A couple of years ago, there were scattered conversations about ethical issues with AI. Not the cognitive robots or singularity, but the more mundane stuff of data science and machine learning.
Stories began to creep out about horribly-biased AI applications in recruiting, law enforcement and criminal justice.
Today, issues of ethics are out in the open, and their activity in addressing it is ramping up. However, there is some confusion, or maybe conflation between AI Ethics and “data for good,” the movement to orient tech workers to provide more to support sustainability, humanity, the environment, and fairness.
AI Ethics and “data for good” seem like related, but separate issues. For example, in “data for good,” one could assume that the motivation would involve ethical principles and therefor ethical principles and frameworks would energize this movement. However, a lot of what is promoted as "for good" is just for show, or worse, cynical.
For example, a manufacturer of GMO seeds and herbicides sought an award “for good” recognition for protecting the world’s food supply (a few billion $’s in jury awards already would tend to dispute that notion.) Alternatively, an esteemed (and costly) medical clinic sought recognition for "using technology to save lives." Isn't that what they are supposed to do, or to invest in their crumbling inner-city neighborhood with the millions of tax breaks they collect every year while opening an exclusive clinic in Dubai.
From an article about the Data For Good Exchange meeting in San Francisco, D4GX San Francisco in February 2018:
Over 2,000 data practitioners, ethicists, advocates, and academics participated at the conference or virtually through the Slack channel to discuss, consolidate, and develop the ethical principles. They developed a 5-point framework around“Fairness, Openness, Reliability, Trust and Social Benefit” (aka FORTS).
This is where I think there is some confusion. "Social Benefit," for example, should be a principle for “data for good,” but not an overarching principle for ethical AI. After all, most AI developed will not be designed for Social Benefit, but it still should be ethical.On the other hand, when attempting to lay down an ethical framework, the goal is to illuminate the ethical risks in all AI development (for good or not) and provide guide rails for execution.
I’ve developed my own set of five principles for ethical AI, explained in a previous article:
- Responsibility: for what you develop and what you use
- Transparency: Logic of an AI must be viewable
- Predictability: perhaps not its workings but surely its effect
- Auditability: like a new drug is monitored for adverse effects
- Incorruptibility: An AI in place must absolutely be incorruptible
The FORTS principles, while reasonable, lack a great deal in creating actual direction for developers, such as the five principles of Responsibility, Transparency, Predictability, Auditability, and Incorruptibility do. In particular, the FORTS definition of Reliability is:
I ensure that every effort is made to glean a complete understanding of what is contained within data, where it came from, and how it was created. I also extend this effort for future users of all data and derivative data.
This is problematic. Organizations have been struggling with this problem for decades, and it gets more difficult every year. The naïve assumption is that if you can gather data, you can understand it. The problem is so complicated that very few organizations achieve a level of completeness, such as the Single Version of the Truth movement in data warehousing, or Master Data Management which rarely gets past customer or product.
The technology gods have favored us with an almost unimaginable capacity to expand our reach for data, but like an Achilles Heel, the gift comes with a weakness. We know yet how to manage this gift. For that reason, this ethical principle is mostly pointless. In fact, it begs the question, if that can’t be done, should you proceed anyway?
However, the group has also developed, “10 Global Data Ethics Principles for data practitioners to apply to their work with data," which you find here. It is hard to argue with any of these, but how would they be implemented?
The greasiest hurdle to making progress on ethical application of technology is that most people are compensated for advancing the organization’s strategy (and ethics). One could argue that most commercial organizations do not have a discernable ethical policy, and even if they do, it’s subordinate to their objectives.
Because of that, individual codes or allegiance to external practices are often left at the front door, for economic reasons. Do you believe that every employee of Purdue Pharmaceutical was devoted to the program of Oxycontin addiction, or did they not see the connection, or did they make a choice to disregard the ethical issues? Nevertheless, they most likely went about their jobs. What’s the answer?
Ethics codes need teeth. Without teeth, they’re corporate mission statements.
Professional organizations frequently have ethical codes of conduct, but they are ineffective unless they can enforce penalties for violations. Lawyers can be fined by their bar association, suspended or even disbarred. Doctors can face similar discipline. Even the American Association of Equine Practitioners can fine, suspend, or revoke licenses for violations, including ethical violations. Other organizations, like actuarial associations, do not have that power, but they so thoroughly stress professionalism that it is a community enforced practice. A dismissed actuary for ethical violations would find it difficult to find another position.
Good intentions are helpful, to a point, but there are so many mundane opportunities for AI to go awry. None of the "principles" of the various groups or even regulatory proposals mention one grave threat: the desire of developers to take on tasks they lack the requisite knowledge or capability to accomplish. Naivete. Amateurish development fostered by the increasing emergence of DIY (Do-it-Yourself) AI, AI workbenches, unregulated data and open source.
What’s the solution?
I'm coming around to Accenture's proposition of a smart machine monitoring and analyzing the central system for ethical risks. I'll be hunting for candidates. Accenture proposes a separate AI system that monitors the actions and effects of other AI systems:
What is needed is another application to work as an observer analyzing the main system to decide if it is ethically sound. It is a challenge to decide what is right and wrong in a fast evolving field, but using the guiding principles of accountability, transparency, fairness, and honesty are a must.
I'd have to agree that Fairness and Honesty should be on that list. Those are good principles, but how or where would an organization find a comprehensive application to monitor AI systems? Besides, the system couldn’t operate without being able to determine if a system is ethically sound, which brings us back t the original problem, what is an ethical system? We’re not there yet.
However, In a recent article in Wired, it described Facebook’s annual developer conference, where scientist Isabel Kloumann described a kind of automatic adviser for the company’s engineers called Fairness Flow.
“It measures how machine-learning software analyzing data performs on different categories - say men and women, or people in different countries - to help expose potential biases. Research has shown that machine-learning models can pick up or even amplify biases against certain groups,” when trained on images or text collected online.
In the same article, Wendell Wallach, a scholar at Yale University's Interdisciplinary Center for Bioethics, said when asked about the early corporate experiments with AI ethics boards and other processes:
None have got it right so far. There aren’t any good examples yet. There’s a lot of high-falutin talk but everything I’ve seen so far is naive in execution.
Wallach’s reasoning says that internal processes, like Microsoft's, Google’s and Facebook’s, “Are hard to trust, particularly when they are opaque to outsiders and don't have an independent channel to a company's board of directors.” His solution is to employ “AI ethics officers,” and the implementation of "ethics review boards,” which in my opinion is not a solution. He does, however, point out that external governance, national and international regulations, agreements, or standards will also be needed.
AI, more than any technology that preceded it, represents a clear and present danger to the world. Biased applications have been around much longer than AI, but the reach of AI has to be understood. There are too many governments in the world, and as the old saying, goes, the chain is only as strong as the weakest link. It only takes one unregulated country to release a massive problem. My money is on AI to control AI.