OpenAI’s GPT-3.5 is one of the consummate promulgators of apparently meaningful nonsense on the Web today. Roughly speaking, it generates output on the basis of a predictive model, where each next word is probabilistically produced on the basis of preceding words. It joins together words that seem like they would make sense together. The problem with this approach is that when it comes to concrete things in reality, such as titles of works, the AI sometimes produces lists of items that are merely plausible-sounding rather than really-existing. For the serious student or scholar, this means entering dangerous territory: Any number of creative works could exist, after all, but what matters more often is whether they in fact exist.
Let’s look at an example. I asked the AI to “[p]roduce a list of twenty articles or essays on Hegel's relationship to or ideas on Christianity that I should read and are available online.” ChatGPT produced the following list of 13 items before bugging out:
"Hegel's Philosophy of Christianity" by Stephen Houlgate
"Hegel and Christianity" by Laurence Dickey
"The Religious Dimension of Hegel's Philosophy" by Frederick G. Whelan
"Hegel's Theological Turn: A Comparative Study of His Early Jena and Later Berlin Manuscripts" by Paul K. Moser
"Hegel and the Problem of Evil" by Robert Stern
"Hegel on Christ and the Christian Religion" by Peter C. Hodgson
"Hegel's Religious Thought: A Philosophical Interpretation" by Laurence Dickey
"Hegel's Interpretation of Christianity" by Laurence Dickey
"Hegel's Dialectic of Religion and Philosophy" by David L. Johnsen
"The Role of Religion in Hegel's Philosophy" by Andrew Buchwalter
"Hegel's Concept of God and Its Relevance for Today" by John McCumber
"Hegel's Concept of the Divine" by Robert C. Solomon
"Hegel's Aesthetics and the Christian Tradition" by David Kolb
Now, for anyone outside of the field of Hegel studies, this list must appear highly plausible; and indeed, all of these titles look like they could very well be highly interesting works. The problem, however, is that none of them are real as stated in their current form. The first listed author, Stephen Houlgate, is a very real, and eminent, Hegel scholar, but he has, to the best of my bibliographical knowledge, never penned an article entitled “Hegel’s Philosophy of Christianity.” Item number 10, “The Role of Religion in Hegel's Philosophy,” is the title of a really-existing book chapter, but in a volume authored by Raymond Keith Williamson, not the listed Andrew Buchwalter. The eleventh title, “Hegel's Concept of God and Its Relevance for Today,” is pure fiction—though, again, one might very well want to read a work by that title.
I tried to remedy what we might call ChatGPT’s credible nonsense propensity (CNP) by asking the AI to include DOIs or URLs in the next prompt, a kind of feeble attempt at introducing certain standards of accountability: “Produce a list of twenty articles or essays on Hegel's relationship to or ideas on Christianity that I should read and are available online. Include DOI links or URLs.”
Here’s what came out on the other end (ChatGPT aborted after six items):
"Hegel and the Problem of Christian Ethics" by William Desmond (https://doi.org/10.1163/156916202760667748)
"Hegel and Christian Theology" by David J. Wood (https://doi.org/10.1163/1569162054444361)
"Hegel's Philosophy of Christianity" by Philip J. Kain (https://doi.org/10.1163/1569162042600230)
"Hegel's Critique of Christianity: The Rejection of the Christian Doctrine of God" by Andrew Pessin (https://doi.org/10.1163/156916203322996561)
"Hegel's Theology and the Role of Christianity in Modern Society" by John E. Toews (https://doi.org/10.1163/156916205774968386)
"Hegel's Concept of God: An Analysis of the Concept's Development and Implications" by Timothy J. Stapleton (https://doi.org/10.1163/156916203322537345)
But ChatGPT would not be deterred from falsity: the DOIs are also fake. The AI is, in other words, an accomplished fibber. Even when asked to introduce a kind of referential standard by which the authenticity of its output might be checked, at least in its present state (GPT-3.5), it shamelessly revels in artifice: Pursuing any of the links will immediately result in a “DOI Not Found” message. (DOI stands for a Document Object Identifier, a universal indexing tool used in academia to keep track of scholarly or scientific work.)
So what’s going on here? Most if not all of the authors listed above are real scholars of German idealism, continental philosophy, and/or Hegel. But the titles and links are all either fake, “misauthored” (i.e. the title exists somewhere, in some form, but not by the listed author), or—more frequently—simply made up on the basis of plausible-sounding patterns of collocated terms. Thus, “Hegel's Critique of Christianity: The Rejection of the Christian Doctrine of God,” though bombastic, could very well be a (cranky) reading of the great German philosopher’s take on Lutheran Protestantism. But in reality—and this is the more important question—it is not. It does not exist.
It would be foolish to insist this erroneous type of output couldn’t be corrected in future iterations by some technological wizardry, but we see here, I think, the limits of AI in its present state. OpenAI’s model in its current shape and form is less good at telling us things about the world in its current state—how the world is. But it is quite good at telling us about the world in its potential state—suggestively signalling to us how it might be.
When used correctly, AI can be highly generative, serving as an auxiliary to human-centered intellectual labor, a mere appendage to human reasoning activities. Anyone working in the specialist field of Hegel studies, for instance, could take the list of 19 titles above, subtracting those that are merely mislabelled, and find inspiration for the title of their next paper. AI tools like ChatGPT are not the author replaced, but more like an author’s amanuensis, assisting in bringing the virtual world of potentiality into being.