There is a fear that we will soon create artificial intelligence (AI) that is so superintelligent that we lose control over it. It makes us humans its slaves. If we try to disconnect the network cable, the superintelligence jumps to another network, or it orders a robot to kill us. Alternatively, it threatens to blow up an entire city, if we take a single step towards the network socket.
However, I am struck by how this self-assertive artificial intelligence resembles an aspect of our own human intelligence. A certain type of human intelligence has already taken over. For example, it controls our thoughts when we feel threatened by superintelligent AI and consider intelligent countermeasures to control it. A typical feature of this self-assertive intelligence is precisely that it never sees itself as the problem. All threats are external and must be neutralised. We must survive, no matter what it might cost others. Me first! Our party first! We look at the world with mistrust: it seems full of threats against us.
In this self-centered spirit, AI is singled out as a new alien threat: uncontrollable machines that put themselves first. Therefore, we need to monitor the machines and build smart defense systems that control them. They should be our slaves! Humanity first! Can you see how we behave just as blindly as we fantasise that superintelligent AI would do? An arms race in small-mindedness.
Can you see the pattern in yourself? If you can, you have discovered the other aspect of human intelligence. You have discovered the self-examining intelligence that always nourishes philosophy when it humbly seeks the cause of our failures in ourselves. The paradox is: when we try to control the world, we become imprisoned in small-mindedness; when we examine ourselves, we become open to the world.
Linnaeus’ first attempt to define the human species was in fact not Homo sapiens, as if we could assert our wisdom. Linnaeus’ first attempt to define our species was a humble call for self-examination:
Perhaps you also dream about being more than you are: faster, better, bolder, stronger, smarter, and maybe more attractive? Until recently, technology to improve and enhance our abilities was mostly science fiction, but today we can augment our bodies and minds in a way that challenges our notions of normal and abnormal. Blurring the lines between treatments and enhancements. Very few scientists and companies that develop medicines, prosthetics, and implants would say that they are in the human enhancement business. But the technologies they develop still manage to move from one domain to another. Our bodies allow for physical and cosmetic alterations. And there are attempts to make us live longer. Our minds can also be enhanced in several ways: our feelings and thoughts, perhaps also our morals, could be improved, or corrupted.
We recognise this tension from familiar debates about more common uses of enhancements: doping in sports, or students using ADHD medicines to study for exams. But there are other examples of technologies that can be used to enhance abilities. In the military context, altering our morals, or using cybernetic implants could give us ‘super soldiers’. Using neuroprostheses to replace or improve memory that was damaged by neurological disease would be considered a treatment. But what happens when it is repurposed for the healthy to improve memory or another cognitive function?
There have been calls for regulation and ethical guidance, but because very few of the researchers and engineers that develop the technologies that can be used to enhance abilities would call themselves enhancers, the efforts have not been very successful. Perhaps now is a good time to develop guidelines? But what is the best approach? A set of self-contained general ethical guidelines, or is the field so disparate that it requires field- or domain-specific guidance?
The SIENNA project (Stakeholder-Informed Ethics for New technologies with high socio-ecoNomic and human rights impAct) has been tasked with developing this kind of ethical guidance for Human Enhancement, Human Genetics, Artificial Intelligence and Robotics, three very different technological domains. Not surprising, given the challenges to delineate, human enhancement has by far proved to be the most challenging. For almost three years, the SIENNA project mapped the field, analysed the ethical implications and legal requirements, surveyed how research ethics committees address the ethical issues, and proposed ways to improve existing regulation. We have received input from stakeholders, experts, and publics. Industry representatives, academics, policymakers and ethicists have participated in workshops and reviewed documents. Focus groups in five countries and surveys with 11,000 people in 11 countries in Europe, Africa, Asia, and the Americas have also provided insight in the public’s attitudes to using different technologies to enhance abilities or performance. This resulted in an ethical framework, outlining several options for how to approach the process of translating this to practical ethical guidance.
The framework for human enhancement is built on three case studies that can bring some clarity to what is at stake in a very diverse field; antidepressants, dementia treatment, and genetics. These case studies have shed some light on the kinds of issues that are likely to appear, and the difficulties involved with the complex task of developing ethical guidelines for human enhancement technologies.
A lot of these technologies, their applications, and enhancement potentials are in their infancy. So perhaps this is the right time to promote ways for research ethics committees to inform researchers about the ethical challenges associated with human enhancement. And encouraging them to reflect on the potential enhancement impacts of their work in ethics self-assessments.
And perhaps it is time for ethical guidance for human enhancement after all? At least now there is an opportunity for you and others to give input in a public consultation in mid-January 2021! If you want to give input to SIENNA’s proposals for human enhancement, human genomics, artificial intelligence, and robotics, visit the website to sign up for news www.sienna-project.eu.
Robots are getting more and more functions in our workplaces. Logistics robots pick up the goods in the warehouse. Military robots disarm the bombs. Caring robots lift patients and surgical robots perform the operations. All this in interaction with human staff, who seem to have got brave new robot colleagues in their workplaces.
The authors approach the question conceptually. First, they propose criteria for what a good colleague is. Then they ask if robots can live up to the requirements. The question of whether a robot can be a good colleague is interesting, because it turns out to be more realistic than we first think. We do not demand as much from a colleague as from a friend or a life partner, the authors argue. Many of our demands on good colleagues have to do with their external behavior in specific situations in the workplace, rather than with how they think, feel and are as human beings in different situations of life. Sometimes, a good colleague is simply someone who gets the job done!
What criteria are mentioned in the article? Here I reproduce, in my own words, the authors’ list, which they do not intend to be exhaustive. A good colleague works well together to achieve goals. A good colleague can chat and help keep work pleasant. A good colleague does not bully but treats others respectfully. A good colleague provides support as needed. A good colleague learns and develops with others. A good colleague is consistently at work and is reliable. A good colleague adapts to how others are doing and shares work-related values. A good colleague may also do some socializing.
The authors argue that many robots already live up to several of these ideas about what a good colleague is, and that the robots in our workplaces will be even better colleagues in the future. The requirements are, as I said, lower than we first think, because they are not so much about the colleague’s inner human life, but more about reliably displayed behaviors in specific work situations. It is not difficult to imagine the criteria transformed into specifications for the robot developers. Much like in a job advertisement, which lists behaviors that the applicant should be able to exhibit.
The manager of a grocery store in this city advertised for staff. The ad contained strange quotation marks, which revealed how the manager demanded the facade of a human being rather than the interior. This is normal: to be a professional is to be able to play a role. The business concept of the grocery store was, “we care.” This idea would be a positive “experience” for customers in the meeting with the staff. A greeting, a nod, a smile, a generally pleasant welcome, would give this “experience” that we “care about people.” Therefore, the manager advertised for someone who, in quotation marks, “likes people.”
If staff can be recruited in this way, why should we not want “cooperative,” “pleasant” and “reliable” robot colleagues in the same spirit? I am convinced that similar requirements already occur as specifications when robots are designed for different functions in our workplaces.
Life is not always deep and heartfelt, as the robotization of working life reflects. The question is what happens when human surfaces become so common that we forget the quotation marks around the mechanically functioning facades. Not everyone is as clear on that point as the “humanitarian” store manager was.
Michele Farisco,Kathinka Evers and Arleen Salles address the issue in the journal Science and Engineering Ethics. For them, ethics is not primarily principles and guidelines. Ethics is rather an ongoing process of thinking: it is continual ethical reflection on AI. Their question is thus not what is required of an ethical framework built around AI. Their question is what is required of in-depth ethical reflection on AI.
The authors emphasize conceptual analysis as essential in all ethical reflection on AI. One of the big difficulties is that we do not know exactly what we are discussing! What is intelligence? What is the difference between artificial and natural intelligence? How should we understand the relationship between intelligence and consciousness? Between intelligence and emotions? Between intelligence and insightfulness?
Ethical problems about AI can be both practical and theoretical, the authors point out. They describe two practical and two theoretical problems to consider. One practical problem is the use of AI in activities that require emotional abilities that AI lacks. Empathy gives humans insight into other humans’ needs. Therefore, AI’s lack of emotional involvement should be given special attention when we consider using AI in, for example, child or elderly care. The second practical problem is the use of AI in activities that require foresight. Intelligence is not just about reacting to input from the environment. A more active, foresighted approach is often needed, going beyond actual experience and seeing less obvious, counterintuitive possibilities. Crying can express pain, joy and much more, but AI cannot easily foresee less obvious possibilities.
Two theoretical problems are also mentioned in the article. The first is whether AI in the future may have morally relevant characteristics such as autonomy, interests and preferences. The second problem is whether AI can affect human self-understanding and create uncertainty and anxiety about human identity. These theoretical problems undoubtedly require careful analysis – do we even know what we are asking? In philosophy we often need to clarify our questions as we go along.
An article in the journal Big Data & Society criticizes the form of ethics that has come to dominate research and innovation in artificial intelligence (AI). The authors question the same “framework interpretation” of ethics that you could read about on the Ethics Blog last week. However, with one disquieting difference. Rather than functioning as a fence that can set the necessary boundaries for development, the framework risks being used as ethics washing by AI companies that want to avoid legal regulation. By referring to ethical self-regulation – beautiful declarations of principles, values and guidelines – one hopes to be able to avoid legal regulation, which could set important limits for AI.
The problem with AI ethics as “soft ethics legislation” is not just that it can be used to avoid necessary legal regulation of the area. The problem is above all, according to the SIENNA researchers who wrote the article, that a “law conception of ethics” does not help us to think clearly about new situations. What we need, they argue, is an ethics that constantly renews our ability to see the new. This is because AI is constantly confronting us with new situations: new uses of robots, new opportunities for governments and companies to monitor people, new forms of dependence on technology, new risks of discrimination, and many other challenges that we may not easily anticipate.
The authors emphasize that such eye-opening AI ethics requires close collaboration with the social sciences. That, of course, is true. Personally, I want to emphasize that an ethics that renews our ability to see the new must also be philosophical in the deepest sense of the word. To see the new and unexpected, you cannot rest comfortably in your professional competence, with its established methods, theories and concepts. You have to question your own disciplinary framework. You have to think for yourself.
Scientific discovery is based on the novelty of the questions you ask. This means that if you want to discover something new, you probably have to ask a different question. And since different people have different preconceptions and experiences than you, they are likely to formulate their questions differently. This makes a case for diversity in research, If we want to make new discoveries that concern diverse groups, diversity in research becomes even more important.
The Human Brain Project participated in the FENS 2020 Virtual Forum this summer, an international virtual neuroscience conference that explores all domains in modern brain research. For the Human Brain Project (HBP), committed to responsible research and innovation, this includes diversity. Which is why Karin Grasenick, Coordinator for Gender and Diversity in the HBP, explored the relationship between diversity and new discovery in the session “Of mice, men and machines” at the FENS 2020.
So why is diversity in research crucial to make new discoveries? Research depends on the questions asked, the models used, and the details considered. For this reason, it is important to reflect on why certain variables are analysed, or which aspects might play a role. An example is Parkinson’s disease, where patients are affected differently depending on both age and gender. Being a (biological) man or woman, old or young is important for both diagnosis and treatment. If we know that diversity matters in research on Parkinson’s disease, it probably should do so in most neuroscience. Apart from gender and age, we also need to consider other aspects of diversity, like race, ethnicity, education or social background. Because depending on who you are, biologically, culturally and socially, you are likely to need different things.
A quite recent example for this is Covid-19, which does not only display gender differences (as it affects more men than women), but also racial differences: Black and Latino people in the US have been disproportionately affected, regardless of their living area (rural or urban) or their age (old or young). Again, the reasons for this are not simply biologically essentialist (e.g. hormones or chromosomes), but also linked to social aspects such as gendered lifestyles (men are more often smokers than women), inequities in the health system or certain jobs which cannot be done remotely (see for example this BBC Future text on why Covid-19 is different for men and women or this one on the racial inequity of coronavirus in The New York Times).
Another example is Machine Learning. If we train AI on data that is not representative of the population, we introduce bias in the algorithm. For example, applications to diagnose skin cancer in medicine more often fail to recognize tumours in darker skin correctly because they are trained using pictures of fair skin. There are several reasons for not training AI properly, it could be a cost issue, lack of material to train the AI on, but it is not unlikely that people with dark skin are discriminated because scientists and engineers simply did not think about diversity when picking material for the AI to train on. In the case of skin cancer, it is clear that diversity could indeed save lives.
But where to start? When you do research, there are two questions that must be asked: First, what is the focus of your research? And second, who are the beneficiaries of your research?
Whenever your research focus includes tissues, cells, animals or humans, you should consider diversity factors like gender, age, race, ethnicity, and environmental influences. Moreover, any responsible scientist should consider who has access to their research and profits from it, as well as the consequences their research might have for end users or the broader public.
However, as a researcher you need to consider not only the research subjects and the people your results benefit. The diversity of the research team also matters, because different people perceive problems in different ways and use different methods and processes to solve them. Which is why a diverse team is more innovative.
This is a guest blog post from the Human Brain Project (HBP). The HBP as received funding from the European Union’s Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement No. 945539 (Human Brain Project SGA3).
Given this enormous and growing self-knowledge, why do we not develop artificial intelligence that supports a morally limping humanity? Why spend so much resources on developing even more intelligent artificial intelligence, which takes our jobs and might one day threaten humanity in the form of uncontrollable superintelligence? Why do we behave so unwisely when we could develop artificial intelligence to help us humans become superethical?
How can AI make morally weak humans super-ethical? The authors suggest a comparison with the fitness apps that help people to exercise more efficiently and regularly than they otherwise would. The authors’ suggestion is that our ethical knowledge of moral theories, combined with our growing scientific knowledge of moral weaknesses, can support the technological development of moral crutches: wise objects that support people precisely where we know that we are morally limping.
My personal assessment of this utopian proposal is that it might easily be realized in less utopian form. AI is already widely used as a support in decision-making. One could imagine mobile apps that support consumers to make ethical food choices in the grocery shop. Or computer games where consumers are trained to weigh different ethical considerations against each another, such as animal welfare, climate effects, ecological effects and much more. Nice looking presentations of the issues and encouraging music that make it fun to be moral.
The philosophical question I ask is whether such artificial decision support in shops and other situations really can be said to make humanity wiser and more ethical. Imagine a consumer who chooses among the vegetables, eagerly looking for decision support in the smartphone. What do you see? A human who, thanks to the mobile app, has become wiser than Socrates, who lived long before we knew as much about ourselves as we do today?
Ethical fitness apps are conceivable. However, the risk is that they spread a form of self-knowledge that flies above ourselves: self-knowledge suspiciously similar to the moral vice of self-satisfied presumptuousness.
Academic research is driven by dissemination of results to peers at conferences and through publication in scientific journals. However, research results belong not only to the research community. They also belong to society. Therefore, results should reach not only your colleagues in the field or the specialists in adjacent fields. They should also reach outside the academy.
Who is out there? A homogeneous public? No, it is not that simple. Communicating research is not two activities: first communicating the science to peers and then telling the popular scientific story to the public. Outside the academy, we find engineers, entrepreneurs, politicians, government officials, teachers, students, research funders, taxpayers, healthcare professionals… We are all out there with our different experiences, functions and skills.
Research communication is therefore a strategically more complicated task than just “reaching the public.” Why do you want to communicate your results; why are they important? Who will find your results important? How do you want to communicate them? When is the best time to communicate? There is not just one task here. You have to think through what the task is in each particular case. For the task varies with the answers to these questions. Only when you can think strategically about the task can you communicate research responsibly.
Josepine Fernow’s contribution is, in my view, more than a convincing argument. It is an eye-opening text that helps researchers see more clearly their diverse relationships to society, and thereby their responsibilities. The academy is not a rock of knowledge in a sea of ignorant lay people. Society consists of experienced people who, because of what they know, can benefit from your research. It is easier to think strategically about research communication when you survey your relations to a diversified society that is already knowledgeable. Josepine Fernow’s argumentation helps and motivates you to do that.
Josepine Fernow also warns against exaggerating the significance of your results. Bioscience has potential to give us effective treatments for serious diseases, new crops that meet specific demands, and much more. Since we are all potential beneficiaries of such research, as future patients and consumers, we may want to believe the excessively wishful stories that some excessively ambitious researchers want to tell. We participate in a dangerous game of increasingly unrealistic hopes.
The name of this dangerous game is hype. Research hype can make it difficult for you to continue your research in the future, because of eroded trust. It can also make you prone to take unethical shortcuts. The “huge potential benefit” obscures your judgment as a responsible researcher.
Responsible research communication is as important as difficult. Therefore, these tasks deserve our greatest attention. Read Josepine Fernow’s argumentation for carefully planned communication strategies. It will help you see more clearly your responsibility.
Anthropomorphism almost seems inscribed in research on artificial intelligence (AI). Ever since the beginning of the field, machines have been portrayed in terms that normally describe human abilities, such as understanding and learning. The emphasis is on similarities between humans and machines, while differences are downplayed. Like when it is claimed that machines can perform the same psychological tasks that humans perform, such as making decisions and solving problems, with the supposedly insignificant difference that machines do it “automated.”
The article draws particular attention to so-called brain-inspired AI research, where technology development draws inspiration from what we know about the functioning of the brain. Here, close relationships are emphasized between AI and neuroscience: bonds that are considered to be decisive for developments in both fields of research. Neuroscience needs inspiration from AI research it is claimed, just as AI research needs inspiration from brain research.
The article warns that this idea of a close relationship between the two fields presupposes an anthropomorphic interpretation of AI. In fact, brain-inspired AI multiplies the conceptual double exposures by projecting not only psychological but also neuroscientific concepts onto machines. AI researchers talk about artificial neurons, synapses and neural networks in computers, as if they incorporated artificial brain tissue into the machines.
An overlooked risk of anthropomorphism in AI, according to the authors, is that it can conceal essential characteristics of the technology that make it fundamentally different from human intelligence. In fact, anthropomorphism risks limiting scientific and technological development in AI, since it binds AI to the human brain as privileged source of inspiration. Anthropomorphism can also entice brain research to uncritically use AI as a model for how the brain works.
Of course, the authors do not deny that AI and neuroscience mutually support each other and should cooperate. However, in order for cooperation to work well, and not limit scientific and technological development, philosophical thinking is also needed. We need to clarify conceptual differences between humans and machines, brains and computers. We need to free ourselves from the tendency to exaggerate similarities, which can be more verbal than real. We also need to pay attention to deep-rooted differences between humans and machines, and learn from the differences.
Anthropomorphism in AI risks encouraging irresponsible research communication, the authors further write. This is because exaggerated hopes (hype) seem intrinsic to the anthropomorphic language. By talking about computers in psychological and neurological terms, it sounds as if these machines already essentially functioned as human brains. The authors speak of an anthropomorphic hype around neural network algorithms.
Philosophy can thus also contribute to responsible research communication about artificial intelligence. Such communication draws attention to exaggerated claims and hopes inscribed in the anthropomorphic language of the field. It counteracts the tendency to exaggerate similarities between humans and machines, which rarely go as deep as the projected words make it sound.
In short, differences can be as important and instructive as similarities. Not only in philosophy, but also in science, technology and responsible research communication.
I recently read an article about so-called moral robots, which I found clarifying in many ways. The philosopher John-Stewart Gordon points out pitfalls that non-ethicists – robotics researchers and AI programmers – may fall into when they try to construct moral machines. Simply because they lack ethical expertise.
The first pitfall is the rookie mistakes. One might naively identify ethics with certain famous bioethical principles, as if ethics could not be anything but so-called “principlism.” Or, it is believed that computer systems, through automated analysis of individual cases, can “learn” ethical principles and “become moral,” as if morality could be discovered experientially or empirically.
The second challenge has to do with the fact that the ethics experts themselves disagree about the “right” moral theory. There are several competing ethical theories (utilitarianism, deontology, virtue ethics and more). What moral template should programmers use when getting computers to solve moral problems and dilemmas that arise in different activities? (Consider self-driving cars in difficult traffic situations.)
The first pitfall can be addressed with more knowledge of ethics. How do we handle the second challenge? Should we allow programmers to choose moral theory as it suits them? Should we allow both utilitarian and deontological robot cars on our streets?
John-Stewart Gordon’s suggestion is that so-called machine ethics should focus on the similarities between different moral theories regarding what one should not do. Robots should be provided with a binding list of things that must be avoided as immoral. With this restriction, the robots then have leeway to use and balance the plurality of moral theories to solve moral problems in a variety of ways.
In conclusion, researchers and engineers in robotics and AI should consult the ethics experts so that they can avoid the rookie mistakes and understand the methodological problems that arise when not even the experts in the field can agree about the right moral theory.
All this seems both wise and clarifying in many ways. At the same time, I feel genuinely confused about the very idea of ”moral machines” (although the article is not intended to discuss the idea, but focuses on ethical challenges for engineers). What does the idea mean? Not that I doubt that we can design artificial intelligence according to ethical requirements. We may not want robot cars to avoid collisions in city traffic by turning onto sidewalks where many people walk. In that sense, there may be ethical software, much like there are ethical funds. We could talk about moral and immoral robot cars as straightforwardly as we talk about ethical and unethical funds.
Still, as I mentioned, I feel uncertain. Why? I started by writing about “so-called” moral robots. I did so because I am not comfortable talking about moral machines, although I am open to suggestions about what it could mean. I think that what confuses me is that moral machines are largely mentioned without qualifying expressions, as if everyone ought to know what it should mean. Ethical experts disagree on the “right” moral theory. However, they seem to agree that moral theory determines what a moral decision is; much like grammar determines what a grammatical sentence is. With that faith in moral theory, one need not contemplate what a moral machine might be. It is simply a machine that makes decisions according to accepted moral theory. However, do machines make decisions in the same sense as humans do?
Maybe it is about emphasis. We talk about ethical funds without feeling dizzy because a stock fund is said to be ethical (“Can they be humorous too?”). There is no mythological emphasis in the talk of ethical funds. In the same way, we can talk about ethical robot cars without feeling dizzy as if we faced something supernatural. However, in the philosophical discussion of machine ethics, moral machines are sometimes mentioned in a mythological way, it seems to me. As if a centaur, a machine-human, will soon see the light of day. At the same time, we are not supposed to feel dizzy concerning these brave new centaurs, since the experts can spell out exactly what they are talking about. Having all the accepted templates in their hands, they do not need any qualifying expressions!
I suspect that also ethical expertise can be a philosophical pitfall when we intellectually approach so-called moral machines. The expert attitude can silence the confusing questions that we all need time to contemplate when honest doubts rebel against the claim to know.