Category Archives: science, technology and society

can AI solve “wicked problems”?

Posted on by
Reply

I’ve been reading predictions that artificial intelligence will wipe out swaths of jobs–see Josh Tyrangiel in The Atlantic or Jan Tegze. Meanwhile, this week, I’m teaching Rittel & Webber (1973), the classic article that coined the phrase “wicked problems.” I started to wonder whether AI can ever resolve wicked problems. If not, the best way to find an interesting job in the near future may be to specialize in wicked problems. (Take my public policy course!)

According to Rittel & Webber, wicked problems have the following features:

  1. They have no definitive formulation.
  2. There is no stopping rule, no way to declare that the issue is done.
  3. Choices are not true or false, but good or bad.
  4. There is no way to test the chosen solution (immediate or ultimate).
  5. It is impossible, or unethical, to experiment.
  6. There is no list of all possible solutions.
  7. Since each problem is unique, inductive reasoning can’t work.
  8. Each problem is a symptom of another one.
  9. You can choose the explanations, and they affect your proposals.
  10. You have no “No right to be wrong.” (You are affecting other people, not just yourself. And the results are irreversible.)

Rittel and Webber argue that those features of wicked problems deflate the 20th-century ideal of a “planning system” that could be automated:

Many now have an image of how an idealized planning system would function. It is being seen as an on-going, cybernetic process of governance, incorporating systematic procedures for continuously searching out goals; identifying problems; forecasting uncontrollable contextual changes; inventing alternative strategies, tactics, and time-sequenced actions; stimulating alternative and plausible action sets and their consequences; evaluating alternatively forecasted outcomes; statistically monitoring those conditions of the publics and of systems that are judged to be germane; feeding back information to the simulation and decision channels so that errors can be corrected–all in a simultaneously functioning governing process. That set of steps is familiar to all of us, for it comprises what is by now the modern-classical mode planning. And yet we all know that such a planning system is unattainable, even as we seek more closely to approximate it. It is even questionable whether such a planning system is desirable (p. 159)

Here they describe planning systems that would have been very labor-intensive in 1973, but many people today imagine that this is how AI works, or will work.

why are problems wicked?

Some of the 10 reasons that some problems are “wicked,” according to Rittel & Webber, relate to the difficulty of generating knowledge. Policy problems involve specific things that have many features or aspects and that relate to many other specific things. For example, a given school system has a vast and unique set of characteristics and is connected by causes and effects to other systems and parts of society. These qualities make a school system difficult to study in conventional, scientific ways. However, could a massive LLM resolve that problem by modeling a wide swath of the society?

Another reason that problems are wicked is that they involve moral choices. In a policy debate, the question is not what would happen if we did something but what should happen. When I asked ChatGPT whether AI will be able to resolve wicked problems, it told me no, because wicked problems “are value-laden.” It added, “AI can optimize for values, but it cannot choose them in a legitimate way. Deciding whose values count, how to weigh them, and when to revise them is a normative, political act, not a computational one.”

Claude was less explicit about this point but emphasized that “stakeholders can’t even agree on what the problem actually is.” Therefore, an AI agent cannot supply a definitive answer.

A third source of the difficulty of wicked problems involves responsibility and legitimacy. In their responses to my question, both ChatGPT and Claude implied that AI models should not resolve wicked problems because they don’t have the right or the standing to do so.

what’s our underlying theory of decision-making?

Here are three rival views of how people decide value questions:

First, perhaps we are creatures who happen to want some things and abhor other things. We experience policies and their outcomes with pleasure, pain, or other emotions. It is better for us to get what we want–because of our feelings. Since an AI agent doesn’t feel anything, it can’t really want anything; and if it says it does, we shouldn’t care. Since we disagree about what we want, we must decide collectively and not offload the decision onto a computer.

Some problems with this view: People may want very bad things–should their preferences count? If we just happen to want various things, is there any better way to make decisions than to maximize as many subjective preferences as possible? Couldn’t a computer do that? But would the world be better if we did maximize subjective preferences?

In any case, you are not going to find a job making value-judgments. Today, lots of people are paid to make decisions, but only because they are assumed to know things. Nobody will pay for preferences. Life works the other way around: you have to pay to get your preferences satisfied.

Second, perhaps value questions have right and wrong answers. A candidate for the right answer would be utilitarianism: maximize the total amount of welfare. Maybe this rule needs constraints, or we should use a different rule. Regardless, it would be possible for a computer to calculate what is best for us. In fact, a machine can be less biased than humans.

Some problems with this view: We haven’t resolved the debate about which algorithm-like method should be used to decide what is right. Furthermore, I and others doubt that good moral reasoning is algorithmic. For one thing, it appears to be “holistic” in the specific sense that the unit of assessment is a whole object (such as a school or a market), not separate variables.

Third, perhaps all moral opinions are strictly subjective, including the opinion that we should maximize the satisfaction of everyone’s subjective opinions. Then it doesn’t matter what we do. We could outsource decisions to a computer, or just roll a die.

The problem with this view: It certainly does matter what we do. If not, we might as well pack it in.

AI as a social institution

I am still tentatively using the following model. AI is not like a human brain; it is like a social institution. For instance, medicine aggregates vast amounts of information and huge numbers of decisions and generates findings and advice. A labor market similarly processes a vast number of preferences and decisions and yields wages and employment rates. These are familiar examples of entities that are much larger than any human being–and they can feel impersonal or even cruel–but they are composed of human inputs, rules, and some hardware.

Another interesting example: integrated assessment models (IAMs) for predicting the global impact of carbon emissions and the costs and benefits of proposed remedies. These models have developed collaboratively and cumulatively for half a century. They take in thousands of peer-reviewed findings about specific processes (deforestation in Brazil, tax credits in Germany) and integrate them mathematically. No human being can understand even a tiny proportion of the data, methods, and instruments that generate the IAMs as a whole. But an IAM is a human product.

A large language model (LLM) is similar. At a first approximation, it is a machine that takes in lots of human generated text, processes it according to rules, and generates new text. Just the same could be said of science or law. This description actually understates the involvement of humans, because we do not merely produce the text that the LLM processes to generate output. We also conceive the idea of an LLM, write the software, build the hardware, construct the data centers, manage the power plants, pour the cement, and otherwise work to make the LLM.

If this is the case, then a given AI agent is not fundamentally different from a given social institution, such as a scientific discipline, a market, a body of law, or a democracy. Like these other institutions, it can address complexity, uncertainty, and disagreements about values. We will be able to ask it for answers to wicked problems. If current LLMs like ChatGPT and Claude refuse to provide such answers, it is because their authors have chosen–so far–to tell them not to.

However, AI’s rules are different from those in law, democracy, or science. I am biased to think that its rules are worse, although that could be contested. The threat is that AI will start to generate answers to wicked problems, and we will accept its answers because our own responses are not definitively better and because it responds instantly at low cost. But then we will lose not only the vast array of jobs that involve decision-making but also the intrinsic value of being decision-makers.

Source: Rittel, Horst WJ, and Melvin M. Webber. “Dilemmas in a general theory of planning.” Policy sciences 4.2 (1973): 155-169. See also: the human coordination involved in AI; the difference between human and artificial intelligence: relationships; the age of cybernetics; choosing models that illuminate issues–on the logic of abduction in the social sciences and policy

why policy debates continue

Posted on by
Reply

I’m at Stanford today to discuss a paper, Policy Models as Networks of Beliefs. After circulating my draft, I realized that the following is really my argument. …

We use mental models to think about and discuss contested questions of policy. Worthy models typically have these features:

  1. They have many components, not just a few. A model might include a causal inference, such as “spending more on x produces better outcomes.” But those two components (the spending and the outcomes) must be part of a much larger model that also explains why certain outcomes are valuable, where the money would come from, what else effects the system, and so on.
  2. The components should be connected, and the resulting structure matters. Structures can take various forms (e.g., root-cause analysis, vicious cycles). There is no single best structure.
  3. Pieces of models may prove regular. For instance, maybe spending more on x regularly produces better outcomes, all else considered. But such regularities only apply to small aspects of good models. The science-like effort to find regularities can only get us so far.
  4. Some components of any worthy model should be values or normative claims. Some normative components have regular significance in all models. However, many value components change their significance depending on the context. Equality, for example, does not consistently mean the same thing and may not always be desirable.
  5. If a model proves influential, it can change the world, which can require a new model. For example, arguing that more money should be spent on X could cause more funds to be allocated to X, at which point it would no longer be wise to increase the funding. Models are dynamic in this sense.

I believe this account supports a pluralistic, polycentric, pragmatist, and deliberative approach to policymaking, as opposed to a positivistic one.

See also: choosing models that illuminate issues–on the logic of abduction in the social sciences and policy; different kinds of social models; social education as learning to improve models; etc.

Cezanne’s portait of Gustave Geffroy

Posted on by
Reply

In “Cézanne’s Doubt” (1946), Maurice Merleau-Ponty discusses Paul Cézanne’s portrait of the critic Paul Geffroy (1895-6), which led me to some congruent reflections.

Merleau-Ponty notes that the table “stretches, contrary to the laws of perspective, into the lower part of the picture.” In a photograph of M. Geffroy, the table’s edges would form parallel lines that would meet at one point, and the whole object would be more foreshortened. That is how an artist who followed what we call “scientific perspective” would depict the table. Why does Cézanne show it otherwise?

Imagine that you actually stood before Paul Geffroy in his study. You would not instantly see the whole scene. Your eye might settle on your host’s face, then jump to the intriguing statuette next to him. The shelves would at first form a vague pattern in the background. Objects for which you have names, such as books, would appear outlined, as borders filled with color. On the other hand, areas of the fireplace or wall would blend into other areas.

You would know that you could move forward toward M. Geffroy, in which case the table would begin to move below you. Just as you see a flying ball as something moving–not as a round zone of color surrounded by other colors–so you might see the table as something that could shift if you moved your body forward.

A photograph of this real-world scene would be a representation of it, very useful for knowing how M. Geffroy looked in his study, and possibly an attractive object in its own right. But the photo would not represent anyone’s experience of the scene. Instead, it would be something that you could experience, rather like the scene itself, by letting your eye move around it, identifying objects of interest, and gradually adding information. You would experience the photograph somewhat differently from the actual scene because you would know that everything was fixed and your body could not move into the space.

A representation of this scene using perspective’s “laws” would make the image useful for certain purposes–for instance, for estimating the size of the table. Michael Baxandall (1978) argued that Renaissance perspective originated in a commercial culture in which patrons enjoyed estimating the size, weight, and value of objects represented in paintings.

But other systems have different benefits. Here is a print in which Toyoharu Kunichika (1835-1900) uses European perspective for the upper floor and a traditional Chinese system (with lines that remain parallel and objects placed higher if they are further away) for the lower floor. As Toshidama writes, this combination is useful for allowing us to see as many people and events as possible.

Print by Toyoharu Kunichika from Toshidama Japanese Prints

Perspective does not tell us how the world is–not in any simple way. The moon is not actually the size of a window, although it is represented as such in a perspectival picture (East Asian or European). Perspective is a way of representing how we experience the world. And in that respect, it is partial and sometimes even misleading. It overlooks that for us, important things seem bolder; objects can look soft, cold or painful as well as large or small; and some things appear in motion or likely to move, while others seem fixed. We can see a whole subject (such as a French intellectual in his study) and parts of it (his beard), at once and as connected to each other.

Merleau-Ponty writes:

Gustave Geoffrey’s [sic] table stretches into the bottom of the picture, and indeed, when our eye runs over a large surface, the images it successively receives are taken from different points of view, and the whole surface is warped. It is true that I freeze these distortions in repainting them on the canvas; I stop the spontaneous movement in which they pile up in perception and in which they tend toward the geometric perspective. This is also what happens with colors. Pink upon gray paper colors the background green. Academic painting shows the background as gray, assuming that the picture will produce the same effect of contrast as the real object. Impressionist painting uses green in the background in order to achieve a contrast as brilliant as that of objects in nature. Doesn’t this falsify the color relationship? It would if it stopped there, but the painter’s task is to modify all the other colors in the picture so that they take away from the green background its characteristics of a real color. Similarly, it is Cézanne’s genius that when the over-all composition of the picture is seen globally, perspectival distortions are no longer visible in their own right but rather contribute, as they do in natural vision, to the impression of an emerging order, of an object in the act of appearing, organizing itself before our eyes.

The deeper point is that a science of nature is not a science of human experience. Third-person descriptions or models of physical reality are not accounts of how we experience things. And even when we are presented with a scientific description, it is something that we experience. For instance, we actively interpret a photograph or a diagram; we do not automatically imprint all of its pixels. And we listen to a person lecture about science; we do not simply absorb the content.

There are truths that can be expressed in third-person form–for example, that human eyes and brains work in certain ways. But there are also truths about how we experience everything, including scientific claims.

And Cézanne is a scientist of experience.

Quotations from Maurice Merleau-Ponty, “Cézanne’s Doubt” (1946), in Sense and Non-sense, translated by Hubert L. Dreyfus and Patricia Allen Dreyfus (Northwestern University Press 1964); image by Paul Cézanne, public domain, via Wikimedia Commons. The image on the Mus?e d’Orsay’s website suggests a warmer palette, but I don’t know whether it’s open-source. I also refer to Michael Baxandall, Painting and Experience in Fifteenth Century Italy : A Primer in the Social History of Pictorial Style (Oxford, 1978).

See also: Svetlana Alpers, The Art of Describing; trying to look at Las Meninas; Wallace Stevens’ idea of order; an accelerating cascade of pearls (on Galileo and Tintoretto); and Rilke, “The Grownup.” My interactive novel, The Anachronist, is about perspective.

The post Cezanne’s portait of Gustave Geffroy appeared first on Peter Levine.

how thinking about causality affects the inner life

Posted on by
Reply

For many centuries, hugely influential thinkers in each of the Abrahamic faiths combined their foundational belief in an omnipotent deity with Aristotle’s framework of four kinds of causes. Many believers found solace when they discerned a divine role in the four causes.

Aristotle’s framework ran afoul of the Scientific Revolution. Today, there are still ways to be an Abrahamic believer who accepts science, and classical Indian thought offers some alternatives. Nevertheless the reduction of causes from Aristotle’s four to the two of modern science poses a spiritual and ethical challenge.

(This point is widely understood–and by no means my original contribution–but I thought the following summary might be useful for some readers.)

To illustrate Aristotle’s four causes, consider my hands, which are currently typing this blog post. Why are they doing that?

  • Efficient cause: Electric signals are passing along nerves and triggering muscles to contract or relax. In turn, prior electrical and mechanical events caused those signals to flow–and so on, back through time.
  • Material cause: My hand is made of muscles, nerves, skin, bones, and other materials, which, when so configured and stimulated, move. A statue’s hand that was made of marble would not move.
  • Formal cause: A hand is defined as “the terminal part of the vertebrate forelimb when modified (as in humans) as a grasping organ” (Webster’s dictionary). I do things like grasp, point, and touch with my hand because it is a hand. Some hands do not do these things–for instance, because of disabilities–but those are exceptions (caused by efficient causes) that interfere with the definitive form of a hand.
  • Final cause: I am typing in order to communicate certain points about Aristotle. I behave in this way because I see myself as a scholar and teacher whose words might educate others. In turn, educated people may live better. Therefore, I move my fingers for the end (telos, in Greek) of a good life.

Aristotle acknowledges that some events occur only because of efficient and material causes; these accidents lack ends. However, the four causes apply widely. For example, not only my hand but also the keyboard that I am using could be analyzed in terms of all four causes.

The Abrahamic thinkers who read Aristotle related the Creator to all the causes, but especially to the final cause (see Maimonides, Guide for the Perplexed, 2:1 or Aquinas, Summa Theologiae, I, Q44). In a well-ordered, divinely created universe, everything important ultimately happens for a purpose that is good. Dante concludes his Divine Comedy by invoking the final cause of everything, “the love that moves the sun and other stars.”

These Jewish and Christian thinkers follow the Muslim philosopher Avicenna, who even considers cases–like scratching one’s beard–that seem to have only efficient causes and not to happen for any end. “Against this objection, Avicenna maintains that apparently trivial human actions are motivated by unconscious desire for pleasure, the good of the animal soul” (Richardson 2020), which, in turn, is due to the creator.

However, writing in the early 1600s, Francis Bacon criticizes this whole tradition. He assigns efficient and material causes to physics, and formal and final causes to metaphysics. He gestures at the value of metaphysics for religion and ethics, but he doubts that knowledge can advance in those domains. His mission is to improve our understanding and control of the natural world. And for that purpose, he recommends that we keep formal and final causes out of our analysis and practice only what he calls “physics.”

It is rightly laid down that true knowledge is that which is deduced from causes. The division of four causes also is not amiss: matter, form, the efficient, and end or final cause. Of these, however, the latter is so far from being beneficial, that it even corrupts the sciences, except in the intercourse of man with man (Bacon, Novum Organum. P. F. Collier, 1620, II;2).

In this passage and others related to it, Bacon proved prescient. Although plenty of scientists after Bacon have believed in final causes, including divine ends, they only investigate efficient and material causes. Perhaps love moves all the stars, but in Newtonian physics, we strive to explain physical motion in terms of prior events and materials. This is a methodological commitment that yields what Bacon foresaw, the advancement of science.

The last redoubt of final causes was the biological world. My hand moves because of electrical signals, but it seemed that an object as complicated as a hand must have come into existence to serve an end. As Kant writes, “it is quite certain that in terms of purely mechanical principles of nature we cannot even adequately become familiar with, much less explain, organized beings and how they are internally possible.” Kant says that no Isaac Newton could ever arise who would be able to explain “how even a mere blade of grass is produced” using only “natural laws unordered by intention” (Critique of Judgment 74, Pluhar trans.). But then along came just such a Newton in the form of Charles Darwin, who showed that efficient and material explanations suffice in biology, too. A combination of random mutation plus natural selection ultimately yields objects like blades of grass and human hands.

A world without final causes–without ends–seems cold and pointless if one begins where Avicenna, Maimonides, and Aquinas did. One option is to follow Bacon (and Kant) by separating physics from metaphysics, aesthetics, and ethics and assigning the final causes to the latter subjects. Indeed, we see this distinction in the modern university, where the STEM departments deal with efficient causes, and final causes are discussed in some of the humanities. Plenty of scientists continue to use final-cause explanations when they think about religion, ethics, or beauty–they just don’t do that as part of their jobs.

However, Bacon’s warning still resonates. He suspects that progress is only possible when we analyze efficient and material causes. We may already know the final causes relevant to human life, but we cannot learn more about them. This is fine if everyone is convinced about the purpose of life. However, if we find ourselves disagreeing about ethics, religion, and aesthetics, then an inability to make progress becomes an inability to know what is right, and the result can be deep skepticism.

Michael Rosen (2022) reads both Rousseau and Kant as “moral unanimists”–philosophers who believe that everyone already knows the right answer about moral issues. But today hardly anyone is a “moral unanimist,” because we are more aware of diversity. Nietzsche describes the outcome (here, in a discussion of history that has become a science):

Its noblest claim nowadays is that it is a mirror, it rejects all teleology, it does not want to ‘prove’ anything any more; it scorns playing the judge, and shows good taste there, – it affirms as little as it denies, it asserts and ‘describes’ . . . All this is ascetic to a high degree; but to an even higher degree it is nihilistic, make no mistake about it! You see a sad, hard but determined gaze, – an eye peers out, like a lone explorer at the North Pole (perhaps so as not to peer in? or peer back? . . .). Here there is snow, here life is silenced; the last crows heard here are called ‘what for?’, ‘in vain’, ‘nada’ (Genealogy of Morals, Kaufman trans. 2:26)

Earlier in the same book, Nietzsche recounts how, as a young man, he was shaped by Schopenhauer’s argument that life has no purpose or design. But Nietzsche says he detected a harmful psychological consequence:

Precisely here I saw the great danger to mankind, its most sublime temptation and seduction – temptation to what? to nothingness? – precisely here I saw the beginning of the end, standstill, mankind looking back wearily, turning its will against life, and the onset of the final sickness becoming gently, sadly manifest: I understood the morality of compassion [Mitleid], casting around ever wider to catch even philosophers and make them ill, as the most uncanny symptom of our European culture which has itself become uncanny, as its detour to a new Buddhism? to a new Euro-Buddhism? to – nihilism? (Genealogy of Morals, Preface:6)

After mentioning Buddhism, Nietzsche critically explores the recent popularity of the great Buddhist virtue–compassion–in Europe.

Indeed, one of the oldest and most widely shared philosophical premises in Buddhism is “dependent origination,” which is the idea that everything happens because of efficient causes alone and not for teleological reasons. (I think that formal causes persist in Theravada texts but are rejected in Mahayana.)

Dependent origination is taken as good news. By realizing that everything we believe and wish for is the automatic result of previous accidental events, we free ourselves from these mental states. And by believing the same about everyone else’s beliefs and desires, we gain unlimited compassion for those creatures. Calm benevolence fills the mind and excludes the desires that brought suffering while we still believed in their intrinsic value. A very ancient verse which goes by the short title ye dharma hetu says (roughly): “Of all the things that have causes, the enlightened one has shown what causes them, and thereby the great renouncer has shown how they cease.”

I mention this argument not necessarily to endorse it. Much classical Buddhist thought presumes that a total release from the world of causation is possible, whether instantly or over aeons. If one doubts that possibility, as I do, then the news that there are no final causes is no longer consoling.

Secondary sources: Richardson, Kara, “Causation in Arabic and Islamic Thought”, The Stanford Encyclopedia of Philosophy (Winter 2020 Edition), Edward N. Zalta (ed.); Michael Rosen, The Shadow of God: Kant, Hegel, and the Passage from Heaven to History, Harvard University Press, 2022. See also how we use Kant today; does skepticism promote a tranquil mind?; does doubting the existence of the self tame the will?; spirituality and science; and the progress of science.

The post how thinking about causality affects the inner life appeared first on Peter Levine.