Category: AI and Ethics

Your brain on algorithms

Article by Andrew Sears, Founder, Technovirtuism and Advisor, All Tech is Human

A quote attributed to the linguist Benjamin Lee Whorf claims that “language shapes the way we think and determines what we can think about.” Whorf, along with his mentor Edward Sapir, argued that a language’s structure influences its speakers’ perception and categorization of experience. Summarizing what he called the “linguistic relativity principle,” Whorf wrote that “language is not simply a reporting device for experience but a defining framework for it.”[1]

Over the past several decades, a strange new language has increasingly exerted its power over our experience of the world: the language of algorithms. Today, millions of computer programmers, data scientists, and other professionals spend the majority of their time writing and thinking in this language. As algorithms have become inseparable from most people’s day-to-day lives, they have begun to shape our experience of the world in profound ways. Algorithms are in our pockets and on our wrists. Algorithms are in our hearts and in our minds. Algorithms are subtly reshaping us after their own image, putting artificial boundaries on our ideas and imaginations.   

Our Languages, Our Selves

To better understand how language effects our experience of the world, we can consider the example depicted in the 2016 science fiction movie Arrival. In the film, giant 7-legged aliens arrive on Earth to give humanity a gift – a “technology,” as they call it. This technology turns out to be their language, which is unlike any human language in that it is nonlinear: sentences are represented by circles of characters with no beginning or end. Word order is not relevant; each sentence is experienced as one simultaneous collection of meanings.

As the human protagonist becomes fluent in this language, her experience of the world begins to adapt to its patterns. She begins to experience time nonlinearly. Premonitions and memories become indistinguishable as her past, present, and future collapse into one another. The qualities of her language become the qualities of her thought, which come to define her experience.

A similar, if less extreme, transformation is occurring today as our modes of thought are increasingly shaped by the language of algorithms. Technologists spend most of their waking hours writing algorithms to organize and analyze data, manipulate human behavior, and solve all manner of problems. To keep up with rapidly-changing technologies, they exist in a state of perpetual language learning. Many are evaluated based on how many total lines of code they write.

Even those of us who do not speak the language of algorithms are constantly spoken to in this language by the devices we carry with us, as these devices organize and mediate our day-to-day experience of the world. To understand how this immersion in the language of algorithms shapes us, we need to name some of the peculiar qualities of this language.

Traditional human languages employ a wide range of linguistic tools to express the world in its complexity, such as narrative, metaphor, hyperbole, dialogue, or humour. Algorithms, by contrast, express the world through formulas. They break up complex problems into small, well-defined components and articulate a series of linear steps by which the problem can be solved. Algorithms – or, more accurately, the utility of algorithms – depend on the assumption that problems and their solutions can be articulated mathematically without losing any meaningful qualities in translation.

They assume that all significant variables can be accounted for, that the world is both quantifiable and predictable. They assume that efficiency should be preferred to friction, that elegance is more desirable than complexity, and that the optimized solution is always the best solution. For many speakers of this language, these assumptions have become axiomatic. Their understanding of the world has been re-shaped by the conventions and thought-patterns of a new lingua franca.

Algorithms and Technological Solutionism

Immersion in the language of algorithms inclines us to look for technological solutions to every kind of problem. This phenomenon was observed in 2013 by Evgeny Morozov, who described the folly of “technological solutionism” in his book, To Save Everything, Click Here:

“Recasting all complex social situations either as neat problems with definite, computable solutions or as transparent and self-evident processes that can be easily optimized–if only the right algorithms are in place!–this quest is likely to have unexpected consequences that could eventually cause more damage than the problems they seek to address.”[2]

An algorithmic approach to problem solving depends on the assumption that the effects of the algorithm – the outputs of the formula – can be predicted and controlled. This assumption becomes less plausible as the complexity of the problem increases. When technologists seek to write an algorithmic solution to a nuanced human problem – like crime prevention or biased hiring, for example – they are compelled to contort the problem into an oversimplified form in order to express it in the language of algorithms. This results, more often than not, in the “unintended consequences” that we’ve all heard so much about. Perhaps Wordsworth said it best, 200 years before Morozov:

“Our meddling intellect

Mis-shapes the beauteous forms of things:—

We murder to dissect.”[3]

A helpful example of all this is furnished by Amazon, who in 2014 implemented an algorithmic tool to automate the vetting of resumes. The tool was designed to rate job candidates on a scale from one to five in order to help managers quickly identify their top prospects. After one year of use, this tool was shelved for discrimination against female candidates.

This parable contains all the components of the technosolutionist narrative: a problem rife with human complexity is assigned an algorithmic solution in the pursuit of efficiency; the problem is oversimplified in a hubristic attempt to express it numerically; the algorithm fails to achieve its objective while creating some nasty externalities along the way.

Immersion in the language of algorithms doesn’t just shape our understanding of the world; it shapes our understanding of the self. Members of the “quantified self” community use a combination of wearable devices, internet-of-things sensors, and – in extreme cases – implants to collect and analyze quantitative data on their physical, mental, and emotional performance. Adherents of this lifestyle pursue “self-knowledge through numbers,” in dramatic contrast to their pre-digital forbearers who may have pursued self-knowledge through philosophy, religion, art, or other reflective disciplines.

The influence of algorithms on our view of the self finds more mainstream expressions as well. Psychologists and other experts have widely adopted computational metaphors to describe the functions of the human brain.[4] Such conventions have made their way into our common vernacular; we speak of needing to “process” information, of feeling “at-capacity” and needing to “recharge,” of being “programmed” by education and “re-wired” by novel experiences.     

To understand the extent to which algorithms have shaped our world, consider that even Silicon Valley’s rogues and reformers struggle to move beyond algorithmic thought patterns. This is evidenced by the fact that tech reformers almost uniformly adopt a utilitarian view of technological ethics, evaluating the ethical quality of a technology based on its perceived consequences.

The algorithmic fixation with quantification, measurement, outputs, and outcomes informs their ethical thinking. Other ethics frameworks – such as deontology or virtue ethics – are seldom applied to technological questions except by the occasional academic who has been trained in these disciplines. 

A Language Fit for Our World

When it comes to complex social problems, algorithmic solutions tend to function at best as band-aids, covering up human shortcomings but doing nothing to address the problem’s true causes. Unfortunately, technological solutionism’s disastrous track record has done little to dampen its influence on the tech industry and within society as a whole.

Contact tracing apps look to be an important part of the solution to the COVID-19 pandemic; but will they be able to compensate for the lack of civic trust and responsibility that has crippled America’s virus containment efforts? Predictive and surveillance policing, powered by artificial intelligence tools, promises to create a safer society; can these atone for the complex social and economic inequalities that drive crime and dictate who gets called a criminal? As November 2020 approaches, social media companies assure us that they’re ready to contain the spread of fake news this time around; are we as citizens ready to engage in meaningful civil discourse?

Benjamin Lee Whorf, mirroring the language of Wordsworth’s poem, wrote that “we dissect nature along lines laid down by our native language.”[5] “We cut nature up, organize it into concepts, and ascribe significances,” as our language dictates and allows. It is therefore critical that the languages we adopt are capable of appreciating the world in all of its messy, tragic, beautiful complexity. Algorithms are well-suited to many kinds of problems, but they have important limits. Rather than shrinking our world to fit within these limits, we must broaden our minds to comprehend the intricacies of the world. We must stop thinking like algorithms and start thinking like humans.

References

[1] Whorf B.L. (1956). Language, Thought, and Reality. Cambridge, MA: MIT Press.

[2] Morozov, Evgeny. To Save Everything, Click Here.

[3] Wordsworth, William (2004). The Tables Turned. In Selected Poems, London: Penguin Books.

[4] For an excellent discussion and refutation of this trend, see Nicholas Carr’s book, The Shallows.

[5] Whorf, Benjamin (1956), Carroll, John B. (ed.), Language, Thought, and Reality: Selected Writings of Benjamin Lee Whorf. MIT Press: 212 – 214.

Translating AI ethics principles into action through digital cooperation

Article by Prateek Sibal, Specialist Consultant on AI at UNESCO and Lecturer in Digital Innovation and Transformation at Sciences Po, Paris

As a general-purpose technology, Artificial Intelligence (AI) impacts all spheres of our lives. A recent study on the role of AI in achieving the Sustainable Development Goals shows that AI could enable the accomplishment of 134 targets out of the total 169 targets under the 17 Sustainable Development Goals (SDGs) that cut across society, economy and the environment. For instance, AI can help deliver food and water security and education and health services efficiently. 

At the same time, AI has the potential to hinder progress on 59 targets concerning environment and society. On the environment front, some projections suggest that by 2030, information and communication technologies (ICTs) may account for up to 20 percent of global electricity consumption as compared to 1 per cent at present. If more environmentally friendly ICTs are not developed, it would adversely impact the achievement of SDGs on Climate Change. 

On the societal aspects, the development and use of AI is raising concerns about privacy, freedom of expression, discrimination, transparency and unequal access to technology. These factors have the potential to exacerbate existing social and economic inequalities between and within countries. 

How are governments responding to AI?

Governments across the world are scrambling to leverage the large economic potential of AI that includes an estimated USD 13 trillion in additional economic activity by 2030. At the same time, they are attempting to maximize the gains and minimizing the harms when dealing with the implications of AI for democracy, labour and the environment. As of 2019, governments, inter-governmental organizations, businesses, NGOs and academia have articulated over 84 AI guidelines or ethical principles.

These ethical guidelines, originating mostly in developed countries, cluster around principles of transparency, justice and fairness, non-maleficence, responsibility and privacy.  The principles have been adopted by countries at inter-governmental platforms like the OECD, European Union, G20, G7, and discussed by technical experts and academics at international conferences, civil society organizations through participatory processes with involvement of citizens or through expert groups constituted by private sector companies and alliances. 

Why ethical guidelines are not enough?

While the large number of ethical guidelines emerging at the national and regional level and by different stakeholder groups help in crystallizing a broad consensus on the values that should be respected in the development and use of AI globally, it leaves the question of translating principles to action insufficiently addressed. 

Ethical guidelines for AI have been criticized for being too high level, conflicting with each other or even meaningless in informing concrete actions that governments, engineers, businesses or justice systems are expected to take.  

Moving from principles to real questions that AI designers and regulators face, researchers at the University of Cambridge, highlight some of the tensions between different principles that remain unresolved. For instance:

— Should data be used to improve the quality and efficiency of services while compromising the privacy and autonomy of individuals?

— Should AI developers write algorithms that make decisions and predictions more accurate but may rank low on ensuring fairness and equal treatment?

— How to balance the benefits of increased personalization in the digital sphere with enhancing solidarity and citizenship?

— How to balance the use of automation to make people’s lives more convenient with empowering people by promoting self- actualization and dignity?

As an illustration, it may be useful to consider the principle of transparency that features in many AI ethics guidelines. The meaning of transparency vis. a vis. AI systems changes with the level of abstraction or the context in which it is applied.  

In the development of AI systems, transparency may be facilitated through transparent procedures that record decisions taken by developers and businesses concerning the objective function of the algorithms, the data used for training and arriving at decisions taken by algorithms, among others. 

However, when it comes to the use of AI systems, the norms of transparency may concern disclosures of how the user’s information would be processed, user consent and maintenance of decision-making logs for ex-post analysis of the decision for any bias and discrimination. 

Finally, transparency principle may also apply to internal technological processes of AI systems pertaining to the black-box problem of AI, where the decisions taken by neural networks or support vector machines are not understandable by humans. In this case, any measures to ensure transparency may involve technical solutions in the form of explainable AI among others. 

How to translate AI ethics principles to action?

Several governments, international organizations and standards organizations are alive to these concerns and are working towards translating high-level commitments into actionable requirements and norms. For instance, the OECD,  through its OECD Network of Experts and the AI Observatory is translating its AI Principles into concrete advice for policymakers. The standards body, IEEE is developing standards and certifications for AI systems as part of its Ethics of Autonomous and Intelligent Systems programme.

Some inspiration is also sought from the field of medicine, where high-level ethical principles are translated into action through tools like “professional societies and boards, ethics review committees, accreditation and licensing schemes, peer self-governance, codes of conduct, and other mechanisms supported by strong institutions help determine the ethical acceptability of day-to-day practice by assessing difficult cases, identifying negligent behaviour and sanctioning bad actors”.

However, the general purpose nature of AI, difficulty in creating professional bodies and institutional mechanisms for oversight makes implementation of ethical guidelines difficult to implement. 

Therefore, solutions may lie in translating high-level principles into “mid-level norms and low-level requirements”. Furthermore, these “norms and requirements cannot be deduced directly from mid-level principles without accounting for specific elements of the technology, application, context of use or relevant local norm”. 

Articulation of a hierarchical framework that translates principles to mid-level norms and low-level requirements as per technology, application and context of use remains a challenge because of the evolving nature of technology and the difficulty in anticipating relevant contexts. Further challenges are posed by gaps in human and institutional capacities for technology governance at the national and local levels.

Therefore, the work of translating AI ethics guidelines at the international level requires enhanced international cooperation and a robust mechanism for provision, monitoring and accountability for digital public goods. Implementation of these guidelines would require programmes for raising awareness, and strengthening capacities for monitoring and evaluation at all levels.

The UN Secretary General’s Roadmap on Digital Cooperation, developed over a two-year long consultative process, provides a good starting point to address some of the governance challenges related to AI. The roadmap aims to facilitate digital cooperation amongst transnational actors through its different working groups on digital public goods and human-centered AI, among others. Furthermore, it proposes strengthening policy advice and extending capacity building support to governments through digital help desks at the national level. 

Note: The views expressed in this article are those of the author and not of any affiliated organization.

Is no privacy the best AI privacy policy?

Article by James Brusseau, Director, Data Ethics Site, Pace University, New York

“You have zero privacy anyway,” Scott McNealy, CEO of Sun Microsystems scolded a concerned journalist in 1999, “Get over it.” He was wrong then. And if he said it again today, he would still be wrong. Tomorrow, though, is not so clear.

Most AI policy discussions about privacy concern strategies for its preservation, and cutouts for the relaxation of safeguards. So, we have policies banning facial recognition, and contact tracers that strip privacy in the name of fighting a pandemic.

No doubt these are the most important and currently pressing privacy issues, but they also hide the deeper and decisive question as technology increasingly pries into our lives: Should we be willing to accept full exposure, life without privacy?

No answer will be proposed here, but to ask the question seriously, we first need to get a sense of what the human experience would be if privacy were gone. That is the subject of this article, which begins about a decade before Apple computers went mainstream, in the 1970s with a philosophical thought experiment authored by Robert Nozick.

What he imagined was a was a floating tank of warm water, a sensory deprivation chamber where electrodes wrapped around our heads to feed our synapses a tantalizing experience indistinguishable from lived reality. You could choose your experiences beforehand – like choosing a book from the library – and then go into the tank.

Presumably, you could choose as many experiences as you would like and a single narrative could go on for years, decades. Nozick called this an Experience Machine, and the reason he formulated it was to ask a difficult hypothetical question: Would you permanently trade your outside life for an existence floating inside the tank?

If the answer is yes, every moment would be as thrilling or heroic or opulent as you could possibly wish. But, it would all only happen in your mind. It is a hard call. What is at stake, though, is easy to see: mental gratification versus personal freedom. You get prefabricated episodes guaranteed to feel good, exciting, or luxurious, at least in your mind. However, you lose the ability to create experiences and form a unique identity for yourself out in the unpredictable world.

A post-privacy reality offers an analogous choice: pleasure for freedom. It promises satisfactions, but requires relinquishing control over our own destinies. First, what does post-privacy mean? Private companies can purchase not only information about our age, location, recent credit card purchases and webpage visits, but also access all the text messages, telephone calls and pixels exchanged with every friend and romance.

There would be full access to medical records and work histories, and not just the topline statistics, but the history of every moment of every day as captured by video and audio surveillance. More could be added – all this is just the beginning of full exposure – but it is enough to begin sketching the pleasure or freedom dilemma.

As an initial and crude example, there are Netflix movies. By combing personal information about the viewer with predictive analytics, a film or episode is selected to begin rolling even before the previous one ends. The transition is almost seamless. Now, if Netflix knows everything about you – if there is no privacy – then it is probably going to be a good choice, one that holds your attention and wraps you in contentment as you keep watching.

It is also true, though, that your autonomy is getting limited. It is because you get what you want before making any choices: on one level, you don’t choose another movie from a list and then, above that, you don’t even choose whether to watch a movie at all because it’s already going. You are trapped in front of the screen. It’s a good trap because the confinement is your own data and entertainment, but it remains confinement.

One series that has been selected by the mechanisms of surveillance capitalism for many of the people reading this sentence is Black Mirror. There is an episode depicting a couple in a restaurant getting served their dishes just before asking to see the menu, and in a big data future of embraced transparency, that anticipatory provisioning should not be disconcerting but expected. Stronger, it is a central reason for relinquishing privacy and exposing ourselves to the uses and satisfactions of AI. It is good to get what you want before you want it.

Generalizing, it will not only be movie selections and the dinner choices. All our wants will be answered so immediately that we won’t even have time to understand why they are right, when we started wanting them, or to ask what it is that we wanted in the first place.

If the big data experience machine is functioning as it should – if it is instantly transforming complete personal information and predictive analytics into total consumer and user satisfactions – then we are not choosing anymore and, far more significantly, we’re not choosing to not choose. When we always already have what we want, no question about making a selection can even arise.

For that reason, one of the twisted curiosities about life after privacy is that the way we realize something is what we want is: we already have it. And that’s the only way we know that we want something. More, if we do feel the urge for a slice of pizza, or to binge Seinfeld, or to incite a romantic fling, what that really means is: we don’t actually want it. We can’t, since the core idea of the AI service economy is that it knows us transparently and so responds to our urges so perfectly that they’re answered without even a moment of suffering an unfulfilled craving.

It would be interesting to know whether something can be truly pleasurable if we get it before realizing a hunger for it, but no matter the answer, the drowning of personal desire in convenience and comfort is a significant temptation. On the other hand, there is no avoiding the cost that we no longer have personal freedom to make anything of ourselves.

Active self determination is gone since there is no room for experimenting with new possibilities, or for struggling with what’s worth pursuing and having. There are only the tranquil satisfactions that initially and powerfully recommended that we expose ourselves to big data reality and predictive analytics. Total exposure – reality without privacy – means a life so perfectly satisfying that we cannot do anything with it.

The unconditional surrender of our personal information is a big data Stockholm syndrome. Viewed from outside, users must seem enamored of the information sets and algorithms that control their experiences: those whose personal freedom has been arrested are actually grateful to their captors for the pleasures their minds experience.

But, from within the experience the very idea of being captive does not make sense since it is impossible to encounter – or even conceptualize – any kind of restraint. If we always get everything we want – movies, dinners, jobs, lovers, everything – before we even know we want them, how could we feel anything but total liberation?

What does Transparent AI mean?

Article by Fenna Woudstra, former intern at Filosofie in actie (an organization based in Arnhem, the Netherlands advising technology companies and projects on the ethical use of data)

Over the last couple of years many companies, institutions, governments, and ethicists have proposed their own principles for the development of Artificial Intelligence (or AI) in an ethical way. Most of the guidelines include principles about fairness, accountability, and transparency, but there is a wide range of interpretations and explanations of what these principles entail. This article delves into the transparency of AI.

What is transparency?

Because of the growing use of Artificial Intelligence (AI) in many different fields, problems like discrimination caused by automated decision-making algorithms are becoming more pressing. The underlying problem is the unknown existence of biases, due to the opacity of the systems.[1]

The logical response to reduce the problems caused by opacity is a greater demand for transparency. Transparency is all about an information exchange between a subject and an object, where the subject receives information about the performance of a system or organization the object is responsible for.[2]

However, putting transparency into practice is easier said than done. Firstly, not everything can simply be made publicly available. Other important values like privacy, national security, and corporate secrecy have to be taken into account.[3] Secondly, even after considering these points it remains unclear how transparency should be realized; what information about which aspects of the AI system should be disclosed for the system to deemed transparent?

Results of comparing 18 Ethical AI guidelines

To find this out, eighteen guidelines for Ethical AI have been analyzed on their principles and practical implementations of transparency, all retrieved from the AI Ethics Guidelines Global Inventory maintained by AlgorithmWatch.

As a result, a new framework has been created that consists of nine main topics: environmental costs, employment, business model, user’s rights, data, algorithms, auditability, trustworthiness, and openness (see Table 1 below). Each topic is explained by multiple practical specifications that were extracted from the existing guidelines. The topics cover the whole process of developing, using, and reviewing AI systems.

Table 1. Framework for Transparent AI

Some critical notes

Unexpectedly, the principles of transparency do not only address algorithmic transparency or the explainability of the AI system, but also the transparency of the company itself. This is most notable in the first two topics, namely, environmental costs and employment. They show how a system is made, at what costs, and under which circumstances, are also a part of the development of a system.

Being transparent about these aspects could probably result in a greater trust and acceptance; Virginia Dignum, Professor of Social and Ethical Artificial Intelligence at University of Umeå in Sweden has already mentioned that “trust in the system will improve if we can ensure openness of affairs in all that is related to the system. As such, transparency is also about being explicit and open about choices and decisions concerning data sources and development processes”.[4]

Furthermore, it should be mentioned that not all existing guidelines have been analyzed and neither can we be sure that the existing guidelines already include all the important principles, due to the novelty of this subject. So clearly, additions to this framework can be made as and when important principles are found to be missing.

That is why this framework is not meant to provide strict rules, but it rather aims at providing an overview of the possible topics to be transparent about. Organizations can decide which aspects are relevant for their system and organization.

This framework also tries to show that being transparent does not necessarily mean that everything must be made publicly available; explaining why a principle cannot be fulfilled is also a form of transparency. Not publishing the used data can be perfectly ethical if it is done to ensure people’s privacy or the national security, as the American economist, Joseph Stiglitz has already noted.[5]

Finally, this framework shows that even when an algorithm’s working is highly complex, there could be still many aspects to be transparent and many ways to achieve transparency that would help reveal or prevent biases in AI systems. There is much more to an AI system than only a black box.

Fenna Woudstra did research on the Transparency of AI during her internship at Filosofie in actie (the Netherlands). This article is a summary of her findings. You can read her full research paper here.

References

[1] O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Broadway Books. See also: Lepri, B., Oliver, N., Letouzé, E., Pentland, A. and Vinck, P. (2018). Fair, transparent, and accountable algorithmic decision-making processes. Philosophy & Technology, 31(4), 611-627.

[2] Meijer, A. (2013). Understanding the complex dynamics of transparency. Public Administration Review, 73(3), 429-439.

[3] Stiglitz, J. E. (1999). On liberty, the right to know, and public discourse: the role of transparency in public life. Globalizing rights: The Oxford amnesty lectures, 149.

[4] Dignum, V. (2019). Responsible Artificial Intelligence: How to Develop and Use AI in a Responsible Way. Springer International Publishing.

[5] Stiglitz, J. E. (1999). On liberty, the right to know, and public discourse: the role of transparency in public life. Globalizing rights: The Oxford amnesty lectures, 149.