The Limits of Computer Decision-Making: Why Higher-Order Decisions Should Remain Human
An answer to the question 'What decisions should computers make?'
I wrote this essay in response to the below question for an essay competition. I don’t stand by most of the things I said—probs cuz I didn’t think much about it so you (yes, you the reader) should come up with a better idea than this.
‘What decisions should computers not make?’
Almost completely, computers are, or will be, better than humans at performing tasks. Computers solve mathematical proofs we cannot (Wikipedia contributors (2023), Bansal et al. (2019)); they play better chess (McIlroy-Young et. al (2020)); they outperform doctors in some diagnoses and clinical research (Longoni & Morewedge (2019))—and will make ground in others (Shaheen (2021)); and they excel in plenty of other domains (Mellers et al. (2023), Tetlock, (2017)). Even in purported areas where unique qualities were supposed to keep us more suitable than the computer, confidence is faltering: there is evidence computers can give more empathetic responses (Ayers et al. (2023)), and with the recent progress in generative AI, now only the best of us are more creative (Koivisto & Grassini (2023)). I will argue it is misdirected, therefore, to categorically suggest computers should not make decisions for these tasks. When trained for the tasks’ aims and goals, computers are more accurate at achieving them. This should not be surprising: computers are programmed to maximize an end. What I will argue here is a higher-order decision-making, setting an end, should be off-limits. The computational process of moving from an ‘is’ to an ‘ought’ is not a problem of removing bias, more computational power, or better training. What would one be removing bias for, adding power after, or training it to be? The very process of programming computers presupposes an end.
By missing this, principles for restricting AI decision-making are misdirected and often redundant. Floridi and Cowls (2019) compile five ethical principles for AI based on converging suggestions in the literature, but principles such as ‘Beneficence’, ‘Non-Maleficence’ and ‘Justice’ are liable to mistreatment. Setting ends for computers towards these is certainly a must. But this is not restriction on computer decision-making: it is restriction on human decision-making to not set poor ends. Actually applying these restrictions to computers making decisions about ends, however, makes them redundant. If we do intend to restrict computers accurately and without bias, we would have to already know what is, say, just to ascertain whether the computer violated it. If this is the case, the computer’s decision-making is no longer useful. Therefore, insisting all AI projects concord with these principles may merely enforce a particular conception of justice and beneficence over more pervasive and life-encompassing systems.
The other principles of ‘Autonomy’ and ‘Explicability’ also fail to be applicable considering higher-order decision-making. However, despite these being more emphasized in AI ethics, they are, because of this flaw, also unjustified. Ceding decision-making to computers raises concerns that it 'may undermine the flourishing of human autonomy’ (Floridi and Cowls (2019), p7), and ‘Explicability’ ensures computers are ‘understandable and interpretable’. While the intention to minimize the diminishing of human autonomy is good, in practice, it is contradictory. For example, ceding diagnoses to computers creates a trade-off between autonomy/explicability and effectiveness: efficiency is improved, but some may argue it unjustly removes the autonomy of doctors. Following Railton (1984), however, sticking with the doctor inadvertently decreases the autonomy of those who could have been correctly diagnosed by causing delimitations, disabilities, and loss of life, which prevent them from living full, autonomous lives. Therefore, these principles are often limiting or redundant and, moreover, do not prevent poor higher-order decisions from perniciously warping the content of these principles.
Higher-order decision-making also inherently introduces bias. Computers, requiring an end to be inputted, cannot move from an ‘is’ to an ‘ought’ except via inputted biases of humans. As said, restraint to principles is redundant if the end is principles. Thus the computer’s end cannot be constrained by any firm principles and becomes representations of the programmer’s desires. Unless the programmer’s desires are infallible—which, given the examples in history of how culture hides our moral flaws, would be unlikely—, this must, inherently, introduce bias. This is highly susceptible to exacerbation, as, unlike the laws of mathematics, what follows from a person’s belief system does not necessarily align with that belief system. Much of the philosophy of ethics is producing counterexamples wherein a decision that aligns with the system has an outcome that even the believers in the system would not accept. An AI following such a belief system would likely produce outcomes more extreme than its biassed input. At best, then, this bias would be a progress-inhibiting force, instilling and embedding the moral norms of today (or the dominant group’s norms) through an endless introduction of ends based on the previous ones. At worst, the bias is a degenerating force that corrupts our moral ends.
This is a bias that cannot be eliminated through increased computing power or by fine-tuning algorithms because programming computers presupposes an end. J.S. Mill distinguishes between art, which identifies ends, and science, which studies how to attain them (Mill (1858), A System of Logic, Book 6, 12.2). As machines are programmed to maximize an end, computers, as said, excel in science. However, if for arts, ‘a clear and precise conception of what we are pursuing would seem to be the first thing we need, instead of the last we are to look forward to’ (Mill (1864), Utilitarianism, 1.2), and computers’ required ends for setting ends rely on the biassed desires of their programmers, then it suggests computers should not be the ones deducing these general principles and setting ends. More efficient algorithms will not help: algorithms, qua presupposing an end, are just unsuited. Greater efficiency will help produce better means, better theories, to an end, but for producing ends, this means merely more closely aligning to the end it was already given.
A clear and precise conception so algorithms’ biases are not overly magnified is unrealistic since a person correctly gaining such an understanding is unlikely. Understanding what is socially good is far more arduous than what naturally is. The social sciences, v.g., face far more challenges compared to natural sciences due to difficulty in verifying and falsifying hypotheses. For example, in building bridges, the applicability of theories can be quickly tested, while in social sciences, testing theories on, say, community bridging is a lot harder to ascertain. Such lacks quantifiability, making objective falsification challenging. Phenomena are society-wide, so it is hard to study enough people, but differentiated amongst people, so determining effects from what, is hard to control for—and for those that are studied, selection bias is introduced. Measuring societal outcomes takes generations, as effects take sometimes lifetimes to manifest, and society evolves so rapidly the society one measures is different from the society that produced the hypothesised effect. Compared to the verifiability of constructing a bridge, the prevalence and persistence of immorality—and its modern residue—in humanity’s most horrendous regimes suggest such social goods are much harder to discover, understand and imbue. Given the exacerbated effect of bias, its persistence within society should caution against making this the foundation procedure whereby computers calculate ends.
Implementing this form of AI, then, would be very detrimental. Computers are more efficient calculators than us. They would be able to mass produce ‘oughts’ from these biases at immense speeds and with immense complexities: thousands of papers a day could be written on ends for thousands of situations. This would far surpass the rate scientists could scour through all the computer produced to identify and write criticism of its underlying bias so harmful effects could be foreseen. Moreover, scientists’ bias when attempting to independently assess work is already significant (Mahoney (1977), Bornmann et al. (2010)) so the ability of academics to find and criticize would already be too weak. Designing AI to avoid these criticisms would just make the bias harder to detect. It internalizes the externality of bias! Training AI to evade the biases pointed out by academics causes the AI’s bias to both become more complex and harder to detect, and incorporate the bias of academics. This creates a vicious cycle where humans, as AI ends up training on them—either naively to reduce bias or because programmers have incentive for their AI to have less criticism—get more and more outperformed as computers are, or will be, better at performing any task. Adding an extra computer to criticize does not help. Another computer will be able to match the speed of production with criticism but it remains as internally biased. Pitting computers against each other generates endless unengageable critiques based on the same human-instilled bias. There is no progress: the bias the humans started with remains, just escalated with higher production.
Allowing computers to make higher-order decisions means either entrenching society’s biases (or the dominant group within society) into computers that outcompete any human, or the biases are exacerbated, and the AI diminishes our moral ends. As Kant suggests (Kant (1785) Groundwork of the Metaphysics of Morals, S2), all acts are for some end and thereby are conferred relative worth. As organic life, we are end setters: the inorganic are set ends. Programming it to produce ends is futile: its output will be a withered reflexion of the end that is set, its worth that which is conferred by us. Thus, it should not be making these decisions: we should.
Bibliography:
· Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., ... & Smith, D. M. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA internal medicine.
· Bansal, K., Loos, S. M., Rabe, M. N., Szegedy, C., & Wilcox, S. (2019). HOList: An environment for machine learning of higher-order theorem proving. arXiv preprint arXiv:1904.03241.
· Bornmann, L., Mutz, R., & Daniel, H. D. (2010). A reliability-generalization study of journal peer reviews: A multilevel meta-analysis of inter-rater reliability and its determinants. PloS one, 5(12), e14331.
· Floridi, Luciano and Cowls, Josh, A Unified Framework of Five Principles for AI in Society (September 20, 2019). Available at SSRN: https://ssrn.com/abstract=3831321 or http://dx.doi.org/10.2139/ssrn.3831321
· Kant, I. (1785) Groundwork of the Metaphysics of Morals.
· Koivisto, M., & Grassini, S. (2023). Best humans still outperform artificial intelligence in a creative divergent thinking task. Scientific Reports, 13(1), 13601.
· Longoni, C., & Morewedge, C. K. (2019). AI can outperform doctors. So why don’t patients trust it. Harvard Business Review, 30.
· McIlroy-Young, R., Sen, S., Kleinberg, J., & Anderson, A. (2020, August). Aligning superhuman ai with human behavior: Chess as a model system. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 1677-1687).
· Mahoney, M. J. (1977). Publication prejudices: An experimental study of confirmatory bias in the peer review system. Cognitive therapy and research, 1, 161-175.
· Mellers, B. A., McCoy, J. P., Lu, L., & Tetlock, P. E. (2023). Human and Algorithmic Predictions in Geopolitical Forecasting: Quantifying Uncertainty in Hard-to-Quantify Domains. Perspectives on Psychological Science, 17456916231185339.
· Mill, J. S. (1864). Utilitarianism.
· Mill, J. S. (1872). A System of Logic, Ratiocinative and Inductive, 8th ed.
· Railton, P. (1984). Alienation, consequentialism, and the demands of morality. Philosophy & Public Affairs, 134-171.
· Shaheen, M. Y. (2021). Applications of Artificial Intelligence (AI) in healthcare: A review. ScienceOpen Preprints.
· Tetlock, P. E. (2017). Expert Political Judgment: How Good Is It? How Can We Know? - New Edition. United Kingdom: Princeton University Press.
· Wikipedia contributors. (2023, September 7). Computer-assisted proof. In Wikipedia, The Free Encyclopedia. Retrieved 14:01, September 30, 2023, from https://en.wikipedia.org/w/index.php?title=Computer-assisted_proof&oldid=1174243728