Is ethical AI actually artificial?

So many companies around the world rely on artificial intelligence (AI), without really considering the implications of how AI programs can act and affect outcomes (positively and negatively) - this point is too important to avoid or not consider. And people, not technology, have the answers.

What is ethical AI?

In 2017, the Select Committee on Artificial Intelligence was appointed by the House of Lords, to consider the economic, ethical and social implications of advances in artificial intelligence. As part of their 2017-2019 session report, they suggested the following code would be suitable when considering AI:

1. Artificial intelligence should be developed for the common good and benefit of humanity.

2. Artificial intelligence should operate on principles of intelligibility and fairness.

3. Artificial intelligence should not be used to diminish the data rights or privacy of individuals, families or communities.

4. All citizens have the right to be educated to enable them to flourish mentally, emotionally and economically alongside artificial intelligence.

5. The autonomous power to hurt, destroy or deceive human beings should never be vested in artificial intelligence.

Isaac Asimov’s three laws of robotics was created in 1942, and provides striking similarities to the above code. Seventy-six years down the line, narrow artificial intelligence now actually exists and has progressed further than it would’ve been possible to imagine so long ago. Asimov’s laws were:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.

2. A robot must obey orders given to it by human beings except where such orders would conflict with the first law.

3. A robot must protect its own existence as long as such protection does not conflict with the first or second law.

Much of Asimov’s writing in iRobot went on to examine how these rules are not sufficient.

The ethics of technology, specific to robots and other artificially intelligent programs, is divided into two parts: robot ethics and machine ethics.

Robot ethics looks at the morality of the human input into the design, construction, usage and treatment of robots and other AI programs. Machine ethics focuses on the moral behaviour of artificial moral agents, for example, can an artificially intelligent being hold moral responsibility and be punished for not acting in a morally obligatory manner? Although certain AI programs are far more advanced than others, it still needs the human (artificial) touch to ensure the ethical oversight of any AI project.

What impact does ethical AI have on the law?

A fundamental part of the law focuses on justice and ethics. At a recent Law Society event, Professor Richard Susskind recommended that AI practitioners become familiar with the study of ethics, as he did as a law student. Susskind suggests that one cannot look to regulation, to the letter of the law, to tell you what is ethical and what is not; you can only use the law to find out what is “legal” and what is not.

There’s also a question of morality—whether you “should” do something, compared to whether it is permitted.

When designing a legal AI system, outcomes must be described in terms of features— what the system will do in response to an action. In any system, unexpected outputs are treated as bugs—undesirable outcomes. In regular programming, one then reviews the code to find the logical programming error that a human has made, and changes it, so the bug is fixed. In AI, specifically machine learning, the artificial intelligence has created its own logic, so the code cannot be changed.

So we have to either restrict the techniques used to build artificially intelligent programs to those that can be interrogated to understand their reasoning, or restrict the domain we’re working on to one where there is significant human oversight.

In practical terms, that means focusing on the extremes in any given case, not on average or typical inputs when trying to find bugs, because an ethical bug is most likely to appear at the margins.

How can we ensure that AI is ethical?

The test criteria for AI systems dealing with matters should be explicitly defined to include scenarios revolving around the marginalised in society.

Companies dealing with volume insurance claims are a great example. In order to make their processes as efficient as possible, these companies create an AI system that automatically classifies claims as Settle or Contest. The system is trained based on thousands of previous examples. They’re confident that it can predict which category a claim should sit in.

It must be ensured that their newly formed AI has not accidentally learned the wrong lessons, therefore putting an unethical AI agent to deal with real human dilemmas. The issue could be that the AI has found some hidden metadata, or patterns, in their training data, meaning it made the right outcome, but for the wrong reasons.

For example, let’s say that the AI system learned that cases from particular postcodes should always be contested, rather than settled. To avoid this, before the AI agent is created, the firm should check with their ethics panel—a management panel with oversight to approve—the terms of this AI agent, and what constraints there should be. If there isn’t an ethics panel in place, the Law Society recommends that there should be. This would be similar to the COLP (compliance officer for legal practice) role that reviews proposed AI solutions, discusses their dimensions, and approves them if they meet ethical standards.

Following the creation of the AI agent, there should be a round of ethical testing, alongside functional testing, to ensure that the AI has not learned the wrong lessons. Humans need to think about unethical outcomes of the newly created AI, and therefore must test to ensure that ethical boundaries are not being breached. This is a prime law for lawyers who are trained in ethics, not for AI designers and engineers, who are trained in computer science, maths and physics.

The aim of this is to uncover unethical behaviour by the AI, and try and trip it up by giving it deliberately biased data, to see if it comes up with biased outcomes. By retrospectively uncovering the rules that the machine learning algorithm has internalised, you can make sure that they’re going to pay attention to the correct facets, and not to bias or unethical features, such as a claimant’s race or gender.

The Law Society is one of many organisations focusing on the ethical application of AI. Recommendations include having an ethical panel or officer who will review the application of AI and the results of ethical testing in a legal context.

What do you do to make sure that your AI programs are ethical?

Comment on the post to let me know.