Developing ethics for automated systems
If A.I. were to act like a human, it would create choices and make decisions guided by a moral compass similar to our own. But, how do we imbue A.I. with such a human morality?
For a clue, we can look for inspiration in how we each learn morality for ourselves. To a large extent, we are taught by our teachers, parents, and leaders to reflect the morality of the cultures they identify with. Insomuch as those cultures are the byproduct of the input of millions of people, in a way, we can view our own morality as being sourced from a crowd. And, indeed, the idea of crowdsourcing is now being considered as a viable way to teach AI human morality.
Most AI systems today learn via large datasets. Some of the pioneering work (by MIT researchers) on crowdsourcing morality has come in the form of Moral Machine — a tool which allows large groups of people to suggest how machines should make decisions in moral dilemmas. The output is a massive public data set which captures the moral positions of humans around the globe.
Criticism
There are several critics of establishing morality through crowdsourcing, though, including Cornell Law professor James Grimmelmann, who noted that individuals are often influenced by the crowd opinion, and argued therefore that consensus by a group on a moral decision does not make the decision moral, but popular. This criticism suggests that moral decision making happens privately or anonymously, and that revision of current public systems would be a step toward greater moral understanding.
Another issue in investigating tools to crowdsource morality is that the ethics of any automated system are treated like the ethics of software (meaning the ethical system is built around an object rather than a human), which are made by the people who made the product. This leaves automated systems vulnerable to the injection of immoral views.
There are a number of instances in conversational AI development where these issues could arise. Capital One and Salesforce recently acquired conversational AI companies. Microsoft announced new conversational AI capabilities to make Cortana more intelligent and useful. Amazon is developing a wearable, AI-powered device that will soon be able to understand emotional tones in the human voice.
Ghost workers
Companies that continue to acquire conversational AI capabilities at scale could begin implementing ghost work, or automated work performed by unrecognised workers and passed off as “software magic.” A particular case of this is Google Duplex, which uses human labour to fill in for most of its “automated” restaurant reservation calls. More often than not, reservations made from Google Duplex are performed solely by humans – but this isn’t advertised.
In this case it’s not the work that’s inherently moral or immoral, but the working conditions. This scenario could involve making the supply chain less transparent to consumers, and prevent laborers from organising and drawing attention to their needs.
An additional risk revolves around the morality of scenario choosing, which is a reflection of the data an artificially intelligent system is given. In popular opinion, one example of this is the AI diversity crisis – meaning that the specific set of hiring data provided to the AI system encompasses human bias – particularly in workplaces that have historically been male-dominated, like tech.
The specific set of hiring data provided to the AI system encompasses human bias – particularly in workplaces that have historically been male-dominated, like tech.
Since 2014, automation tycoon Amazon has trained hiring algorithms to spot patterns across the resumes of successful candidates, only to recently discover that their hiring process favoured men, and therefore so did the algorithm. If unchecked, this type of training bias could reinforce systematic oppression; if embraced, this same underlying data could reveal unconscious bias in other scenarios to create a more diverse workforce. To ensure the latter, the data used to train A.I. systems must be vetted for historical bias before use and corrected for future use – the responsibility of which should fall on the developers of the system.
In general use, this bias discovery could present data reports on person-specific practices – like a moral report card – and make recommendations toward a greater moral understanding.
Predictive policing
A similar training bias also manifests in predictive policing algorithms increasingly used by police departments across the country, which use historical data on arrest records and police reports to forecast crime. While this kind of predictive analysis might currently be low risk for companies like The Weather Channel – where inaccurate or biased data is more intertwined with image and credibility than human life – the consequences surrounding use by law enforcement are enormous. This is an instance in which crowdsourcing could be less biased than existing data – that is, if we believe the bias of collective cultural opinion in the present is more acceptable than that of the past.
Assumed observation (bias at first sight) is tightly connected, but less mature than conversational AI and application of scenario-choosing to AI. This is evident in the rise of facial recognition software startups (like Clarifai) specialising in race categorisation. While this might be useful to market researchers looking for a demographic-based sample, this type of AI could also be used to systematically reinforce racial profiling.
It’s easy to imagine GoPro footage of police officers pulling over a person of colour being used to train other officers on visual indicators of suspicious activity. Using the technology this way is why some states, like Michigan, have considered banning the use of it by law enforcement. This limitation of AI – and its reputation as truly limitless – is a weakness in its moral makeup, or lack thereof.
A novel way
As humans, our own ethics are determined by the experiences we were exposed to and the culture of the people who tried to teach us right and wrong. The complexity here is that moral decisions are often private, and so crowdsourcing is best utilised when users interact individually – whether through platforms that allow that privacy or through historical sources that draw from emotional and ancestral histories, not just hard data. Google Brain, for example, swallowed 11,000 novels with the intention of sharpening its conversational capabilities through emotionally rich and diverse sentences and grammatical cadences. The same method could possibly be used to privately crowdsource a variety of moral frameworks, since classic (and classical) literature has a history of spotlighting controversial ethical dilemmas.
Antigone could teach civil disobedience and justice. The Fountainhead could teach complexity in selfishness.
We learned (and do learn) by generalising, rather than case by case scenarios. This is where assumed observation training – which depends on a case by case scenarios – falls short at this cultural moment. A.I. can not yet shortcut the process of learning to apply one lesson abstractly. Until then, trusting A.I. to make morally sound decisions independent of human intervention is risky. Once A.I. reliably understands generalised learning, observation could (potentially) be a reasonable path forward. This could be one solution toward that path.
Moral roadmaps
Perhaps the difficulty in training AI to act, observe, and make decisions like a human is that common practice relies on the assumption that AI is not human. We should instead assume that historical data – although essential – stripped of emotional and ancestral context isn’t sufficient enough to inform a reliable moral system.
In this line of thought, the makeup of our moral roadmaps are crowdsourced cultural stories that develop over our lifetime. Crowdsourcing is the natural mechanism by which we inject beliefs into a large model or system – humanity – and the clear path toward moral AI.
In looking toward the future of this technology, it is possible that human morality might be the highest standard currently imaginable to which we can hold A.I. The best we can do is what we’ve established over time. In measuring the consequences of unbridled AI, then, the question is not, can we trust a robot? But, can we trust ourselves?
