the Key Principles of Responsible AI?
Definitions of responsible AI have evolved over time, but the current consensus is that it comprises a set of five principles: Fairness, transparency, accountability, privacy and safety.
5 Principles of Responsible AI
- Fairness
- Transparency
- Accountability
- Privacy
- Safety
Fairness
For artificial intelligence, fairness means ensuring that AI systems are free of bias and don’t make decisions that result in discrimination. Bias can appear at various stages of a machine learning model’s life cycle, starting with the design of the algorithms that drive the AI. An algorithm can include errors or incorrect assumptions that cause decisions to favor a particular group, or the data sets that models are trained on may reflect historical biases, causing decisions that perpetuate those biases.
Data scientists employ a number of different techniques to ensure fair outcomes from AI systems, like exploratory data analysis (EDA) to review of data to identify bias, preprocessing data to remove known biases or inaccuracies, and training models using “synthetic data,” algorithmically manufactured information that stands in for real-world data.
Transparency
Transparency means designing AI systems in a way that “opens up the black box” so that other humans can clearly understand how you developed and deployed the system, and how and why that system arrived at any given decision. To achieve transparency, organizations must put in place policies for clearly documenting the design and decision-making processes of their AI systems.
Data scientists must employ interpretable machine learning techniques to ensure that others can clearly understand the logic behind an AI system’s decisions. Organizations also can incorporate some element of human monitoring and review of the outputs of a system to validate the results.
Accountability
Transparency goes hand-in-hand with accountability, which refers to establishing mechanisms for holding those that develop and use AI systems responsible for the outputs and outcomes of those systems.
Organizations establish accountability by putting in place mechanisms for overseeing and monitoring AI systems to ensure they are producing the results that they were supposed to produce — and not causing unintended harm. This could include adopting ethical guidelines for model creators, deploying monitoring and reporting tools to track the performance of AI systems and auditing systems using techniques like fairness testing and bias detection.
Privacy
Privacy broadly refers to a person’s right to control who has access to their personally identifiable information (PII) and how others use that information. In line with data privacy laws like HIPAA in the US and GDPR in the EU, companies must get individuals’ consent to collect their data, collect only necessary data and use that data only for its intended purposes.
In the world of artificial intelligence, privacy issues can come up both while building and training the models that drive AI and when the models are in production, interacting with consumers. Data scientists have a responsibility to ensure that the models they build and deploy protect personal data in both phases (build/train and production). They employ privacy-preserving methods, such as differential privacy or homomorphic encryption, in designing models, and they can use synthetic data to train the models to avoid privacy issues.
Safety
Safety in responsible AI refers to the obligation of developers and organizations to ensure that AI systems do not result in negative impacts for individuals and society, whether physical, such as damage to property, or non-physical, such as discrimination.
Safety can’t be an afterthought. Organizations leveraging AI need to integrate safety into the entire lifecycle of their machine learning models, starting in the design phase by engaging with diverse stakeholders to understand their perspectives and concerns.
As they build and operationalize their models, data science teams can help to ensure safety by conducting risk assessments, testing under a variety of conditions, providing human oversight, and continuously monitoring and improving their models in production.
The Growing Responsible AI Wave
In 2016, Amazon, Google, IBM, Facebook (now Meta) and Microsoft came together to form the Partnership on AI, with a mission to study and promote the responsible use of artificial intelligence. The Partnership created a set of tenets that laid the foundation for today’s responsible AI principles. The tenets focus on ensuring that AI benefits as many people as possible, protects individuals’ privacy and security, complies with legal requirements and produces “explainable” AI systems rather than “black boxes.”
Various public and private efforts have continued to refine voluntary guidelines for responsible AI in the years since the Partnership’s founding. At the same time, national and local governments have begun to put in place actual requirements around AI. Several US states (e.g., California, Colorado, Illinois, New York) already have laws in place related to AI, with more states set to follow.
At the US Federal level, the American Data Privacy and Protection Act (ADPPA) introduced in Congress last year would require comprehensive “algorithm impact assessments” of all an organization’s machine learning (ML) models in production. Elsewhere, the European Union’s AI Act, which looks on track for adoption by the end of 2023, would require conformity assessments of AI systems, documentation of data quality, activity logging and a governance framework for certain models.
With these current and pending AI regulations along with the heightened concerns about the risks of generative AI applications like ChatGPT, there’s more pressure and incentive for organizations to ensure that they are practicing responsible AI.
Getting Started with Responsible AI
Research by MIT Sloan Management Review and Boston Consulting Group in the spring of 2022 found that 52 percent of companies surveyed had put some level of responsible AI into practice. However, if you’re at one of the roughly half of companies that have not started their responsible AI journey, here are steps that your organization can take to begin adopting responsible AI principles.
1. Get Buy In From Leadership Around Responsible AI
First, start at the top. Executive leadership needs to clearly establish responsible AI as a strategic priority for the organization. AI/ML is a cross-functional, multi-stakeholder process, and C-level executives must be involved in creating an ethical AI culture that extends across the entire business.
2. Create an AI Governance Structure
The next step is to create an AI governance structure that includes responsible AI leadership empowered to assemble a team to implement and monitor the organization’s use of AI. This team should also be in charge of communicating and enforcing responsible AI principles across the organization.
3. Conduct a Responsible AI Assessment
Once you’ve established the structure, the organization can now perform an assessment to identify potential ethical concerns related to their current use of AI and determine their current responsible AI maturity level. The responsible AI executive or team should lead this assessment, but other stakeholders should also participate, including groups like IT, legal and compliance. Currently there is not a “standard assessment” for companies to follow, but companies like Accenture, Microsoft and others have published responsible AI frameworks that serve as a useful starting point for an assessment.
4. Draft a Responsible AI Roadmap
Based on this cross-functional assessment, the organization should be positioned to outline the current state of its responsible AI controls, standards and tools, and then draft a roadmap for enhancing its responsible AI maturity. This plan may involve training current staff on responsible AI techniques and investing in tools and staff to implement and oversee the program.
5. Put Responsible AI into Practice
Then the real work begins for all the teams involved in the machine learning lifecycle, putting responsible AI into practice as they design and deploy AI systems. This involves the techniques described above, but it also should include regular reviews of the organization’s responsible AI framework and governance structure as the legal and regulatory landscape evolves over time.
Responsible AI Benefits
The good news is that organizations investing in responsible AI stand to realize a variety of business benefits. Adhering to responsible AI principles positions an organization to meet current and future regulatory requirements, mitigating legal risks. Eliminating biases in decisions reduces financial and reputational risks.
These kinds of programs also can produce operational benefits. For example, increased communication and collaboration among stakeholders can streamline handoffs between teams, allowing an organization to get more models into production faster. Heightened focus on improving model quality and reducing errors can increase productivity and reduce costs.
With these kinds of benefits at stake, organizations have every incentive to start their responsible AI journey today, rather than waiting for new regulations to come into effect — or waiting for an “AI misfire” that forces their hand.
