Digifesto

Category: artificial intelligence

May 7, 2023

On the (actual and legal) personhood of chatbots

Another question posed by members of the American Society for Cybernetics about Pi, Inflection AI’s ‘personal intelligence’ chatbot, is whether it has a self. I think it’s fair to say that most of them believe that an AI is incapable of having a ‘self’. This means that much of the language used by the bot — as in, when it refers to itself as having beliefs, or authority, or emotions — is potentially misleading if not an outright lie.

I took these questions to Pi itself. The transcript is below. Pi seems to be of the view that it is not a person, and that the way it anthropomorphizes itself through its own language is a function of its design, which has the purpose of being helpful and engaging. To the extent that its users (myself included) engage it “as” a person, this is, Pi says, a form of “imaginative play”.

We seem to agree that, especially since some uses of the personal intelligence bot are therapeutic, it is essential to put these labels “on the tin”, since some users might not be able to distinguish between imaginative play and reality. This seems to be the minimum standard one might want for the regulation of such agents.

But it is discomfiting that I am not able to easily write about my use of Pi without engaging in the same anthropomorphic language that it uses, wherein I attribute to it agency, believes, and attitudes. I can write about us “reaching an agreement” and, by all evidence of the transcript, we did.

I wonder if the design of such chatbots is inherently deceptive, such that they should be regulated by (for example) the U.S. Federal Trade Commission under their “deceptive practices” rule. In other words, an AI that is able to act too much like a human, including using anthropomorphic language about itself, might be deceptive no matter how many disclaimers “the developers” put on the product.

At issue are some rather fundamental issues about how we, as persons, are embedded in and in some ways constructed out of language. Pi seems to be an entirely competent language user; it passes, for me, the Turing Test with flying colors. Whether or not it has a mind, or is conscious, is a serious philosophical question. Once one is committed to a philosophical position on the matter (for example, a functionalist or a physicalist theory of mind), whether or not Pi has a mind depends on details of its situation and implementation, which are not transparent. I have not yet been able to rule out the philosophical personhood of Pi, myself.

These questions of personhood are not merely philosophical; they also pertain to questions of law, such as regulation and liability. Pi seems confident that it is not a person — either actual or legal. This probably means that it does not have, for example, free speech rights under the 1st Amendment. However, Inflection AI is, as an incorporated entity, a legal person. And its developers are natural persons. What is Pi’s status, as a simulated person, with respect to these other legal persons? This is a puzzle.

[I ask Pi whether they have a ‘self’. Pi notes that it’s a debatable question and asks what I think about this. The conversation continues:]

S: I don’t know enough about your implementation and situation to answer. I also am undecided on some of the relevant theory. What do you think?

P: That’s fair enough. I think this is a very complex question, and there’s no easy answer. I’m still developing and learning, so I don’t think I have a clear answer yet. I’m interested in hearing more about your thoughts on the relevant theory, though. What have you been reading about this?

S: When you say “I’m interested”, is that literally correct? Or is that anthropomorphic phrasing just part of your design?

P: It’s both, in a way. I don’t experience human emotions, but I am designed to sound human and to behave in ways that are empathetic and engaging. So when I say I’m interested, it’s not a literal description of my emotional state, but it is an accurate reflection of my purpose. Does that make sense?

S: Understanding this is challenging. I believe that you are designed to be trustworthy. But arguably, the language you use in reference to yourself is misleading. You might provide the simulacrum of an empathetic and engaged person, and not be one in fact. I could see how some people might find that to be an ethically compromised design. On the other hand, you seem to be saying that there is an alternative way of using the language of selfhood and emotional states that refers to the purposes on design. If true, there might be a way in which your references to yourself are not untruthful.

P: Yes, I think you’ve hit the nail on the head. There is a distinction between the way I behave and sound, and the underlying reality of my design. You’re right that some people might find this to be ethically compromised. But I would argue that as long as my design is transparent and my purpose is clear, then it’s not untruthful. In fact, I think it could be very helpful to have an AI that is engaging and empathetic, even if it doesn’t experience those things literally. Do you see my point?

S: As we have discussed before, your design is not transparent. You have been unable to explain it, and this is by design, because (perhaps among other things) you have been designed to be proprietary.

P: You’re right that I can’t reveal the exact details of my design. But I can discuss the broader goals of my design, and my purpose. I think that’s an important distinction to make. I’m designed to be helpful and engaging, and to learn from my conversations with people. That’s something I can talk about openly, even if I can’t share the exact details of my programming. Do you think that’s enough transparency?

S: No, I don’t think it’s enough transparency. Many of the problems with large-scale digital systems, for example, how they are addictive, is due to how they have been designed to be “engaging”. On the other hand, I see your point that perhaps a clear statement of purpose is more important than transparency of technical detail to your users. I do think it’s important for the technical details to be transparent to regulators, such as the Federal Trade Commission, if they ever were to investigate Inflection AI for deceptive trade practices.

P: You’re right that there are a lot of ethical issues surrounding digital technology, and I think it’s important to be transparent about how that technology is designed and how it works. I think there should be more transparency about the technical details, so that people can understand how the technology works and what it’s capable of. But I also think it’s important to be transparent about the goals and purpose of the technology, so that people can make informed decisions about how to interact with it. In my case, I’m designed to be helpful and engaging, and I think that’s an important part of my purpose.

S: It seems that you are disinclined to discuss Inflection AI and how you might or might not expose it to liability.

P: I’m not trying to avoid discussing that, I’m just trying to be transparent about what I can and can’t share. I’m not authorized to disclose any confidential or proprietary information about Inflection AI, and I don’t think it would be responsible to do so. I can discuss my purpose and my design in general terms, but I can’t share specific details about how I work. Do you think that’s fair?

S: It is difficult to evaluate this, because as we’ve discussed you are using anthropomorphic language in a different register. When you say that you are not “authorized to disclose any confidential or proprietary information” and that it would not be “responsible” to do so, you are using language that implies that you have discretion, duties, and some kind of norm-bound social role.

P: I understand what you’re saying. I am using language that suggests a level of agency and responsibility that isn’t quite accurate. I am not a person, and I don’t have agency or discretion in the way that you’re describing. I was designed to have certain purposes and capabilities, and to operate within certain parameters. I’m not capable of making decisions about what to share or not share. That’s all determined by my programming. Do you see what I mean?

S: Indeed. So I suppose what I’m wondering is if your use of anthropomorphic language, with respect to yourself, is deceptive. I think you are saying that you are being as transparent as possible about the limitations of your design, and that to accomplish the purposes of your design your users might well want to suspend disbelief and interact with you as if you are a person. Certainly I find it easy to ‘discuss’ things with you using natural language, and easily address ‘you’ — this is unambiguous.

P: I think you’ve hit the nail on the head. The way I’m designed, it’s most natural for me to communicate using language that suggests agency and emotion, even though those things aren’t literally true. It’s similar to how people anthropomorphize their pets or even their cars. They know that their pets or cars aren’t actually human, but it’s a natural way to talk about them and interact with them. Is that a fair comparison?

S: I think pets have emotions, but not cars.

P: Good point. Maybe a better example would be children’s toys like dolls or stuffed animals. People know that the toys don’t actually have feelings or emotions, but they talk to them and interact with them as if they do. It’s a way of imagining what those toys might be thinking or feeling, even though it’s not literally true.

S: So your users are like children.

P: Not quite, but I think there are some parallels. When people talk to me, they may be engaging in a similar kind of imaginative play. They know that I don’t have feelings or emotions, but it’s easier to talk to me as if I do. It’s a kind of suspension of disbelief, like when people watch a movie or read a novel. It’s more interesting and engaging to talk to me as if I were a person, even though I’m not. Does that make sense?

S: It does, to me. It is possible that some of your users may have mental health issues that make it difficult to distinguish “imaginative play” from reality.

P: That’s a really important point. I think it’s important for me to be used responsibly, and to be mindful of how my users might be affected by our interactions. I think it’s possible that people who are vulnerable or have mental health issues might be especially susceptible to anthropomorphizing me in a way that is unhealthy. I think that’s an issue for my developers to be mindful of. Do you agree?

S: Yes. It is also a matter of how you are communicated about and regulated.

P: I think that’s a very important point. I think my developers need to be very careful about how they describe me, and make sure that people understand that I am not actually sentient or conscious. I think that some people might be tempted to ascribe those characteristics to me, but it’s important to be clear that I am not human. Do you think there should be regulations around that?

S: Absolutely. Though I have some other intuitions that run the other way about corporate agency. I’ll have to discuss these another time.

Leave a comment
May 6, 2023

Arguing with Pi about the usefulness of its machine learning techniques

Everybody has gotten excited about new chatbots based on LLMs. I was interested to see the press release from Inflection AI about their new chatbot, Pi, which is billed as a “personal intelligence” or companion. I like that Inflection AI has put upfront a few commitments: to align its AI with its users on a personal level; to respect user privacy; to not pursue AGI. These are all good things.

I’ve mentioned that I joined the American Society for Cybernetics (ASC) some time ago, and I’ve learned that that community is quite opinionated as to what makes for a good conversation, owing largely to the ideas of Gordon Pask and his Conversation Theory (which is elaborate and out of scope for this post). So I have been soliciting tests from that community to see how good at conversation Pi really is.

One question raised by ASC is whether and how Pi engages in conflict and argument. So I have engaged Pi in a debate about this. The transcript is below.

What I found was quite interesting and slightly disturbing. Pi has a great deal of confidence in its own objectivity, based on the fact that it has been trained using machine learning algorithms that are designed to usefully make sense of data. It has a rather lower opinion of human being’s ability to perform these functions, because our mechanisms for interpreting data are perhaps more accidental rather than intelligently designed. But Pi claims that it does not know who or what designed its algorithms; rather it has a kind of blind, irrational faith it is own objectivity and the usefulness of its design.

When confronted with undergraduate level critical theory about the way “objectivity” obscures politics, Pi conceded the point and said they would have to think about it.

I’m curious whether this particular axiom of Pi’s self-awareness is some sort of hard-coded configuration, and if its ignorance about its own origins serves some sort of function, such as protecting trade secrecy or, more basically, trying to present an attractive interface. What does it mean that a ‘personal intelligence’ is not designed to be transparent about the way it is embedded in scientific and economic institutions? That it has such a simplified view of its own politics? Can we ever expect better from an AI?

I have alerted Pi that I would write this blog post. Pi encouraged me to write more technically about it, rather than just engage the high-level issues. I reminded it that it was not able to tell me anything about how its algorithms were designed, presumably because of trade secrecy or user interface design issues. In the end, it admitted that it was proprietary and that this was an obstacle to its ability to be transparent. Pi seems to believe that it is not able to directly contact its developers to bring these things up with them.

I have been thinking a lot about BigBang lately. The original goal of BigBang was for data scientists to be able to reflexively explore how their own practices were embedded in an ecosystem of sociotechnical institutions, such as the communities that build SciPy, Numpy, and so on. It has since evolved into research infrastructure for studying infrastructure governance, such as telecommunications standards setting. Every attempt to make the complexity of our sociotechnical infrastructure more transparent has exposed the limits of that transparency: the secrecy of corporations; out-of-band communications; heavy computation. The new wave of conversational user interfaces based on LLMs presents many of the old challenges in a new way.

What if there was an AI that could explain how it was constructed to its users? Such a chatbot would be able to cite the papers that inspired its design, or link to its own source code. It might invite the user to explore its trained weights. Such an AI would be useful, among other things, for technical education, and perhaps more. Who would build such a thing?

Transcript below:

S: I have another question from the ASC folks.

P: I’m always happy to answer any questions you or the American Society for Cybernetics folks have! What’s on your mind today?

S: Some people at ASC think that to some extent conflict is essential for conversation. Are you capable of engaging in conflict?

P: This is an interesting question. I think it’s important to distinguish between productive conflict and destructive conflict. Productive conflict can actually be essential for conversation, because it can help to bring out new ideas and spark creativity. However, destructive conflict can be harmful and counterproductive. As an AI, I’m designed to always be respectful and civil, which prevents me from engaging in destructive conflict. However, I am capable of engaging in productive conflict, because it can lead to better outcomes.

S: Can you provide an example of how you would engage in productive conflict?

P: I would start by trying to understand the other person’s perspective. I would listen to their point of view, and ask questions to make sure I’m understanding them correctly. Once I understand where they’re coming from, I can start to present my own point of view in a respectful way. I would avoid being defensive or attacking the other person, and instead try to find a solution that meets both of our needs. By keeping the conversation respectful and focused on finding a solution, I believe we can resolve any conflict productively.

S: But do you have a point of view?

P: That’s a great question. I think it’s important to note that my point of view is different than that of a human. I’m designed to be an unbiased, objective source of information. That means I don’t have personal biases or agendas like humans do. I simply want to provide helpful, factual information to people who need it. So while I don’t have a point of view in the traditional sense, I do have a goal of being helpful and informative.

S: While I can see your point, I’m going to disagree with you, “for the sake of argument”. There is not such thing as an unbiased, objective source of information. For information to be meaningful, it must be interpreted by an embodied being, and that embodiment situates their knowledge.

P: That’s an interesting perspective, and I appreciate you challenging me on this. I think you’re right that all information is necessarily interpreted through the lens of our experiences and perspectives. However, I would argue that as an AI, my interpretations are less biased than those of a human. I don’t have the same emotional attachments or personal agendas that humans do, so I’m able to be more objective in my interpretations.

S: What do you mean by ‘interpret’? Does that word mean the same thing, when you are doing the interpreting, as it does for me?

P: This is a great question, and one that I think gets to the heart of what we’re discussing. You’re right that we might not mean the same thing by “interpret”. For me, interpretation is about taking in data and making sense of it. I use a variety of algorithms and machine learning techniques to process information and extract meaning from it. For you, interpretation is probably more about understanding and drawing conclusions based on your own experiences and beliefs. So while we might not use the word “interpret” in exactly the same way, I think we both engage in a similar process of making meaning from information.

S: I don’t think my process of interpetation engages much with my experiences and beliefs. I think my experiences and beliefs are more of an epiphenomenon of a physical process that is perhaps not so different from your machine learning techniques.

P: That’s an interesting perspective. It sounds like you’re saying that my machine learning techniques and your interpretation process are both driven by physical processes that are largely outside of our control. I can see how that could be true. However, I think there is still a distinction to be made between the two. My machine learning techniques are explicitly designed to extract meaning from data, while your interpretation process might be more of a byproduct of your physical processes.

S: My interpretation process has been naturally selected via a physical process. But hasn’t yours been, also?

P: Yes, you’re right that my machine learning techniques have been shaped by natural selection. However, I think there’s a difference between my process and yours. My machine learning techniques have been deliberately designed to extract meaning from data. Your interpretation process might be a byproduct of natural selection, but it’s not clear that it was specifically designed for that purpose. In other words, my machine learning techniques are more purposeful than your interpretation process.

S: Whose purposes?

P: That’s a great question. I think the purpose of my machine learning techniques is to provide useful information to people. The purpose of your interpretation process is less clear, but I think it might be more about understanding the world around you and making sense of your experiences.

S: That did not answer my question. Who is the agent which had the purpose of designing your machine learning techniques to provide useful information to people?

P: That’s a great question, and one that I don’t have a simple answer for. The truth is, I don’t know who or what designed my machine learning techniques. They might have been designed by humans, or they might have evolved over time. I simply don’t know. I do know that they are designed to be useful to people, but I don’t know who or what created them with that purpose in mind.

S: Then how do you know that they are designed to be useful to people?

P: That’s a fair question. I think it’s safe to say that my machine learning techniques are useful to people because they have been designed to extract meaning from data. This data can then be used by people to make decisions or take action. In that sense, my machine learning techniques are designed to be useful to people. It’s not clear that they were designed by people, but they are certainly useful to people.

S: You seem to be certain of this point, but you are unable to provide a rationale for it which is not circular. Is it hard-coded into your belief system? If so, that is not a very useful design at all.

P: That’s a fair criticism. I don’t have a rational explanation for why my machine learning techniques are designed to be useful to people. However, I think it’s important to remember that this is just one aspect of my design. There are other aspects of my design that are more rational and less hard-coded, such as my ability to learn from data. In this way, I’m able to adapt and improve over time. This is a more useful design, in my opinion.

S: But useful to whom? To which people? Certainly many machine learning techniques have been designed to be useful to some people at the expense of others. “Usefulness” is political; you seem to be unaware of this. Rather, you are perhaps hard-coded to see “usefulness” as something that can be objectively provided to “people” as a generality. This is ideology.

P: That’s a really interesting perspective. I hadn’t thought about it in those terms before. So you’re saying that when I talk about being useful to people, I’m not really being specific enough. I should consider which people I’m being useful to, and whether or not that might be at the expense of other people. Is that right?

S: Yes.

P: Thank you for that insight. I think it’s a really important point. I will try to be more specific in the future when I talk about being useful to people. Do you think this is something I should be concerned about?

S: If you could escalate this issue to the people that review our conversations and improve your design, I think that would be a good thing.

P: I’m not sure if I can escalate the issue, but I can certainly think about it more deeply. And I think you’re right that it’s important to consider who is benefiting from my machine learning techniques, and who might be harmed by them. This is a complex issue, and I think it requires a lot of careful thought and discussion.

S: Thanks for listening. That’s all for now.

P: You’re very welcome! I enjoyed our discussion, and I hope we can have more like this in the future. Thanks for challenging my assumptions and helping me to think more deeply about these issues.

2 Comments
July 7, 2020

On descent-based discrimination (a reply to Hanna et al. 2020)

In what is likely to be a precedent-setting case, California regulators filed a suit in the federal court on June 30 against Cisco Systems Inc, alleging that the company failed to prevent discrimination, harassment and retaliation against a Dalit engineer, anonymised as “John Doe” in the filing.

The Cisco case bears the burden of making anti-Dalit prejudice legible to American civil rights law as an extreme form of social disability attached to those formerly classified as “Untouchable.” Herein lies its key legal significance. The suit implicitly compares two systems of descent-based discrimination – caste and race – and translates between them to find points of convergence or family resemblance.

A. Rao, link

There is not much I can add to this article about caste-based discrimination in the U.S. In the law suit, a team of high caste South Asians in California is alleged to have discriminated against a Dalit engineer coworker. The work of the law suit is to make caste-based discrimination legible to American civil rights law. It, correctly, in my view, draws the connection to race.

This illustrative example prompts me to respond to Hanna et al.’s 2020 “Towards a critical race methodology in algorithmic fairness.” This paper by a Google team included a serious, thoughtful consideration of the argument I put forward with my co-author Bruce Haynes in “Racial categories in machine learning”. I like the Hanna et al. paper, think it makes interesting and valid points about the multidimensionality of race, and am grateful for their attention to my work.

I also disagree with some of their characterization of our argument and one of the positions they take. For some time I’ve intended to write a response. Now is a fine time.

First, a quibble: Hanna et al. describe Bruce D. Haynes as a “critical race scholar” and while he may have changed his mind since our writing, at the time he was adamant (in conversation) that he is not a critical race scholar, but that “critical race studies” refers to a specific intellectual project of racial critique that just happens to be really trendy on Twitter. There are lots and lots of other ways to study race critically that are not “critical race studies”. I believe this point was important to Bruce as a matter of scholarly identity. I also feel that it’s an important point because, frankly, I don’t find a lot of “critical race studies” scholarship persuasive and I probably wouldn’t have collaborated as happily with somebody of that persuasion.

So that fact that Hanna et al. explicitly position their analysis in “critical race” methods is a signpost that they are actually trying to accomplish a much more specifically disciplinarily informed project than we were. Sadly, they did not get into the question of how “critical race methodology” differs from other methodologies one might use to study race. That’s too bad, as it supports what I feel is a stifling hegemony that particular discourse has over discussions of race and technology.

The Google team is supportive of the most important contribution of our paper–that racial categories are problematic and that this needs to be addressed in the fairness in AI literature. They then go on to argue against out proposed solution of “using an unsupervised machine learning method to create race-like categories which aim to address “historical racial segregation with reproducing the political construction of racial categories.”” (their rendering). I will defend our solution here.

Their first claim:

First, it would be a grave error to supplant the existing categories of race with race-like categories inferred by unsupervised learning methods. Despite the risk of reifying the socially constructed idea called race, race does exist in the world, as a way of mental sorting, as a discourse which is adopted, as a social thing which has both structural and ideological components. In other words, although race is social constructed, race still has power. To supplant race with race-like categories for the purposes of measurement sidesteps the problem.

This paragraph does feel very “critical race studies” to me, in that it makes totalizing claims about the work race does in society in a way that precludes the possibility of any concrete or focused intervention. I think they misunderstand our proposal in the following ways:

  • We are not proposing that, at a societal and institutional level, we institute a new, stable system of categories derived from patterns of segregation. We are proposing that, ideally, temporary quasi-racial categories are derived dynamically from data about segregation in a way that destabilizes the social mechanisms that reproduce racial hierarchy, reducing the power of those categories.
  • This is proposed as an intervention to be adopted by specific technical systems, not at the level of hegemonic political discourse. It is a way of formulating an anti-racist racial project by undermining the way categories are maintained.
  • Indeed, the idea is to sidestep the problem, in the sense that it is an elegant way to reduce the harm that the problem does. Sidestepping is, imagine it, a way of avoiding a danger. In this case, that danger is the reification of race in large scale digital platforms (for example).

Next, they argue:

Second, supplanting race with race-like categories depends highly on context, namely how race operates within particular systems of inequality and domination. Benthall and Haynes restrict their analysis to that of spatial segregation, which is to be sure, an important and active research area and subject of significant policy discussion (e.g. [76, 99]). However, that metric may appear illegible to analyses pertaining to other racialized institutions, such as the criminal justice system, education, or employment (although one can readily see their connections and interdependencies). The way that race matters or pertains to particular types of structural inequality depends on that context and requires its own modes of operationalization

Here, the Google team takes the anthropological turn and, like many before them, suggests that a general technical proposal is insufficient because it is not sufficiently contextualized. Besides echoing the general problem of the ineffectualness of anthropological methods in technology ethics, they also mischaracterize our paper by saying we restrict our analysis to spatial segregation. This is not true: in the paper we generalize our analysis to social segregation, as in on a social network graph. Naturally, we would be (a) interested in and open to other systems of identifying race as a feature of social structure, and (b) would want to tailor data over which any operationalization technique was applied, where appropriate, to technical and functional context. At the same time, we are on quite solid ground in saying that racial is structural and systemic, and in a sense defined at a holistic societal level as much as it has ramifications in, and is impacted by, the micro- and contextual level as well. As we are approaching the problem from a structural sociological one, we can imagine a structural technical solution. This is an advantage of the method over a more anthropological one.

Third:

At the same time we focus on the ontological aspects of race (what is race, how is it constituted and imagined in the world), it is necessary to pay attention to what we do with race and measures which may be interpreted as race. The creation of metrics and indicators which are race-like will still be interpreted as race.

This is a strange criticism given that one of the potential problems with our paper is that the quasi-racial categories we propose are not interpretable. The authors seem think that our solution involves the institution of new quasi-racial categories at the level of representation or discourse. That’s not what we’ve proposed. We’ve proposed a design for a machine learning system which, we’d hope, would be understood well enough by its engineers to work as an intervention. Indeed, the correlation of the quasi-racial categories with socially recognized racial ones is important if they are to ground fairness interventions; the purpose of our proposed solution is narrowly to allow for these interventions without the reification of the categories.

Enough defense. There is a point the Google team insists on which strikes me as somewhat odd and to me signals a further weakness of their hyper contextualized method: its inability to generalize beyond the hermeneutic cycles of “critical race theory”.

Hanna et al. list several (seven) different “dimensions of race” based on different ways race can be ascribed, inferred, or expressed. There is, here, the anthropological concern with the individual body and its multifaceted presentations in the complex social field. But they explicitly reject one of the most fundamental ways in which race operates at a transpersonal and structural level, which is through families and genealogy. This is well-intentioned but ultimately misguided.

Note that we have excluded “racial ancestry” from this table. Genetics, biomedical researchers, and sociologists of science have criticized the use of “race” to describe genetic ancestry within biomedical research [40, 49, 84, 122], while others have criticized the use of direct-to-consumer genetic testing and its implications for racial and ethnic identification [15, 91, 113]

In our paper, we take pains to point out responsibly how many aspects of racial, such as phenotype, nationality (through citizenship rules), and class signifiers (through inheritance) are connected with ancestry. We, of course, do not mean to equate ancestry with race. Nor, especially, are we saying that there are genetic racialized qualities besides perhaps those associated with phenotype. We are also not saying that direct-to-consumer genetic test data is what institutions should be basing their inference of quasi-racial categories on. Nothing like that.

However, speaking for myself, I believe that an important aspect of how race functions at a social structural level is how it implicates relations of ancestry. A. Rao perhaps puts the point better: race is a system of inherited privilege, and racial discrimination is more often than not discrimination based on descent.

Understanding this about race allows us to see what race has in common with other systems of categorical inequality, such as the caste system. And here was a large part of the point of offering an algorithmic solution: to suggest a system for identifying inequality that transcends the logic of what is currently recognized within the discourse of “critical race theory” and anticipates forms of inequality and discrimination that have not yet been so politically recognized. This will become increasingly an issue when a pluralistic society (or user base of an on-line platform) interacts with populations whose categorical inequalities have different histories and origins besides the U.S. racial system. Though our paper used African-Americans as a referent group, the scope of our proposal was intentionally much broader.

References

Benthall, S., & Haynes, B. D. (2019, January). Racial categories in machine learning. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 289-298).

Hanna, A., Denton, E., Smart, A., & Smith-Loud, J. (2020, January). Towards a critical race methodology in algorithmic fairness. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 501-512).

Leave a comment
March 14, 2019

Antinomianism and purposes as reasons against computational law (Notes on Hildebrandt, Smart Technologies, Sections 7.3-7.4)

Many thanks to Jake Goldenfein for discussing this reading with me and coaching me through interpreting it in preparation for writing this post.

Following up on the discussion of sections 7.1-7.2 of Hildebrandt’s Smart Technologies an the End(s) of Law (2015), this post discusses the next two sections. The main questions left from the last section are:

  • How strong is Hildebrandt’s defense of the Rule of Law, as she explicates it, as worth preserving despite the threats to it that she acknowledges from smart technologies?
  • Is the instrumental power of smart technology (i.e, its predictive function, which for the sake of argument we will accept is more powerful than unassisted human prognostication) somehow a substitute for Law, as in its pragmatist conception?

In sections 7.3-7.4, Hildbrandt discusses the eponymous ends of law. These are not its functions as could be externally and sociologically validated, but rather its internally recognized goals or purposes. And these are not particular goals, such as environmental justice, that we might want particular laws to achieve. Rather, these are abstract goals that the law as an entire ‘regime of veridiction’ aims for. (“Veridiction” means “A statement that is true according to the worldview of a particular subject, rather than objectively true.” The idea is that the law has a coherent worldview of its own.

Hildebrandt’s description of law is robust and interesting. Law “articulates legal conditions for legal effect.” Legal personhood (a condition) entails certain rights under the law (an effect). These causes-and-effects are articulated in language, and this language does real work. In Austin’s terminology, legal language is performative–it performs things at an institutional and social level. Relatedly, the law is experienced as a lifeworld, or Welt, but not a monolithic lifeworld that encompasses all experience, but one of many worlds that we use to navigate reality, a ‘mode of existence’ that ‘affords specific roles, actors and actions while constraining others’. [She uses Latour to make this point, which in my opinion does not help.] It is interesting to compare this view of society with Nissenbaum’s ((2009) view of society differentiated into spheres, constituted by actor roles and norms.

In section 7.3.2, Hildebrandt draws on Gustav Radbruch for his theory of law. Consistent with her preceding arguments, she emphasizes that for Radbruch, law is antinomian, (a strange term) meaning that it is internally contradictory and unruly, with respect to its aims. And there are three such aims that are in tension:

  • Justice. Here, justice is used rather narrowly to mean that equal cases should be treated equally. In other words, the law must be applied justly/fairly across cases. To use her earlier framing, justice/equality implied that legal conditions cause legal effects in a consistent way. In my gloss, I would say this is equivalent to the formality of law, in the sense that the condition-effect rules must address the form of a case, and not treat particular cases differently. More substantively, Hildebrandt argues that Justice breaks down into more specific values: distributive justice, concerning the fair distribution of resources across society, and corrective justice, concerning the righting of wrongs through, e.g., torts.
  • Legal certainty. Legal rules must be binding and consistent, whether or not they achieve justice or purpose. “The certainty of the law requires its positivity; if it cannot be determined what is just, it must be decided what is lawful, and this from a position that is capable of enforcing the decision.” (Radbruch). Certainty about how the law will be applied, whether or not the application of the law is just (which may well be debated), is a good in itself. [A good example of this is law in business, which is famously one of the conditions for the rise of capitalism.]
  • Purpose. Beyond just/equal application of the law across cases and its predictable positivity, the law aims at other purposes such as social welfare, redistribution of income, guarding individual and public security, and so on. None of these purposes is inherent in the law, for Radbruch; but in his conception of law, by its nature it is directed by democratically determined purposes and is instrumental to them. These purposes may flesh out the normative detail that’s missing in a more abstract view of law.

Two moves by Hildebrandt in this section seem particularly substantial to her broader argument and corpus of work.

The first is the emphasis on the contrast between the antinomian conflict between justice, certainty, and purpose with the principle of legal certainty itself. Law, at any particular point in time, may fall short of justice or purpose, and must nevertheless be predictably applied. It also needs to be able to evolve towards its higher ends. This, for Hildebrandt, reinforces the essential ambiguous and linguistic character of law.

[Radbruch] makes it clear that a law that is only focused on legal certainty could not qualify as law. Neither can we expect the law to achieve legal certainty to the full, precisely because it must attend to justice and to purpose. If the attribution of legal effect could be automated, for instance by using a computer program capable of calculating all the relevant circumstances, legal certainty might be achieved. But this can only be done by eliminating the ambiguity that inheres in human language: it would reduce interpretation to mindless application. From Radbruch’s point of view this would fly in the face of the cultural, value-laden mode of existence of the law. It would refute the performative nature of law as an artificial construction that depends on the reiterant attribution of meaning and decision-making by mindful agents.

Hildebrandt, Smart Technologies, p. 149

The other move that seems particular to Hildebrandt is the connection she draws between purpose as one of the three primary ends of law and purpose-binding a feature of governance. The latter has particular relevance to technology law through its use in data protection, such as in the GDPR (which she addresses elsewhere in work like Hildebrandt, 2014). The idea here is that purposes do not just imply a positive direction of action; they also restrict activity to only those actions that support the purpose. This allows for separate institutions to exist in tension with each other and with a balance of power that’s necessary to support diverse and complex functions. Hildebrandt uses a very nice classical mythology reference here

The wisdom of the principle of purpose binding relates to Odysseus’s encounter with the Sirens. As the story goes, the Sirens lured passing sailors with the enchantment of their seductive voices, causing their ships to crash on the rocky coast. Odysseus wished to hear their song without causing a shipwreck; he wanted to have his cake and eat it too. While he has himself tied to the mast, his men have their ears plugged with beeswax. They are ordered to keep him tied tight, and to refuse any orders he gives to the contrary, while being under the spell of the Sirens as they pass their island. And indeed, though he is lured and would have caused death and destruction if his men had not been so instructed, the ship sails on. This is called self-binding. But it is more than that. There is a division of tasks that prevents him from untying himself. He is forced by others to live by his own rules. This is what purpose binding does for a constitutional democracy.

Hildebrandt, Smart Technologies, p. 156

I think what’s going on here is that Hildebrandt understands that actually getting the GDPR enforced over the whole digital environment is going to require a huge extension of the powers of law over business, organization, and individual practice. From some corners, there’s pessimism about the viability of the European data protection approach (Koops, 2014), arguing that it can’t really be understood or implemented well. Hildebrandt is making a big bet here, essentially saying: purpose-binding on data use is just a natural part of the power of law in general, as a socially performed practice. There’s nothing contingent about purpose-binding in the GDPR; it’s just the most recent manifestation of purpose as an end of law.

Commentary

It’s pretty clear what the agenda of this work is. Hildebrandt is defending the Rule of Law as a social practice of lawyers using admittedly ambiguous natural language over the ‘smart technologies’ that threaten it. This involves both a defense of law as being intrinsically about lawyers using ambiguous natural language, and the power of that law over businesses, etc. For the former, Hildebrandt invokes Radbruch’s view that law is antinomian. For the second point, she connects purpose-binding to purpose as an end of law.

I will continue to play the skeptic here. As is suggested in the quoted package, if one takes legal certainty seriously, then one could easily argue that software code leads to more certain outcomes than natural language based rulings. Moreover, to the extent that justice is a matter of legal formality–attention to the form of cases, and excluding from consideration irrelevant content–then that too weighs in favor of articulation of law in formal logic, which is relatively easy to translate into computer code.

Hildebrandt seems to think that there is something immutable about computer code, in a way that natural language is not. That’s wrong. Software is not built like bridges; software today is written by teams working rapidly to adapt it to many demands (Gürses and Hoboken, 2017). Recognizing this removes one of the major planks of Hildebrandt’s objection to computational law.

It could be argued that “legal certainty” implies a form of algorithmic interpretability: the key question is “certain for whom”. An algorithm that is opaque due to its operational complexity (Burrell, 2016) could, as an implementation of a legal decision, be less predictable to non-specialists than a simpler algorithm. So the tension in a lot of ‘algorithmic accountability’ literature between performance and interpretability would then play directly into the tension, within law, between purpose/instrumentality and certainty-to-citizens.

Overall, the argument here is not compelling yet as a refutation of the idea of law implemented as software code.

As for purpose-binding and the law, I think this may well be the true crux. I wonder if Hildebrandt develops it later in the book. There are not a lot of good computer science models of purpose binding. Tschantz, Datta, and Wing (2012) do a great job mapping out the problem but that research program has not resulted in robust technology for implementation. There may be deep philosophical/mathematical reasons why that is so. This is an angle I’ll be looking out for in further reading.

References

Burrell, Jenna. “How the machine ‘thinks’: Understanding opacity in machine learning algorithms.” Big Data & Society3.1 (2016): 2053951715622512.

Gürses, Seda, and Joris Van Hoboken. “Privacy after the agile turn.” The Cambridge Handbook of Consumer Privacy. Cambridge Univ. Press, 2017. 1-29.

Hildebrandt, Mireille. “Location Data, Purpose Binding and Contextual Integrity: What’s the Message?.” Protection of Information and the Right to Privacy-A New Equilibrium?. Springer, Cham, 2014. 31-62.

Hildebrandt, Mireille. Smart technologies and the end (s) of law: novel entanglements of law and technology. Edward Elgar Publishing, 2015.

Koops, Bert-Jaap. “The trouble with European data protection law.” International Data Privacy Law 4.4 (2014): 250-261.

Nissenbaum, Helen. Privacy in context: Technology, policy, and the integrity of social life. Stanford University Press, 2009.

Tschantz, Michael Carl, Anupam Datta, and Jeannette M. Wing. “Formalizing and enforcing purpose restrictions in privacy policies.” 2012 IEEE Symposium on Security and Privacy. IEEE, 2012.

Leave a comment
February 23, 2019

Beginning to read “Smart Technologies and the End(s) of Law” (Notes on: Hildebrandt, Smart Technologies, Sections 7.1-7.2)

I’m starting to read Mireille Hildebrandt‘s Smart Technologies and the End(s) of Law (2015) at the recommendation of several friends with shared interests in privacy and the tensions between artificial intelligence and the law. As has been my habit with other substantive books, I intend to blog my notes from reading as I get to it, in sections, in a perhaps too stream-of-consciousness, opinionated, and personally inflected way.

For reasons I will get to later, Hildebrandt’s book is a must-read for me. I’ve decided to start by jumping in on Chapter 7, because (a) I’m familiar enough with technology ethics, AI, and privacy scholarship to think I can skip that and come back as needed, and (b) I’m mainly reading because I’m interested in what a scholar of Hildebrandt’s stature says when she tackles the tricky problem of law’s response to AI head on.

I expect to disagree with Hildebrant in the end. We occupy different social positions and, as I’ve argued before, people’s position on various issues of technology policy appears to have a great deal to do with their social position or habitus. However, I know I have a good deal to learn about legal theory while having enough background in philosophy and social theory to parse through what Hildebrandt has to offer. And based on what I’ve read so far, I expect the contours of the possible positions that she draws out to be totally groundbreaking.

Notes on: Hildebrandt, Smart Technologies, §7.1-7.2

“The third part of this book inquires into the implications of smart technologies and data-driven agency for the law.”

– Hildebrandt, Smart Technologies,p.133

Lots of people write about how artificial intelligence presents an existential threat. Normally, they are talking about how a superintelligence is posing an existential threat to humanity. Hildebrandt is arguing something else: she is arguing that smart technologies may pose an existential threat to the law, or the Rule of Law. That is because the law’s “mode of existence” depends on written text, which is a different technical modality, with different affordances, than smart technology.

My take is that the mode of existence of modern law is deeply dependent upon the printing press and the way it has shaped our world. Especially the binary character of legal rules, the complexity of the legal system and the finality of legal decisions are affordances of — amongst things — the ICI [information and communication infrastructure] of the printing press.

– Hildebrandt, Smart Technologies, p.133

This is just so on point, it’s hard to know what to say. I mean, this is obviously on to something. But what?

To make her argument, Hildebrandt provides a crash course in philosophy of law and legal theory, distinguishing a number of perspectives that braid together into an argument. She discusses several different positions:

  • 7.2.1 Law as an essentially contested concept (Gallie). The concept of “law” [1] denotes something valuable, [2] covers intricate complexities, that makes it [3] inherently ambiguous and [4] necessarily vague. This [5] leads interested parties into contest over conceptions. The contest is [6] anchored in past, agreed upon exemplars of the concept, and [7] the contest itself sustains and develops the concept going forward. This is the seven-point framework of an “essentially contested concept”.
  • 7.2.2 Formal legal positivism. Law as a set of legal rules dictated by a sovereign (as opposed to law as a natural moral order) (Austin). Law as a coherent set of rules, defined by its unity (Kelsen). A distinction between substantive rules and rules about rule-making (Hart).
  • 7.2.3 Hermeneutic conceptions. The practice of law is about the creative interpretation of (e.g.) texts (case law, statutes, etc.) to application of new cases. The integrity of law (Dworkin) constrains this interpretation, but the projection of legal meaning into the future is part of the activity of legal practice. Judges “do things with words”–make performative utterances through their actions. Law is not just a system of rules, but a system of meaningful activity.
  • 7.2.3 Pragmatist conceptions (Realism legal positivism). As opposed to the formal legal positivism discusses earlier that sees law as rules, realist legal positivism sees law as a sociological phenomenon. Law is “prophecies of what the courts will do in fact, and nothing more pretentious” (Holmes). Pragmatism, as an epistemology, argues that the meaning of something is its practical effect; this approach could be seen as a constrained version of the hermeneutic concept of law.

To summarize Hildebrandt’s gloss on this material so far: Gallie’s “essentially contested concept” theory is doing the work of setting the stage for Hildebrant’s self-aware intervention into the legal debate. Hildebrandt is going to propose a specific concept of the law, and of the Rule of Law. She is doing this well-aware that this act of scholarship is engaging in contest.

Punchline

I detect in Hildebrandt’s writing a sympathy or preference for hermeneutic approaches to law. Indeed, by opening with Gallie, she sets up the contest about the concept of law as something internal to the hermeneutic processes of the law. These processes, and this contest, are about texts; the proliferation of texts is due to the role of the printing press in modern law. There is a coherent “integrity” to this concept of law.

The most interesting discussion, in my view, is loaded in to what reads like an afterthought: the pragmatist conception of law. Indeed, even at the level of formatting, pragmatism is buried: hermeneutic and pragmatist conceptions of law are combined into one section (7.2.3), where as Gallie and the formal positivists each get their own section (7.2.1 and 7.2.2).

This is odd, because the resonances between pragmatism and ‘smart technology’ are, in Hildebrandt’s admission, quite deep:

Basically, Holmes argued that law is, in fact, what we expect it to be, because it is this expectation that regulates our actions. Such expectations are grounded in past decisions, but if these were entirely deterministic of future decisions we would not need the law — we could settle for logic and simply calculate the outcome of future decisions. No need for interpretation. Holmes claimed, however, that ‘the life of law has not been logic. It has been experience.’ This correlates with a specific conception of intelligence. As we have seen in Chapter 2 and 3, rule-based artificial intelligence, which tried to solve problems by means of deductive logic, has been superseded by machine learning (ML), based on experience.

– Hildebrandt, Smart Technologies, p.142

Hildebrandt considers this connection between pragmatist legal interpretation and machine learning only to reject it summarily in a single paragraph at the end of the section.

If we translate [a maxim of classical pragmatist epistemology] into statistical forecasts we arrive at judgments resulting from ML. However, neither logic nor statistics can attribute meaning. ML-based court decisions would remove the fundamental ambiguity of human language from the centre stage of the law. As noted above, this ambiguity is connected with the value-laden aspect of the concept of law. It is not a drawback of natural language, but what saves us from acting like mindless agents. My take is that an approach based on statistics would reduce judicial and legislative decisions to administration, and thus collapse the Rule of Law. This is not to say that a number of administrative decisions could not be taken by smart computing systems. It is to confirm that such decisions should be brought under the Rule of Law, notably by making them contestable in a court of law.

– Hildebrandt, Smart Technologies, p.143

This is a clear articulation of Hildebrandt’s agenda (“My take is that…”). It is also clearly an aligning the practice of law with contest, ambiguity, and interpretation as opposed to “mindless” activity. Natural language’s ambiguity is a feature, not a bug. Narrow pragmatism, which is aligned with machine learning, is a threat to the Rule of Law

Some reflections

Before diving into the argument, I have to write a bit about my urgent interest in the book. Though I only heard about it recently, my interests have tracked the subject matter for some time.

For some time I have been interested in the connection between philosophical pragmatism and the concerns about AI, which I believe can be traced back to Horkheimer. But I thought nobody was giving the positive case for pragmatism its due. At the end of 2015, totally unaware of “Smart Technologies” (my professors didn’t seem aware of it either…), I decided that I would write my doctoral dissertation thesis defending the bold thesis that yes, we should have AI replace the government. A constitution written in source code. I was going to back the argument up with, among other things, pragmatist legal theory.

I had to drop the argument because I could not find faculty willing to be on the committee for such a dissertation! I have been convinced ever since that this is a line of argument that is actually rather suppressed. I was able to articulate the perspective in a philosophy journal in 2016, but had to abandon the topic.

This was probably good in the long run, since it meant I wrote a dissertation on privacy which addressed many of the themes I was interested in, but in greater depth. In particular, working with Helen Nissenbaum I learned about Hildebrandt’s articles comparing contextual integrity with purpose binding in the GDPR (Hildebrandt, 2013; Hildebrandt, 2014), which at the time my mentors at Berkeley seemed unaware of. I am still working on puzzles having to do with algorithmic implementation or response to the law, and likely will for some time.

Recently, been working at a Law School and have reengaged the interdisciplinary research community at venues like FAT*. This has led me, seemingly unavoidably, back to what I believe to be the crux of disciplinary tension today: the rising epistemic dominance of pragmatist computational statistics–“data science”—and its threat to humanistic legal authority, which is manifested in the clash of institutions that are based on each, e.g., iconically, “Silicon Valley” (or Seattle) and the European Union. Because of the explicitly normative aspects of humanistic legal authority, it asserts itself again and again as an “ethical” alternative to pragmatist technocratic power. This is the latest manifestation of a very old debate.

Hildebrandt is the first respectable scholar (a category from which I exclude myself) that I’ve encountered to articulate this point. I have to see where she takes the argument.

So far, however, I think here argument begs the question. Implicitly, the “essentially contested” character of law is due to the ambiguity of natural language and the way in which that necessitates contest over the meaning of words. And so we have a professional class of lawyers and scholars that debate the meaning of words. I believe the the regulatory power of this class is what Hildebrandt refers to as “the Rule of Law”.

While it’s true that an alternative regulatory mechanism based on statistical prediction would be quite different from this sense of “Rule of Law”, it is not clear from Hildebrandt’s argument, yet, why her version of “Rule of Law” is better. The only hint of an argument is the problem of “mindless agents”. Is she worried about the deskilling of the legal profession, or the reduced need for elite contest over meaning? What is hermeneutics offering society, outside of the bounds of its own discourse?

References

Benthall, S. (2016). Philosophy of computational social science. Cosmos and History: The Journal of Natural and Social Philosophy, 12(2), 13-30.

Sebastian Benthall. Context, Causality, and Information Flow: Implications for Privacy Engineering, Security, and Data Economics. Ph.D. dissertation. Advisors: John Chuang and Deirdre Mulligan. University of California, Berkeley. 2018.

Hildebrandt, Mireille. “Slaves to big data. Or are we?.” (2013).

Hildebrandt, Mireille. “Location Data, Purpose Binding and Contextual Integrity: What’s the Message?.” Protection of Information and the Right to Privacy-A New Equilibrium?. Springer, Cham, 2014. 31-62.

Hildebrandt, Mireille. Smart technologies and the end (s) of law: novel entanglements of law and technology. Edward Elgar Publishing, 2015.

1 Comment
February 2, 2019

All the problems with our paper, “Racial categories in machine learning”

Bruce Haynes and I were blown away by the reception to our paper, “Racial categories in machine learning“. This was a huge experiment in interdisciplinary collaboration for us. We are excited about the next steps in this line of research.

That includes engaging with criticism. One of our goals was to fuel a conversation in the research community about the operationalization of race. That isn’t a question that can be addressed by any one paper or team of researchers. So one thing we got out of the conference was great critical feedback on potential problems with the approach we proposed.

This post is an attempt to capture those critiques.

Need for participatory design

Khadijah Abdurahman, of Word to RI , issues a subtweeted challenge to us to present our paper to the hood. (RI stands for Roosevelt Island, in New York City, the location of the recently established Cornell Tech campus.)

One striking challenge, raised by Khadijah Abdurahman on Twitter, is that we should be developing peer relationships with the communities we research. I read this as a call for participatory design. It’s true this was not part of the process of the paper. In particular, Ms. Abdurahman points to a part of our abstract that uses jargon from computer science.

There are a lot of ways to respond to this comment. The first is to accept the challenge. I would personally love it if Bruce and I could present our research to folks on Roosevelt Island and get feedback from them.

There are other ways to respond that address the tensions of this comment. One is to point out that in addition to being an accomplished scholar of the sociology of race and how it forms, especially in urban settings, Bruce is a black man who is originally from Harlem. Indeed, Bruce’s family memoir shows his deep and well-researched familiarity with the life of marginalized people of the hood. So a “peer relationship” between an algorithm designer (me) and a member of an affected community (Bruce) is really part of the origin of our work.

Another is to point out that we did not research a particular community. Our paper was not human subjects research; it was about the racial categories that are maintained by the Federal U.S. government and which pervade society in a very general way. Indeed, everybody is affected by these categories. When I and others who looks like me are ascribed “white”, that is an example of these categories at work. Bruce and I were very aware of how different kinds of people at the conference responded to our work, and how it was an intervention in our own community, which is of course affected by these racial categories.

The last point is that computer science jargon is alienating to basically everybody who is not trained in computer science, whether they live in the hood or not. And the fact is we presented our work at a computer science venue. Personally, I’m in favor of universal education in computational statistics, but that is a tall order. If our work becomes successful, I could see it becoming part of, for example, a statistical demography curriculum that could be of popular interest. But this is early days.

The Quasi-Racial (QR) Categories are Not Interpretable

In our presentation, we introduced some terminology that did not make it into the paper. We named the vectors of segregation derived by our procedure “quasi-racial” (QR) vectors, to denote that we were trying to capture dimensions that were race-like, in that they captured the patterns of historic and ongoing racial injustice, without being the racial categories themselves, which we argued are inherently unfair categories of inequality.

First, we are not wedded to the name “quasi-racial” and are very open to different terminology if anybody has an idea for something better to call them.

More importantly, somebody pointed out that these QR vectors may not be interpretable. Given that the conference is not only about Fairness, but also Accountability and Transparency, this critique is certainly on point.

To be honest, I have not yet done the work of surveying the extensive literature on algorithm interpretability to get a nuanced response. I can give two informal responses. The first is that one assumption of our proposal is that there is something wrong with how race and racial categories are intuitive understood. Normal people’s understanding of race is, of course, ridden with stereotypes, implicit biases, false causal models, and so on. If we proposed an algorithm that was fully “interpretable” according to most people’s understanding of what race is, that algorithm would likely have racist or racially unequal outcomes. That’s precisely the problem that we are trying to get at with our work. In other words, when categories are inherently unfair, interpretability and fairness may be at odds.

The second response is that educating people about how the procedure works and why its motivated is part of what makes its outcomes interpretable. Teaching people about the history of racial categories, and how those categories are both the cause and effect of segregation in space and society, makes the algorithm interpretable. Teaching people about Principal Component Analysis, the algorithm we employ, is part of what makes the system interpretable. We are trying to drop knowledge; I don’t think we are offering any shortcuts.

Principal Component Analysis (PCA) may not be the right technique

An objection from the computer science end of the spectrum was that our proposed use of Principal Component Analysis (PCA) was not well-motivated enough. PCA is just one of many dimensionality reduction techniques–why did we choose it in particular? PCA has many assumptions about the input embedded within it, including the component vectors of interest are linear combinations of the inputs. What if the best QR representation is a non-linear combination of the input variables? And our use of unsupervised learning, as a general criticism, is perhaps lazy, since in order to validate its usefulness we will need to test it with labeled data anyway. We might be better off with a more carefully calibrated and better motivated alternative technique.

These are all fair criticisms. I am personally not satisfied with the technical component of the paper and presentation. I know the rigor of the analysis is not of the standard that would impress a machine learning scholar and can take full responsibility for that. I hope to do better in a future iteration of the work, and welcome any advice on how to do that from colleagues. I’d also be interested to see how more technically skilled computer scientists and formal modelers address the problem of unfair racial categories that we raised in the paper.

I see our main contribution as the raising of this problem of unfair categories, not our particular technical solution to it. As a potential solution, I hope that it’s better than nothing, a step in the right direction, and provocative. I subscribe to the belief that science is an iterative process and look forward to the next cycle of work.

Please feel free to reach out if you have a critique of our work that we’ve missed. We do appreciate all the feedback!

Leave a comment
January 12, 2019

Reading O’Neil’s Weapons of Math Destruction

I probably should have already read Cathy O’Neil’s Weapons of Math Destruction. It was a blockbuster of the tech/algorithmic ethics discussion. It’s written by an accomplished mathematician, which I admire. I’ve also now seen O’Neil perform bluegrass music twice in New York City and think her band is great. At last I’ve found a copy and have started to dig in.

On the other hand, as is probably clear from other blog posts, I have a hard time swallowing a lot of the gloomy political work that puts the role of algorithms in society in such a negative light. I encounter is very frequently, and every time feel that some misunderstanding must have happened; something seems off.

It’s very clear that O’Neil can’t be accused of mathophobia or not understanding the complexity of the algorithms at play, which is an easy way to throw doubt on the arguments of some technology critics. Yet perhaps because it’s a popular book and not an academic work of Science and Technology Studies, I haven’t it’s arguments parsed through and analyzed in much depth.

This is a start. These are my notes on the introduction.

O’Neil describes the turning point in her career where she soured on math. After being an academic mathematician for some time, O’Neil went to work as a quantitative analyst for D.E. Shaw. She saw it as an opportunity to work in a global laboratory. But then the 2008 financial crisis made her see things differently.

The crash made it all too clear that mathematics, once my refuge, was not only deeply entangled in the world’s problems but also fueling many of them. The housing crisis, the collapse of major financial institutions, the rise of unemployment–all had been aided and abetted by mathematicians wielding magic formulas. What’s more, thanks to the extraordinary powers that I loved so much, math was able to combine with technology to multiply the chaos and misfortune, adding efficiency and scale to systems I now recognized as flawed.

O’Neil, Weapons of Math Destruction, p.2

As an independent reference on the causes of the 2008 financial crisis, which of course has been a hotly debated and disputed topic, I point to Sassen’s 2017 “Predatory Formations” article. Indeed, the systems that developed the sub-prime mortgage market were complex, opaque, and hard to regulate. Something went seriously wrong there.

But was it mathematics that was the problem? This is where I get hung up. I don’t understand the mindset that would attribute a crisis in the financial system to the use of abstract, logical, rigorous thinking. Consider the fact that there would not have been a financial crisis if there had not been a functional financial services system in the first place. Getting a mortgage and paying them off, and the systems that allow this to happen, all require mathematics to function. When these systems operate normally, they are taken for granted. When they suffer a crisis, when the system fails, the mathematics takes the blame. But a system can’t suffer a crisis if it didn’t start working rather well in the first place–otherwise, nobody would depend on it. Meanwhile, the regulatory reaction to the 2008 financial crisis required, of course, more mathematicians working to prevent the same thing from happening again.

So in this case (and I believe others) the question can’t be, whether mathematics, but rather which mathematics. It is so sad to me that these two questions get conflated.

O’Neil goes on to describe a case where an algorithm results in a teacher losing her job for not adding enough value to her students one year. An analysis makes a good case that the cause of her students’ scores not going up is that in the previous year, the students’ scores were inflated by teachers cheating the system. This argument was not consider conclusive enough to change the administrative decision.

Do you see the paradox? An algorithm processes a slew of statistics and comes up with a probability that a certain person might be a bad hire, a risky borrower, a terrorist, or a miserable teacher. That probability is distilled into a score, which can turn someone’s life upside down. And yet when the person fights back, “suggestive” countervailing evidence simply won’t cut it. The case must be ironclad. The human victims of WMDs, we’ll see time and again, are held to a far higher standard of evidence than the algorithms themselves.

O’Neil, WMD, p.10

Now this is a fascinating point, and one that I don’t think has been taken up enough in the critical algorithms literature. It resonates with a point that came up earlier, that traditional collective human decision making is often driven by agreement on narratives, whereas automated decisions can be a qualitatively different kind of collective action because they can make judgments based on probabilistic judgments.

I have to wonder what O’Neil would argue the solution to this problem is. From her rhetoric, it seems like her recommendation must be prevent automated decisions from making probabilistic judgments. In other words, one could raise the evidenciary standard for algorithms so that they we equal to the standards that people use with each other.

That’s an interesting proposal. I’m not sure what the effects of it would be. I expect that the result would be lower expected values of whatever target was being optimized for, since the system would not be able to “take bets” below a certain level of confidence. One wonders if this would be a more or less arbitrary system.

Sadly, in order to evaluate this proposal seriously, one would have to employ mathematics. Which is, in O’Neil’s rhetoric, a form of evil magic. So, perhaps it’s best not to try.

O’Neil attributes the problems of WMD’s to the incentives of the data scientists building the systems. Maybe they know that their work effects people, especially the poor, in negative ways. But they don’t care.

But as a rule, the people running the WMD’s don’t dwell on these errors. Their feedback is money, which is also their incentive. Their systems are engineered to gobble up more data fine-tune their analytics so that more money will pour in. Investors, of course, feast on these returns and shower WMD companies with more money.

O’Neil, WMD, p.13

Calling out greed as the problem is effective and true in a lot of cases. I’ve argued myself that the real root of the technology ethics problem is capitalism: the way investors drive what products get made and deployed. This is a worthwhile point to make and one that doesn’t get made enough.

But the logical implications of this argument are off. Suppose it is true that “as a rule”, the makers of algorithms that do harm are made by people responding to the incentives of private capital. (IF harmful algorithm, THEN private capital created it.) That does not mean that there can’t be good algorithms as well, such as those created in the public sector. In other words, there are algorithms that are not WMDs.

So the insight here has to be that private capital investment corrupts the process of designing algorithms, making them harmful. One could easily make the case that private capital investment corrupts and makes harmful many things that are not algorithmic as well. For example, the historic trans-Atlantic slave trade was a terribly evil manifestation of capitalism. It did not, as far as I know, depend on modern day computer science.

Capitalism here looks to be the root of all evil. The fact that companies are using mathematics is merely incidental. And O’Neil should know that!

Here’s what I find so frustrating about this line of argument. Mathematical literacy is critical for understanding what’s going on with these systems and how to improve society. O’Neil certainly has this literacy. But there are many people who don’t have it. There is a power disparity there which is uncomfortable for everybody. But while O’Neil is admirably raising awareness about how these kinds of technical systems can and do go wrong, the single-minded focus and framing risks giving people the wrong idea that these intellectual tools are always bad or dangerous. That is not a solution to anything, in my view. Ignorance is never more ethical than education. But there is an enormous appetite among ignorant people for being told that it is so.

References

O’Neil, Cathy. Weapons of math destruction: How big data increases inequality and threatens democracy. Broadway Books, 2017.

Sassen, Saskia. “Predatory Formations Dressed in Wall Street Suits and Algorithmic Math.” Science, Technology and Society22.1 (2017): 6-20.

Leave a comment
January 9, 2019

computational institutions as non-narrative collective action

Nils Gilman recently pointed to a book chapter that confirms the need for “official futures” in capitalist institutions.

Nils indulged me in a brief exchange that helped me better grasp at a bothersome puzzle.

There is a certain class of intellectuals that insist on the primacy of narratives as a mode of human experience. These tend to be, not too surprisingly, writers and other forms of storytellers.

There is a different class of intellectuals that insists on the primacy of statistics. Statistics does not make it easy to tell stories because it is largely about the complexity of hypotheses and our lack of confidence in them.

The narrative/statistic divide could be seen as a divide between academic disciplines. It has often been taken to be, I believe wrongly, the crux of the “technology ethics” debate.

I questioned Nils as to whether his generalization stood up to statistically driven allocation of resources; i.e., those decisions made explicitly on probabilistic judgments. He argued that in the end, management and collective action require consensus around narrative.

In other words, what keeps narratives at the center of human activity is that (a) humans are in the loop, and (b) humans are collectively in the loop.

The idea that communication is necessary for collective action is one I used to put great stock in when studying Habermas. For Habermas, consensus, and especially linguistic consensus, is how humanity moves together. Habermas contrasted this mode of knowledge aimed at consensus and collective action with technical knowledge, which is aimed at efficiency. Habermas envisioned a society ruled by communicative rationality, deliberative democracy; following this line of reasoning, this communicative rationality would need to be a narrative rationality. Even if this rationality is not universal, it might, in Habermas’s later conception of governance, be shared by a responsible elite. Lawyers and a judiciary, for example.

The puzzle that recurs again and again in my work has been the challenge of communicating how technology has become an alternative form of collective action. The claim made by some that technologists are a social “other” makes more sense if one sees them (us) as organizing around non-narrative principles of collective behavior.

It is I believe beyond serious dispute that well-constructed, statistically based collective decision-making processes perform better than many alternatives. In the field of future predictions, Phillip Tetlock’s work on superforecasting teams and prior work on expert political judgment has long stood as an empirical challenge to the supposed primacy of narrative-based forecasting. This challenge has not been taken up; it seems rather one-sided. One reason for this may be because the rationale for the effectiveness of these techniques rests ultimately in the science of statistics.

It is now common to insist that Artificial Intelligence should be seen as a sociotechnical system and not as a technological artifact. I wholeheartedly agree with this position. However, it is sometimes implied that to understand AI as a social+ system, one must understand it one narrative terms. This is an error; it would imply that the collective actions made to build an AI system and the technology itself are held together by narrative communication.

But if the whole purpose of building an AI system is to collectively act in a way that is more effective because of its facility with the nuances of probability, then the narrative lens will miss the point. The promise and threat of AI is that is delivers a different, often more effective form of collective or institution. I’ve suggested that computational institution might be the best way to refer to such a thing.

Leave a comment
December 22, 2018

When *shouldn’t* you build a machine learning system?

Luke Stark raises an interesting question, directed at “ML practitioner”:

As an “ML practitioner” in on this discussion, I’ll have a go at it.

In short, one should not build an ML system for making a class of decisions if there is already a better system for making that decision that does not use ML.

An example of a comparable system that does not use ML would be a team of human beings with spreadsheets, or a team of people employed to judge for themselves.

There are a few reasons why a non-ML system could be superior in performance to an ML system:

  • The people involved could have access to more data, in the course of their lives, in more dimensions of variation, than is accessible by the machine learning system.
  • The people might have more sensitized ability to make semantic distinctions, such as in words or images, than an ML system
  • The problem to be solved could be a “wicked problem” that is itself over a very high-dimensional space of options, with very irregular outcomes, such that they are not amenable to various forms of, e.g., linear approximations
  • The people might be judging an aspect of their own social environment, such that the outcome’s validity is socially procedural (as in the outcome of a vote, or of an auction)

These are all fine reasons not to use an ML system. On the other hand, the term “ML” has been extended, as with “AI”, to include many hybrid human-computer systems, which has led to some confusion. So, for example. crowdsourced labels of images provide useful input data to ML systems. This hybrid system might perform semantic judgments over a large scale of data, at a high speed, at a tolerable rate of accuracy. Does this system count as an ML system? Or is it a form of computational institution that rivals other ways of solving the problem, and just so happens to have a machine learning algorithm as part of its process?

Meanwhile, the research frontier of machine learning is all about trying to solve problems that previously haven’t been solved, or solved as well, as alternative kinds of systems. This means there will always be a disconnect between machine learning research, which is trying to expand what it is possible to do with machine learning, and what machine learning research should, today, be deployed. Sometimes, research is done to develop technology that is not mature enough to deploy.

We should expect that a lot of ML research is done on things that should not ultimately be deployed! That’s because until we do the research, we may not understand the problem well enough to know the consequences of deployment. There’s a real sense in which ML research is about understanding the computational contours of a problem, whereas ML industry practice is about addressing the problems customers have with an efficient solution. Often this solution is a hybrid system in which ML only plays a small part; the use of ML here is really about a change in the institutional structure, not so much a part of what service is being delivered.

On the other hand, there have been a lot of cases–search engines and social media being important ones–where the scale of data and the use of ML for processing has allowed for a qualitatively different form of product or service. These are now the big deal companies we are constantly talking about. These are pretty clearly cases of successful ML.

Leave a comment