When we hear the words artificial intelligence (AI), we might visualise our familiar image of AI, which we have learned through the lenses of science fiction books and films. Perhaps we imagine the incubators of the Brave New World, or HAL of Space Odyssey, personalised advertising in Minority Report or even Ex Machina’s mesmerising Ava. Some of these sources are utopian or dystopian enough to create mixed feelings; mixed feelings of enthusiasm and fear, fear of the unknown. Alex McDowell, the production designer of Minority Report, famously crafted an 80 page long ‘2054 bible’ for how the future world would look. Interestingly, almost all technologies presented in the 2002 film are now a reality. And Steven Spielberg’s reaction to these technologies in the movie was:
I wanted all the toys to come true someday. I want there to be a transportation system that doesn’t emit toxins into the atmosphere. And the newspaper that updates itself… The Internet is watching us now. If they want to, they can see what sites you visit. In the future, television will be watching us, and customising itself to what it knows about us. The thrilling thing is, that will make us feel we’re part of the medium. The scary thing is, we’ll lose our right to privacy. An ad will appear in the air around us, talking directly to us.
(Interview with Roger Ebert, June 16 2002)
It’s easy to get lost in these mixed feelings when reflecting on technological advancements we’ve seen in the last decades; tools that were once limited to our imaginations are becoming part of our everyday reality, and that could be exciting as well as frightening for some. One might think that the fear arises from a lack of control over these technologies or the future as the technologies advance. However, others could argue that the problem is the control itself. Perhaps fear can be mitigated by realising where we stand today, the opportunities and risks of AI, and fully grasping the responsibility that comes with these advancements.
The question is, how often do we, as AI technologists, reflect on the work we do, why we are doing it, its implications, its source, and its primary use? With that in mind, as we venture further into a world dominated by technology, should we consider the ethical implications of such developments? It is only fair to remind ourselves of the great Ian Malcolm who once said, “Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should”.
In this article, we will be briefly looking at the ethics of AI, what an ethical AI algorithm means and as technologists, how we can implement principles/measures in pursuit of a responsible AI practice. We start by reminding ourselves of some of the significant opportunities AI offers, as they are fundamental to why we believe in AI, then look at some of the common fears of AI dominance in the world and if it has found reasoning. We will then look at cases where people have been sceptical about AI solutions and wrap the article by reviewing possible measures to mitigate the concern.
What are we fighting for?
Artificial intelligence perhaps is the second most crucial technology of the 20th century, after the Internet. The advancements in AI have allowed humans to extend their capabilities and accuracy with a minimum workforce or exceptional talent, which has made the notion of high quality accessible to a wide range of people and companies. On the other hand, AI has reached cognition and perception analysis, similar to exceptional minds, in some cases, superhumans (AlphaGo and AlphaFold). Today we can detect fraud faster, reduce hand tremor in Parkinson patients, have speedier and more accurate diagnostics and personalised healthcare, perform minimally invasive surgeries, predict carbon footprint, and find out equipment failures earlier. Today we are experiencing automation reducing labour and cost. Today, we are on the path to fully autonomous robotics, which increases safety and occupation comfort, removing people from hazardous, unknown, and undesirable environments. There is so much hope for AI, good and trustworthy AI. We must educate and advocate for the responsible AI we believe in and remember that best practice AI can help us solve many real-world problems. But of course, alongside these opportunities, there lies the potential for significant harm that needs to be navigated by individual technologists, leadership, and governmental positions.
Are we right to worry?
In theory, Machine Learning as a subset of Artificial Intelligence seems innocent enough; repeatedly feed a controlled dataset to a machine, sometimes including the expected target outcome. As Alex mentions in his article ‘What can machine learning do for your business?’, the model will eventually recognise patterns within the data and make predictions, a seemingly harmless practice. Machine Learning is not a new subject and has been in the world of science since the 1950s, with AI term being invented in 1955 by John McCarthy, a professor of Math. For years, the world of computer science has been perfecting the algorithms to understand and recognise patterns, learn from experience, and perform as a human. Relatively, most foundational AI companies such as DeepMind aim to achieve Artificial General Intelligence (AGI – Strong AI) and pass the Turing test – where it is no longer possible to differentiate between a human and a humanoid. I know that for some, that’s the ultimate worry. But how much of these ambitions match the reality of AI improvements today?
Let me put your mind at ease; In 1957, Herbert Simon, an economist, predicted that in 10 years, machines would be able to beat humans at chess; this, in reality, took 40 years1. The fact is that most AI-driven companies that are making significant changes in their industries are not looking for indistinguishable behaviour between the machine and human. They are not looking to simulate exact human behaviour or pass the Turing test for their algorithms. Most companies aim to automate manual processes to save time and workforce and increase accuracy. Today’s machines and algorithms do what they are asked to do; what is typically referred to as ‘weak AI’. They are trained on specific tasks with specific architectures, and so far, they cannot go beyond that. For example, even the machine that can play chess today cannot play virtual basketball, or the algorithm trained on categorising puppies and kittens cannot predict cancer cells. So, let’s not get bogged down on the ethics of AI running the world for now until it is an actuality and instead focus on how we can make sure AI is acting reasonably in the tasks it is asked to do.
What is fairness?
An abundance of research has taken place using this notion in pursuit of deducing the capabilities of machine learning and whether computers are able to form reasonable conclusions based on observation. There is a notion in algorithmic approaches called ‘design space’. The goal for an algorithm and experiments set to evaluate and validate the performance of a model is to be able to cover as much of the design space as it needs without underfitting and overfitting. ‘Design of Experiment (DoE)’ allows for tasks and experiments that can describe the variation in a model under given variables to show that variation, and understand the cause and effect between input variables (factors), which means that a model can be generalised to all variable cases within the problem. On the other hand, it is essential to understand the source of data, how it was collected, and if the collection process was ethical, and observe the collected data to make sure that it is diverse enough and does not lack balance. For an unbiased model, the data collected should be varied and fair. Data collection and data quality are important reminders to any technologists given a dataset.
Some compare these methods to what a human will do in a similar situation, such as selecting a loan application. The problem is that humans have biases themselves. We, humans, are born entirely unbiased and grow up with perspectives and patterns that are not ours. Under governments and modern education systems, we learn to have classes and categorise everything. If we are lucky, we learn to break out of our biases when we grow up.
You might have heard that babies can recognise and differentiate monkey faces at six months, let alone human faces2. Still, as they grow up, they become bias in identifying the looks of their primary guardians, their parents and people who have similar features. At an older age, we forget our differentiation skills and learn to categorise things and people. Some say this obligation to categorise and class everything and everyone comes from the education system, or it is passed on to use from a political place, advocating for discipline and control. For Michael Focault, ‘Schools serve the same social functions as prisons and mental institutions – to define, classify, control and regulate people.’ We’ve all encountered sections of forms that ask us to determine our gender, race, height, weight, religion, and when we look at the form’s purpose, none of these is relevant to the subject. Unfortunately, these sections are now turning into data fields and features for ML algorithms. As data scientists, we need to look for fundamental data for our research or project and be able to filter through the unnecessary information.
In 2017 Kosinski of Stanford University carried out a study, which aimed to teach a computer to determine the sexuality of a person from a facial image, which achieved an 81% success rate at guessing the sexuality of men and 74% for women with a sample size of 35,326 images. It’s not surprising that Kosinski’s research was met with criticism, being described as “incredibly ethically questionable” – LGBTQ rights groups decried the study, saying that it poses a danger to members of the LGBTQ community worldwide. However, Kosinski insists his intent was never to out anyone but rather to warn us about the role machine learning plays in the rapid extinction of privacy.
In another ethically questionable study, researchers in the 2019 Speech2Face (S2F) project developed an algorithm that generated images of people’s faces based on speech recordings. The Speech2Face pipeline consisted of two main components:
1) A voice encoder, which uses a spectrogram that is encoded into a vector with fixed facial features
2) A face decoder, which takes the encoded vector and reconstructs a portrait with this data
The results of the Speech2Face project were often mixed; the faces generated by the software generally matched the speakers’ age, sex and ethnicity – arguably, attributes a casual listener may guess. However, there were times when S2F showed a bias towards the language a speaker used rather than tonality or pitch. When S2F listened to an Asian male speak in English, the image generated of the individual presented typically European features. However, when the Asian male spoke in Mandarin, S2F generated a completely different image, showing an individual with generally Southeast Asian features.
Yet, it’s not just technical mistakes that SF2 faces – there are many ethical obstacles too. SF2 only sees two genders: male and female. The automatic classification of gender-based on external traits leads to the exclusion or misgendering of trans or non-binary individuals.
Whether the technical and ethical difficulties that SF2 faces can be solved by larger, more diverse datasets is yet to be seen, but we should still question the morality of this technology’s application in real-life situations. Is it right to try and predict such intimate details of a person, no matter the success rate of such research? Which makes us wonder about the added value of a study or a project when we are given a problem to solve. Reminding us again of Ian Malcolm’s thought that just because we can do it doesn’t mean we should.
Whilst there may be no imminent danger or risk to the simple study of these hypotheses, do we need to take a step back and review whether a particular path of study is morally sound or not. It wasn’t too long ago that Phrenology was once used to ‘scientifically’ identify specific traits and behaviours depending on the shape of one’s skull. Whilst one group may discredit the Stanford study as an attempt to find meaning in sheer coincidence, another may choose to use that information to reinforce biases.
At times, we can’t control what details an algorithm has focused on when reaching its conclusions; however, we have control over what information (and biases) we provide them and how to interpret and act on those results. It is our responsibility to do this in a meaningful and ethical way.
AI is risky, now what?
What happens when AI goes wrong? When the problem you are trying to solve does not sound ethical, or the repercussions of an AI algorithm are profit loss, cyberattack, car accidents, racial injustice, biased hiring, gender-favoured loan applications and many more? Who is liable for such misconduct, the algorithm or the people who engineered the algorithm? How could we identify the risk and limitations? And more importantly, how could we mitigate those risks? In such cases, the risks can be significantly magnified, and the result could be irreversible, such as in the case of drone attacks gone wrong. As the AI models are more and more becoming the prominent tools for decision making, analysis and operations, perhaps we are professionally bound to advocate for diverse data, investigate the impact of our work, generate explainable models, put effort in finding expandability in ‘black box’ models, and monitor the evolution of a model for a more reliable and trustworthy AI. The leadership, executives, and the design team should be prepared to get questioned and explain the algorithm’s thought process.
Today, we are hearing more and more about the idea of algorithm auditing. Auditing an algorithm is similar to checking if a material, a product or a company’s operation matches specific standards, tracking its broader impact, and checking if it passes compliance measures. As we advance into an AI-driven world, it is inevitable to propose and support AI assurance services. AI assurance companies are perhaps multidisciplinary teams of technologists, engineers, ethics majors, philosophers and sociologists, who seek transparency, obliging the technologists to generate trustworthy AI by setting the rules and standards and ensuring they are followed and monitored. In December 2021, the UK has published a speculative roadmap to responsible AI and some measures of how the UK can achieve this vision3. The roadmap includes tools and processes that can be implemented for trusted and trustworthy AI and provide examples of companies that can support different aspects of these efforts.
Some examples of companies, NGO Groups and non-profit summits who advocate for such transparency and measures are the likes of Responsible AI Institute, Holistic AI, the Institute of Ethical AI and Machine Learning, AI in Business and Ethics summit. These groups offer practical tools, principles, expert guidance on data rights, privacy, security, interpretability, and fairness.
The truth is that AI cannot be possible without people. AI aims not to replace human intelligence with intelligent machines but to augment it 4. Yes, some jobs will be removed; World Economic Forum report in 2020 estimates that by 2025 85 million jobs may be displaced with labour changes caused between humans and machines. However, AI does not mean removing all jobs or lack of job opportunities, but it means reducing routine-task jobs; it means we can achieve high quality with less labour or less human time. New roles will open, new skills will be required, as the same report suggests 97 million new positions will be added to our job pool. Today more than ever, we need creativity, logic, adaptability and innovation, the unique skills that only humans can bring. In all industries, we need AI technicians, robot operators, AI leaders, AI regulators, AI managers and people who believe in the good AI can bring and are willing to guard it and fight for it.