Before using artificial intelligence (AI), teams must have a responsible AI framework and a toolbox in place, notwithstanding the many advantages of innovation. AI is a neutral technology; it is neither intrinsically ethical nor immoral. Alternatively, AI is a technology that adheres to societal norms and standards. It is essential to analyze what restrictions, constraints, or standards are in place or should be established to support ethical AI.
What is Ethical AI?
Broadly, ethical AI can be defined as data science algorithms that make predictions and trigger actions that are unbiased in nature — i.e., do not discriminate in terms of gender, sexuality, race, language, disability, or any other demographic feature — and also set the foundations for more equitable business decision-making.
PwC identifies the following attributes of ethical AI:
- Interpretability: Should be able to describe their decision-making process in its entirety.
- Reliability: Should function within the boundaries of its design and produce standardized, repeatable predictions and recommendations.
- Security: Should be secured against cyber risks, particularly those posed by third parties and the cloud.
- Accountability: Should have especially identified owners who are responsible for the ethical consequences of the usage of AI models.
- Beneficiality: Should prioritize the common good, focusing on sustainability, collaboration, and transparency.
- Privacy: Should spread awareness on obtained and data in use.
- Human agency: Should facilitate more human supervision and participation.
- Lawfulness: Should adhere to the law and all applicable guidelines.
- Fairness: Should not be prejudiced against individuals or groups.
- Safety: Should not endanger the physical or mental well-being of individuals.
Unfortunately, ethical AI is NOT the industry standard by default, and several companies are facing hurdles in its implementation. In a recent survey, respondents recognized the significance of ethical AI, but delivering on this promise is more difficult than it seems. Nine out of ten (90%) top executives agree that moral standards in the creation and use of new technologies may provide organizations with a competitive edge. Nonetheless, approximately two-thirds (64%) of top executives have observed bias in AI systems used by their organization.
3 Roads toward Building Ethical AI
There are three typical methodologies for mitigating the ethical risks associated with data and AI: the academic method, the corporate method, and the regulatory method. Ethicists, who are often found in philosophy departments, are good at identifying ethical difficulties, their origins, and how to reason around them.
The “on-the-ground” strategy comes next. Typically, eager technologists, data analysts, and product managers are the ones raising important questions inside organizations. They are familiar with asking business-relevant risk-related questions since they are the ones that create the products to achieve specific business objectives.
There are now corporations (not to mention governments) implementing high-level AI ethical norms. Google and Microsoft, for example, proclaimed their values years ago. Given the diversity of corporate values across dozens of sectors, an information and AI ethics policy must be adapted to the organization’s unique commercial and legal requirements. There are several steps that you, as a business leader, can take to achieve this.
Steps for Building Ethical AI that Works
To construct ethical AI from its inception, (rather than retrofit existing AI systems with ethics), keep in mind the following steps:
- Define a common agreement of what AI ethics means
This description must be precise and practical for all key corporate stakeholders. Creating cross-functional teams of experts to advise all activities about the development, production, and implementation of ethical ML and AI is also an excellent idea.
- Catalog AI’s impact on business systems
An essential component of developing an ethical AI framework is documenting the company’s AI usage. The business is rapidly adopting AI, notably in the avatar of recommender systems, bots, customer segmentation modeling, costing engines, and anomaly detection. Regular monitoring of such AI techniques and the processes or applications in which they are embedded is crucial to preventing logistical, reputational, and financial threats to your firm.
- Create a data and AI ethical risk framework that is tailored to your industry.
An effective framework includes, at its foundation, an articulation of the company’s ethical values, a proposed governance model, and a description of how this configuration will be maintained. It is essential to build KPIs and a QA program in order to evaluate the ongoing efficacy of an ethical AI approach.
A comprehensive framework also elucidates the incorporation of ethical risk management into operations. It should include a clear procedure for reporting ethical issues to senior leadership or an ethics committee.
- Optimize ethical AI guidance and tools for product managers
Although your framework provides guidance at a broader level, recommendations at the product level must be precise. Standard machine-learning algorithms recognize patterns that are too complex for humans to comprehend. The issue is that frequently a conflict arises between rendering results explainable on the one hand and accurate on the other.
Product managers must be able to make this trade-off. If the outputs are subject to constraints demanding explanations, like when financial institutions must explain why a loan application was denied, then precision will be essential. Product managers should have the tools to gauge its importance in a particular use case.
- Monitor impacts and engage stakeholders
Creating corporate awareness, ethical committees, knowledgeable product owners, managers, architects, and data analysts are all components of the developmental process, and, ideally, purchase procedure. Due to resource scarcity, time management, and a larger — and obvious — inability to foresee all the ways in which things may go awry, it is critical to monitor the effects of information and AI products on the market.
Ethical AI Example: Sentiment Analysis
An excellent example of integrating fairness and inclusivity is sentiment assessment — in order to prep an ML model to distinguish both positive and negative sentiment in textual data, one must offer adequate training data in terms of social and linguistic contexts.
In a sociolinguistic scenario, what language do you employ? Are you considering the wider cultural import that will subsist in tandem with your emotion tags? Have you taken regional linguistic variance into account? These problems pertain to both the automated speech recognition (ASR) and natural language processing (NLP) components of ethical artificial intelligence.
If your ASR model is only trained in US English, for instance, you may encounter transcription problems while processing other English variants. In this instance, major variations between American and Australian English include the pronouncing of r in particular linguistic situations and vowel pronunciation variances in words, which must be included into the AI system.
Using AI Ethically
Beyond building ethical AI, its use also has to be considered and regulated. When individuals are financially rewarded for unethical actions, ethical standards are undermined. Remember, the unfair application of a system may cause harm, and not its inequity, opacity, or other technical features.
As an example, take Deepfake algorithms, an AI technique that is often used for malicious purposes. The vast majority of deepfakes online are created without the victims’ permission. While it is possible to ensure that the generative adversarial network utilized construct Deepfakes works equally well on people of all skin tones and genders, these fairness improvements/corrections are inconsequential — given that the same algorithms are in use for other, more pernicious intent.
Ethical AI has to be woven into every step of the artificial intelligence pipeline, right from the time of conceptualizing the algorithm to development, and prolonged use and maintenance. As this article explains, there are five steps one needs to follow when developing ethical AI, along with using ethical datasets for AI model training and user education.
