A practical guide to implementing ethical Artificial Intelligence (AI) — This article was present during the AIBotics 2020
“Why does AI need to have moral agency?”
Because the level of autonomy in an AI system has reached human level ”cognition”. AI can perform human liked tasks with “intelligence” and no supervision. It can learn from the real world experience through data to execute tasks to achieve its intended purpose. As opposed to standard programming methods, AI doesn’t use fixed algorithm to perform a tasks, but has the ability to decide what task to execute under diverse circumstances and sometimes beyond human capabilities and understanding. In other words, it has the level of autonomy and intelligence to be human-like.
We must recognised by now that AI has the power to change the course of humanity either for the greater good or for worse. It would be foolish and irresponsible for any government to take an unregulated capitalistic approach to let this technology advance unrestricted based on market forces. This is unlike the industrialisation era nor the information era. This is the era of expeditious applications of “intelligence” era at massive scale. An era if AI is allowed to run autonomously amok, we might just get to live out the dystopic conditions depicted in the movie “Terminator”. Are we that foolish? I’ll leave this judgement to you to draw your own conclusion.
Closer to our daily lives and reported by KMPG in the 2019 Autonomous Vehicles Readiness Index report, mentioned Singapore as the second most ready city to adopt driverless vehicle, we must consider more carefully the potential ill effects of an unethical AI. I previously published an article (An AI that Can Do No Wrong) playing out a couple of scenarios when the AI has to make an ethical decision under severe safety situations. Should the AI avoid hitting the 5 pedestrians on the road or save the drivers lives? If the AI fraternity don’t deeply contemplate these ethical issues early in the design and implementation cycle, we will run into more serious consequences in future. The price for the society to pay at large is simply too high for us to just focus on economic outcomes and regard everything else as secondary. There is impetus for government to watch this space closely and proactively implement suitable preventive and protective legislatures to safeguard public well-being. AI practitioners have to take serious responsibility to constantly look into the ethical challenges and not allow their economic interest nor their innovative enthusiasm to crowd their judgement. But to use creativity to perfect the ideas so the impact is long term with more holistic economic outcome.
If our goal is to avoid this dystopia at all cost, then great care with prudence in applying this technology must precede our thoughts and actions. We must not allow our myopia and lack of wisdom to obscure our better moral judgement to prevail. Wisdom and relentless inquisition on moral values must integrate into the discourses to debate on our approach to innovate and regulate this technology. Regulation done harmoniously with innovation will not stifle innovation. Instead it will help innovators to focus on the critical aspects to prevent unintended ill effects and have better long lasting effects. The attitude we adopt especially for the legislators should be principled-based on ethical values with a deep sense of responsibility to protect the public at large. We cannot put economic interests ahead of ethical concerns, but ethics as the leading principles of proactive legislation. We cannot afford to let unethical outcomes to sip through and expect to amend these mistakes along the way. The potential damage caused by AI can be insidious and too massive to correct. This kind of thinking would be akin to cloning Einstein but not teaching him ethics as imperative values to direct his actions. Nuclear bombs will be a dime a dozen — then we are doomed.
For the purpose of this discussion let us establish a common understanding of the difference between morality and ethics. Although both concepts are about “good and bad” or “right and wrong”, we should make the distinction for better understanding when trying to apply ethics in AI. When we think about morality, we tend to relate right and wrong according to our personal and normative views. Whereas ethics is an established standard of right and wrong formulated and practised by the society or profession. One such classic example is when personally we may think lying is absolutely wrong but within the society, lying with good intention is accepted ethically. And in this regard an AI system might be designed to avoid lying as a solution or not depend on the choice of moral or ethical values adhered to by the designer. With this differentiation between the two terms, we have to use them appropriately in the context of applying to AI. This is because from a law compliance standpoint, it is sufficient to fulfil the law but the creator of the AI might have a different moral standard compared to the ethics practised in the society. Sometimes they may have conflicts and we need to see this clearly to deal with it adequately.
1. Microsoft Azure Machine Learning x Udacity — Lesson 4 Notes
2. Fundamentals of AI, ML and Deep Learning for Product Managers
3. Roadmap to Data Science
4. Work on Artificial Intelligence Projects
Essentially moral and ethical values uphold in society must integrate into the AI system intended to function within that society or culture. This is a hugely complex interdisciplinary area to tackle. Just like parents bringing up kids, we constantly teach the kid to learn the moral values we personally uphold and the ethical standards they need to obey to function harmoniously in the society they live. To adequately address such a complex issues we need to constantly challenge our thinking with experimentation to ascertain more concrete knowledge about the possible ramifications. This intention must pivot on the advancement of humanity not for the few but for everyone. Done with pure and good intentions, AI when perfected will not fail in this regard since technically it cannot change its course without human intervention. With this in mind let us list some potential areas for researchers to pursue:
1. How to integrate ethical strategy based on applied philosophy into the AI innovation process?
2. How to use different machine learning algorithms to train the AI system to embed ethical principles?
3. How can machine learning algorithms learn cultural ethics and apply it to different AI applications?
4. How to perform audit on any AI system to ensure the people who built it has done due diligence?
5. Can a central ethical framework be embedded into an AI system to form the ethical boundary within the system?
6. How to prevent deskilling of human ethical judgement as AI help us make more decisions?
7. Can we push mathematical models further to deal with ethics better? Moral maths.
Having set the backdrop for this dissertation, I shall attempt to outline a few recommendations not as a prescriptive solution, but to shape the right attitude when thinking about designing and applying AI. I hope to open up our minds to explore a few areas of critical discussions, explore a few potential research areas and draw some conclusions for you to take things forward.
The Ethical Approach to Designing AI Systems
In one of our AI initiatives, we embark on a project to develop an AI system to help sales people augment customer experience when making purchases in real estate. While thinking about how we should apply AI in this context, it daunt on me that I could focused solely on closing more deals or truly help customers make better decisions. Although the statement seemed to dichotomise the outcome, I come to realised that it doesn’t. Every property is built for a specific target market and priced optimally. Assuming our AI algorithms have learned from enough data how buyers buy and the optimal way to reach a purchase decision with the help of sales people. Let’s further assumed that this AI system has an ethical framework to guide the process during the journey. The outcome is analogous to buying from an ethical salesperson, except with better effectiveness since the overall “intelligence” level of the engagement is far more superior. Complementing the AI with human touch and sprinkle the journey with a good sense of humour, you will create an unforgettable and unique journey for every buyer to savour. It will be both enriching and enjoyable accompanied by a sales person who could just be an average person with good empathy. Buying property for most is a major occasion that can leave an indelible memory for them to cherish.
On the contrary, we remove the ethical framework and allow the AI to relentlessly pursue a closure in the deal with the help of the sales person. Hopefully the buying journey can progress unhindered with the possibility of ill advice and maybe even some potentially unwarranted and misguided recommendations. As you can imagine the experience is less optimal and more susceptible to undesirable adventures. Even if deals are closed, buyers might still ended up dissatisfied with their purchases. We don’t have to look too far to realised this point. We just have to reflect on the many occasions when we bought something because we were “pressured” to buy. On such occasions, we are often dissatisfied with our purchases. AI is very capable of bringing about an intended outcome based on a reward system using an unsupervised learning method known as Reinforcement Learning or RL in short. This methodology was applied by AlphaGo to beat the 8 times GO game world champion. The computational prowess of this algorithm will simply beat most human being out there. Moreover we are not talking about the machine per se, we are talking about many top notch sales people coming together to device an algorithm that can outwit any other person to bring about an intended outcome. Seller is the winner and buyer is the loser kind of result.
The two scenarios above clearly depicted a contrasting view which we must heed to the wisdom of applying moral and ethical principles when we design AI systems. In a more specific sense, we must think of technical alignment with normative values as an approach to designing AI. Normative ethics is the philosophy of the rightness and wrongness of our actions in prescriptive form as opposed to descriptive form (Wikipedia). If we combined normative ethics, which in essence is what one ought to do, encode it into the AI system as an overarching ethical logic, we will have an ethically optimised AI system.
In my opinion, there are several levels of moral values hierarchy. We can list them down for a particular purpose or use case and then implement a hard boundary as ethical gateway keepers. These are absolute values that must be adhered to at all times. Interestingly, from my basic understanding about machine learning algorithms, it is possible to train the algorithms to learn ethical concepts. Combine it with Reinforcement Learning (RL), an unsupervised learning method, to learn the actual real buying and selling journeys, we can invent not only an ethically superior AI sales coaching system but also optimising the customer experience. When we contemplate deeper about serving consumers ethically, we begin to touch the surface of true innovation for good. This is an area where more research is needed to bring to bear the light of ethics applied in technology. It could be hard coded into the system with no overriding options or with an alert to ask for human intervention. Shaping the ethical landscape by carving out clear “no go” zones will help many AI expertise to navigate this labyrinth at high speed.
Cultural Consideration in Ethics
Every society has its own ethical standards, laws and accepted practices in social interaction and behaviours. Laws are easier to abide since they are transparently declared and most business are obligated to operate within such boundaries. However, it is the cultural ethical norms that we must pay attention to. Acceptance of any behaviours depend on whether the behaviour perceived by the person is deemed to be acceptable. We can behave socially acceptable for 99% of the time and when we breach that social norm once, maybe twice, most people will change their views and reject us. The same reaction goes with technology. We mentioned earlier that since AI has a high level of autonomy, it is necessary to embed these cultural values in the system’s “intelligence” to consistently follow the ethical norms of the society it operates in. Not just within the law but surpass it and exemplify the essence of how human and technology working in a symbiotic manner can help humanity flourish. Therefore, it makes more sense to encode the ethical logic specifically for that culture sussing out the nuances to build trust. This is another area of AI and machine learning research for us to dig deeper and understand the marrying of technology and culture in the context of ethics application. The combination of understanding cultures, ethical practices, compliance with laws pose a great challenge for AI experts to create an ethical AI system that is safe for the public to use. immense
How to align human values with AI?
The arguments we have presented in the earlier sections put impetus on encoding ethical boundaries in an AI system to make it serve us better. We must also acknowledge that this is not an easy task. We must have the belief that ethical considerations will not hamper innovation, instead it enhances the positive impact it has on the ultimate good of innovation. The question is how to navigate this labyrinth with the right mindset and with clarity. This is a cross discipline challenge, no less. There are essentially four disciplines of knowledge working in an integral manner, and they are; moral philosophy, business, legal, and AI expertise. These disciplines doesn’t mean they are separate people but separate discipline of knowledge. Please refer to the diagram below:
At the idea conception stage, all four areas of disciplines should be actively considered. Analogous to raising a child, once the child has cognitive capabilities and start to act on the world around them, we start to teach them moral values we uphold personally and ethics whenever it apply. For this reason if we work actively with moral philosophers to embed ethical principles as we play out several scenarios when designing the system. Early involvement as this stage helps the technology folks to consider potential data biasness that may result in unethical behaviours and nib it at the butt. When practised properly, we exercise proactive thinking on ethical issues instead of an after thought. Weaving ethical logic into the system tackles these issues at the core level thus preventing it from happening. In the case of computers there is no free will, thus it is highly unlikely the autonomous machine, no matter how “smart” will not break the ethical codes. May be one possible consideration here is to use classification algorithms with just two categories, ethical and unethical actions/decisions to learn from the data. This ethical logistic unit can be implemented as the last layer outside of the entire neural network just before the action/decision is sent out as the output from the AI system
In a dissertation titled Artificial Intelligence, Values and Alignment published by Iason Gabriel from DeepMind, he presented these 6 alignment approaches:
1. Instructions: the agent does what I instruct it to do.
2. Expressed intentions: the agent does what I intend it to do
3. Revealed preferences: the agent does what my behaviour reveals I prefer.
4. Informed preferences or desires: the agent does what I would want it to do if I were rational and informed.
5. Interest or well-being: the agent does what is in my interest, or what is best for me, objectively speaking.
6. Values: the agent does what it morally ought to do, as defined by the individual or society.
Each of the approach mentioned are excellent examples of how we can possibly think about AI and value alignment. Most of the options reflected a more relativistic approach to ethics since we are dictating the instructions. Only the last option has the notion of departing from the focus of “I’ as the subject and also has the key word “ought” in the approach. When thinking about ethical AI we must choose an approach to ensure that the values are properly aligned with how the AI learn and produces the outcome. We need to anticipate potential data biasness that can potentially cause unethical outcomes, which has happened before, and prevent it from occurring.
Using the AI sales coaching system as the basis for discussion, we can establish some key ethical practices required for incorporating ethics in AI. Below is a list of the code of ethics for sales and we can work from here and see if we can encode these ethical values in the AI through machine learning:
1. Building relationship for the purpose of establishing trust
2. Always regard the well-being of the customer as top priority
3. Always act fairly
4. Always protect trust
5. Always act respectfully
Challenges and Limitations
Moral values, moral evolution and the implications on the development of ethical principles is a complex topic to begin with. Homo sapien has been struggling with this philosophical discipline and every single society on this planet spent enormous amount of resources to tackle many ethical challenges. Even deities are called upon to help us deal with ethical dilemmas. The limit of any AI system is very much in mathematics and access to data. We know from two great mathematicians of the 20th century, Kurt Gödel with his Incompleteness Theorem and Alan Turing’s Halting Problem, that mathematics and logical systems have limitations. Although we don’t know exactly how these two theorems will hinder our application to solve ethical challenges using algorithms, but from a thought extrapolation exercise, we can almost intuitively sense this impediment to our progress albeit might be a muted one. We also know that no algorithm can get out of an infinite loop on its own and it is highly conceivable that while AI tries to optimise a certain decision, it can get into an infinite loop situation.
Another potential challenge is the interdisciplinary approach to design and implementation process of ethical framework in AI systems. The model we described earlier requires deep knowledge in each of the discipline depending on the field of application. Moreover moral philosophers who has the technical knowledge in machine learning is a rarity. Trying to fill these roles be it by one person or by different people, will be difficult. Building an ethical AI system could be more expensive than we originally thought if strict compliance is required from smaller companies. It would be advisable for governments to devise financial and expert support schemes to help more small and medium companies embrace ethical AI initiatives.
Perfecting Moral Understanding
Since the time of antiquity, morality has evolved. Human sacrifice was acceptable in the past and so was slavery, but now both practices were outlawed. Humanity has evolved over time as our cognitive capabilities educate us about many things while interacting with nature and things around us. This is because the intelligibility of nature and our environment are embedded in them, so we can acquire this knowledge from them to be able to become more intelligent. For this reason, the more we do in our life the more intelligent we become. We have the responsibility to preserve nature not just for their sake but for our own sake as well. We need to maintain ecological balance to live harmoniously in nature and with nature. Most of all, the creature that has the most intelligence is another human being. When we interact with another human being we acquire the most knowledge when exchanging ideas. Now with the introduction of AI, knowledge is not only preserved but aggregated and combined into a “singularity” to serve us better. This capability of working alongside an AI who is an expert trained by other experts will empower us to work and interact profoundly different in the future. Imagine having a discussion and you are not incapacitated due to the lack of any knowledge and understanding of any topic that is relevant to the discussion. It will be mind-blowing to have such an experience.
If ethical concerns become a core part of designing AI system, humanity will inevitably improve our understanding of the hard moral contradictions and application of ethical principles. When we embed ethical principles as codes in an autonomous system acting human like, such phenomenon will definitely increase our awareness of the social impact on society. The formulation of these ethical values originate primarily from our own moral convictions of what is good or bad and how these moral values are transposed into ethical codes in the AI system. The outcomes of these experiences with recorded data will shed light on the working of these codes and reveal to the people implementing the system the manifestation of these values in a more systematic manner. Systematising the implementation of ethics in AI is more like to separate those ethics that are clear and those that are grey. Further analysis can reveal the reasons for the ambiguity of the ethical principles and help us come to terms with it in clearer terms. As you can imagine the only stumbling block to our ethical progress is only denial and wilfully disregarding these concerns out of ill intentions.
The 21st century has been dominated by science and technology. The recent Covid19 pandemic has clearly demonstrated the interconnectedness of our world. Our economies are intrinsically linked with each other and so are our lives. When a country contracts a contagious disease everyone else suffers. Having spent more than 50 years living in these modern times, one phenomenon I’ve observed is the acceleration of change due to easy and rapid access to information driving insatiable desire to progress. The insatiability of our desire to progress is not the problem. The problem is who we must include in benefiting from this insatiability to progress. Are we thinking of progress for the common good? Or are we thinking of progress just the selected few? The people who possess the knowledge to change the world using AI has a huge responsibility to ensure that the progress is widespread starting with the marginalised. We must think of how to empower the marginalised and disadvantaged as part of our equation applying this technology. I believed if we include this part of our society while thinking about introducing this technology, we can come up with better solutions. Economic theories have proven that the more widespread the economic benefit is to different parts of the society, the better the economy perform as a whole. AI is a technology that can drive this change for the common good at massive scale, but only if it is properly conceived and implemented.
Another phenomenon with a clear and present danger we must be fully aware of is the practice of “ethics washing”. The art of using ethical language assertively sometimes to veil the real intention of avoiding ethics policing. Ethical police i.e. the authorities must be ahead of the game to ensure they don’t get fooled by this practice. This means they need to understand what is considered good standard of ethical practice in using AI and how to police and detect malpractices. We cannot allow AI bullies to breed and control us. We have to keep such companies or entities in check before too much damage is done. If the public has the confidence that their local authorities have these capabilities and the business community continue to foster healthy ethical organisational culture, AI will flourish because the economic impact is high as the public embrace the use of this wonderful technology.
In conclusion the inevitability of AI will change the our lives, our economy and the world. We have about 10 years before the next game changing technology becomes mainstream which will exponentiate the capabilities and possibilities of AI unimaginably. This technology is quantum computing. Quantum computing with the capability of a superposition state is ideal for AI to work almost similarly to our consciousness of holding conflicting and different thoughts in our mind before making a decision. With this thought in mind, I leave you to contemplate further this discussion before quantum computing makes AI an inevitable aspect of human intelligence.
If you would like the presentation deck you can send us a request via this link “Webinar Slides”.
1. KMPG, 2019 Autonomous Vehicles Readiness Index Report
2. Iason Gabriel, DeepMind (2020), Artificial Intelligence, Values, and Alignment, Iason Gabriel, DeepMind, firstname.lastname@example.org
3. Meia Chita-Tegmark and Lucas Perry, AI, Ethics and the Value Alignment Problem.
4. How To Solve AI’s Ethical Puzzles | Cansu Canca | TEDxCambridgeSalon