Philosophy of Computer Science: Online Resources

Further Readings for Chapter 17:

Computer Ethics I:
Should We Trust Computers?

Note 1: Many of these items are online; links are given where they are known. Other items may also be online; an internet search should help you find them.

Note 2: In general, works are listed in chronological order. (This makes it easier to follow the historical development of ideas.)

Part V: Further Reading on Computer Ethics:

• For good introductions to computer ethics in general, see:
  • Johnson, D.G. and Snapper, J.W., editors (1985). Ethical Issues in the Use of Computers. Wadsworth
    • Reviewed in:
      Rapaport, W.J. (1986). Review of Johnson and Snapper 1985. Teaching Philosophy, 9(3):275–278

  • Moor, J.H. (1985). What is computer ethics? Metaphilosophy, 16(4):266–275.

  • Johnson, D.G. (2001). Computer Ethics. Prentice Hall, Upper Saddle River, NJ.

  • Anderson 2006

  • Müller, Vincent C., Ethics of Artificial Intelligence and Robotics, The Stanford Encyclopedia of Philosophy (Summer 2021 Edition), Edward N. Zalta (ed.)

  • Pham, K. and Rudra, A. (2021). Teaching Responsible Computing Playbook

  • Awad, Edmond et al. (2022), "Computational Ethics", Trends in Cognitive Sciences 26(5):388–404.

• AAAI's "AI Topics" website is an excellent source, with many links.

• See also The Research Center on Values in Engineering, Science, and Technology's pages on computer ethics.

• Is ethics computable? For an argument that it is not, see:
  • Johnson, Deborah G. & Verdicchio, Mario (2023), "Ethical AI Is Not about AI", Communications of the ACM 66(2) (February): 32–34.
    • "The concepts at the heart of ethics are not fixed or determinate in their precise meaning. To be applied they must be interpreted, and interpretations vary among individuals and groups, from context to context, and may change over time. … Because ethical concepts require social interpretations, they are subject to disagreement, contestation, and change. No computational model can capture all the possible interpretations that can constitute the social meaning of an ethical concept. As such, ethical concepts are not conducive to computational expression." (p. 33, my italics)

§17.1: Introduction

On automated vehicles, see:

• Halpern, S. (2016). Our driverless future. New York Review of Books, 63(18):18 20.
  • Discusses technical and ethical issues concerning the automated decisions made by driverless vehicles.

• Monticello, M. (2016). The state of the self-driving car" and Will self-driving cars make our road safer? Consumer Reports, 81(5):44–49.
  • Offers observations on their relative safety.

• There are other automated vehicles besides self-driving cars, for which some of the same issues arise:
  • Markoff, J. (2015, 7 April). Planes without pilots. New York Times, pages D1, D4,
    • Asks, in the wake of a commercial-airline pilot who who committed suicide, thereby killing all aboard his plane, whether pilots are necessary at all.
    • See also two follow-up letters to the editor (New York Times (14 April): D3, that point out situations in which:
      1. computers should take control over from humans, or
      2. humans might make better decisions — because more creative or context-dependent — than computers.

§17.2: Decisions and Computers

On game-playing computers:

• On checkers, see:
  • Schaeffer, J., Burch, N., Björnsson, Y., Kishimoto, A., Müller, M., Lake, R., Lu, P., and Sutphen, S. (2007). Checkers is solved. Sciencexpress

  • "Chinook (computer program)" (Wikipedia)

• On chess and other games, see:
  • "Deep Blue (chess computer)" (Wikipedia)

  • Silver, D. et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 539:484–489.

  • Campbell, M. (2018). Mastering board games. Science, 362(6419):1118.

  • Kasparov, G. (2018, 7 December). Chess, a Drosophila of reasoning. Science, 362(6419):1087

  • Silver, D. et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419):1140–1144.

  • Halina, M. (2021). Insightful artificial intelligence. Mind & Language, 36(2):315–329.

§17.5: Should Computers Make Decisions with Us?

• On team efforts, such as the ones that Hutchins discusses, see:
  • Vélez, Natalia; Christian, Brian; Hardy, Mathew; Thompson, Bill D.; & Griffiths, Thomas L. (2023), "How Do Humans Overcome Individual Computational Limitations by Working Together?", Cognitive Science 47(e13232).

  

• On the Question for the Reader concerning whether human-computer interaction constitutes the use of a (human) oracle, especially in the context of the trustworthiness of large language models, see:
  • Faggioli, Guglielmo et al. (2024), "Who Determines What Is Relevant? Humans or AI? Why Not Both?", Communications of the ACM 67(4, April):31–34.
  

§17.6: Should We Trust Decisions Computers Make?

• For a philosophical investigation of the nature of trust, see:
  • Baier, A. (1986). Trust and antitrust. Ethics, 96(2):231–260.

• Heingartner, D. (2006, 18 July). Maybe we should leave that up to the computer. New York Times
  • Discusses whether computers can make better decisions than humans, and observes that "mathematical models generally make more accurate predictions than humans do. … The main reason for computers' edge is their consistency — or rather humans' inconsistency — in applying their knowledge."

• Brachman & Levesque 2022, p. 213, answers a version of Moor's question ("Is it right to trust an AI system with a rock-solid track record that makes decisions we cannot quite follow?") in the negative.

• For scientific research on wrong or illogical explanations by humans, see:
  • Tversky, A. and Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157):1124–1131.

  • Wainer, H. (2007, May-June). The most dangerous equation. American Scientist, 95(3):249ff
    • On humans' difficulty in reasoning about probability and statistics.

• For other readings on humans' difficulty in reasoning, see my bibliography on "Reasoning"

§17.6.1: The Bias Problem

• For introductions to "Deep Learning", see:
  • LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521:436–444.
    • An introduction to "deep" machine learning by three of its pioneering Turing Award winners.

  • Buckner, C. (2019). Deep learning: A philosophical introduction. Philosophy Compass, 14(10).
    • An introduction to deep learning for philosophers.

• On the Bias Problem, see:
  • Savage, N. (2016). When computers stand in the schoolhouse door. Communications of the ACM, 59(3):19–21.
    • A discussion of how "researchers are trying to identify … and root … out" biases found in classification algorithms.

  • Metz, C. (2019, 16 August). A.I. is learning from humans. Many humans. New York Times.
    • Reports on minimally trained non-professionals who supply the training for machine-learning systems, which raises the question of how accurate those systems algorithms could be.

  • Singer, N. and Metz, C. (2019, 19 December). Many facial-recognition systems are biased, says U.S. study. New York Times.
    • Discusses evidence of bias in facial-recognition algorithms.

  • Smith, C.S. (2020, 2 January). Dealing with bias in artificial intelligence. New York Times.
    • Contains interviews with three female AI researchers on how they deal with such bias in algorithms.

  • For a detailed analysis of bias problems in natural-language processing systems (such as GPT-3) — and as opposed to natural-language understanding systems — see:
    • Bender, Emily M.; Gebru, Timnit; McMillan-Major, Angelina; & Shmitchell, Shmargaret (2021), "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜", CFAccT '21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency: 610–623.
      • Note the parrot in the title of this paper!

§17.6.2: The Black Box Problem:

• On the black-box problem, see:
  • Brachman, R.J. (2002). Systems that know what they're doing. IEEE Intelligent Systems, pages 67–71.
    • A leading AI researcher suggests how and why decision-making computers should be able to explain their decisions.

  • Brachman & Levesque 2022 suggest that AI systems should not only be intelligent but also intelligible.

  • There are (at least) two strategies for trying to understand how deep learning systems work. One is to explicate the statistical algorithms that underlie them. Another is to provide a higher-level, "psychological" explanation or description. For discussions of attempts to get systems to be able to "account" for themselves, see:
    • Lipton, Z.C. (2016). The mythos of model interpretability. In 2016 ICML Workshop on Human Interpretability in Machine Learning (WHI 2016).

    • Kuang, C. (2017, 26 November). Can A.I. be taught to explain itself? New York Times Magazine, pages MM46ff

    • Metz, C. (2018, 7 March). Google researchers are learning how machines learn. New York Times, page B3.

    • Cappelen, Herman; & Dever, Josh (2021), Making AI Intelligible: Philosophical Foundations (Oxford: Oxford University Press).

    • Hill, Robin K. (2023, 29 July), "Bigger than a Blackbox", Blog@CACM:
      … we know exactly how deep learning works. It's not a black box. It's a complex data-gobbling number-crunching white box. We just can't force the results into our familiar concepts. When we try, and fail, it all starts to look mysterious.

• Machine-learning programs are not the only technology to suffer from a black-box problem:
  "Medicines without obvious mechanisms of action are still vetted by the FDA."
  — Herrman, John (2023), "The AI Magic Show", New York Magazine/Intelligencer (18 January)

• Although the following observation is not directly related to the black-box problem as it is discussed in this chapter, it is not exactly irrelevant:
  "There are at least three philosophies that are used in the choice of test data. … Most users of a large program are not interested in the details of its functioning; they only wish to obtain answers. That is, they wish to treat the program as a black box. … In the glass-box method of testing, the logical structure of the program is examined, and for each alternative that may occur, test data are derived that will lead to that alternative. … Let us mention one further philosophy of program testing, the philosophy that is, unfortunately, quite widely used. This might be called the ticking-box method. It consists of doing no testing at all after the project is fairly well debugged, but instead turning it over to the customer for trial and acceptance. The result, of course, is a time bomb."
  — Robert L. Kruse, Data Structures and Program Designs (1987)
  [quoted on the cover of Computing Reviews 31(2) (February 1990)]