If you search for papers about 'AI readiness', you'll find dozens of articles and papers that invariably say the same thing, but rarely is there any mention of training people in the ethical implications and setting up the mechanism for compliance.
For example, in AI Readiness, Alex Castrrounis explains readiness regarding Technological, Financial and Cultural Issues. Summarizing some of his points:
Organizational — Need a Chief AI Officer (CAIO) or chief analytics officer (CAO), a vision and strategy for each AI initiative; and key stakeholders must sponsor and support the initiative.
Technological — speaks for itself — tools, methodologies, data quality
Financial — Budget, use cases, storytelling
Cultural — Insights have no value if they can't be acted upon
In an extensive benchmark study by Capgemini Consulting, AI Readiness Benchmark POV (PDF), the term "ethics" appears only once in 30+ pages.
Hopefully, I have demonstrated that data readiness is less about the data and more about readiness. Readiness spans multiple layers of operational maturity, including:
Standardized methods for labeling, validating, cleaning, and organizing (indexing) data across an enterprise;
A data strategy that establishes guidance for effective data management and data usage;
Data governance that spans compliance, risk, and regulation related to data (including privacy, security, and access controls);
Data democratization policies that specify access rights, ‘right to see' authorizations, ethical principles, and acceptable applications for data usage across the organization;
Open data platform that aggregates data and enables automated data ingest, processing, storage, and workflow orchestration;
An organizational assessment of technological infrastructure needs; and
Investment in the infrastructure (eg, cloud, GPUs) to support AI solutions.
All of these articles, and many others like them, make good points. I wouldn't suggest that getting ready for AI should not involve these issues, but there is something glaringly missing. As Paul Virilio famously quipped, "The invention of the ship was also the invention of the shipwreck." Ethics, or trustworthy AI, if you like, require more than a position paper. It's a complicated subject that requires training and monitoring to keep you from foundering on the rocks.
To the extent that AI risks involve people — the social context — we can say that they are ethical risks. In general, ethics refers to standards of morally good or bad, right or wrong conduct. Defining and detecting unethical behavior or processes with AI requires attention to numerous social contexts, including:
Diversity and inclusion
Improper use of health records
Invasion of privacy
Lack of sustainability
Conflicts of interest
If your preparation for AI does not consider these risks, your program will have serious problems. An excellent first step is learning the risks of using prohibited and high-risk data. To understand this, it is helpful to study some well-known failures and what steps could have avoided it. Part and parcel of this are learning to ask the questions to prevent similar failures.
AI ethics depend on social context
We've talked a lot about 'AI Ethics', but a necessary sub-topic is data ethics. It is not a simple issue as proper use of data is subject to current and historical ethical frameworks, regulators, cultural mores and professional organizations. It goes beyond a document you assemble at the beginning. Ethics evolve and can be addressed differently across jurisdictions.
Ethical issues with personal information and ownership of data are the most important to understand. The simplest way to remember this is to think about the social context. Social context refers to the immediate physical and social setting in which people live or in which something happens or develops. It includes the culture that the individual was educated or lived in and the people and institutions they interact with.
Ethics in AI becomes an issue when the social context is involved. Companies can incorporate guidelines into their AI development and implementation processes to limit unintended consequences arising from AI implementation. Formulating and practicing AI's ethical application requires consideration of some simple guidelines that can be useful for formulating a more detailed policy, such as the following:
AI researchers should acknowledge that their work can be used maliciously.
Policymakers need to learn from technical experts about these threats.
The AI world must learn from cybersecurity experts about how to protect its systems.
Ethical frameworks for AI need to be developed and followed.
More people need to be involved in these discussions. It should not just be AI scientists and policymakers, but also businesses and the general public.
Identify data biases and related risks associated with a given dataset, including; data collection/sourcing, data types (proxy, demographic, biographical), data utilizations.
Learn appropriate documentation techniques. Doing this guarantees rreproducibility, interpretability, and peer review.
Modelers have to learn how to deal with various types of bias:
Identify and explain data bias, algorithmic bias, proxy discrimination, disparate treatment, and disparate impact. An early warning system to spot situations in which algorithmic bias or proxy discrimination may arise needs to be devised.
One hazardous approach is to use variables that are proxies for protected classes, or that may lead to unfair discrimination (mask GENDER, but retain variable HOBBY needlepoint)
Understand that even though a feature may seem innocuous during the model build, in practice, it can end up responsible for a model unfairly discriminating.
Common types of statistical bias in predictive analytics (ie, selection, observer, survivorship, cognitive) and how to minimize the risks.
Risks and implications of algorithmic bias and proxy discrimination
Your staff needs to be clear about the difference between measuring variables that are behavior-induced and those that are attributes, and how they behave differently in a model.
Apply relevant regulatory frameworks and professional standards of practice to minimize the risks of algorithmic bias and proxy discrimination.
Teamwork and collaboration reduce the risk of unintended biases and unanticipated outcomes.
Points to consider in AI modeling
AI modeling requires a great deal of skill because there is no procedural code to examine. You have to understand the structure of your data and how the algorithms perform.
Learn how machine algorithms work. Typically, it uses a gradient descent method to optimize the objective function. If it diverges, it can begin to look for other variables, — Latent Value —not in the feature set and return incorrect results without warning.
Learn how to construct model fit and risk measures.
Distinguish causation from association and identify secondary correlation (FICO scores and low income; FICO scores cause low income)
Understand how model fitting with historical data may not predict future outcomes.
Construct model degradation techniques because all models decay. Recognize when to rebuild vs. refresh a model.
Be careful of amateurish development and Do-It-Yourself (DIY) platforms.
A futher consideration is the need for mechanisms that explain how the model has arrived at its results, ie model explainability:
Model explainability is critical, especially where reporting to regulators is involved.
There are limits to what can be done today, but a great deal of work is being researched. Stay on top of it.
Some explainability schemes involve the insertion of monitors within the model. There is tension between model explainability, performance and predictive ability. You will need to understand the tradeoffs.
Process is essential:
Proper project management ensures that interested parties are appropriately given input along the way.
It is not unusual for data scientists and AI developers to have conflicts over data, resources, roles and models. It has to be settled.
Understand why and how to implement a post-deployment monitoring framework as a means to ensure the model is working as intended and blind spots are being guarded against.
AI developers need to learn to explain the model, data, intended uses, cautions and risks to those who do not have technical knowledge.
Of course, to be ready for AI, you need infrastructure, skills, data and lots of it, data management skills, management backdrop…all of these things. But if you think about it, it is pretty similar to getting ready for any data technology. However ethics and a program to insure them is strangely missing from the available advice.
There are so many ways to make a mess of things with AI, and there is no code to inspect to see what went wrong. You have to rely on an understanding of the process. Current DIY tools for machine learning pose the ultimate risk — lacking in knowledge of what the model does and no imbued ethical guidelines.