Ghosts and Dolls

Sometimes, for some people, a piece of writing’s attribution alone suffices to taint or burnish it. The attribution might be to this or that individual writer, or to one or another kind of writer. Those who take a piece of writing to be tainted in this way need no acquaintance with what it contains to support their judgement; attribution suffices.

It wasn’t, for example, uncommon in nineteenth-century England for female writers to publish under male pseudonyms. The Brontë sisters each did that when they began to publish, as did Mary Ann Evans, who is still mostly known as George Eliot. They did it principally because they and their publishers judged that under a female name their works would attract fewer readers; they would be read, if at all, read differently, with condescension or ridicule. Masculine pseudonymity in this case was a protective strategy. Those who used it wanted attention paid not to their persons in a kind but to what they’d written.

What the Brontës and Evans tried to protect themselves against is strictly a prejudice, a prejudgement, in the sense that it isn’t based upon acquaintance with what members of the kind in question have written. Rather, those who adopt it know, or take themselves to know, something about members of the kind that taints whatever they write. Perhaps it is that such writers are incapable of writing anything interesting; perhaps that they’re so heavily implicated in a noxious political or economic or cultural system that reading them serves to endorse something that shouldn’t be endorsed; perhaps that all writers of the kind are corrupt, maleficent, and dangerous, and therefore not to be consorted with by reading them; perhaps that they’re taking good work away from writers who belong to some other, more deserving kind, and therefore shouldn’t be supported by readers. And so on.

A version of this attitude is at the moment widespread with respect to the literary writing done by large-language generative artificial intelligences. Some take such writing to be tainted just because it’s been done by such entities, without respect to its substance. The analogy with the judgement that women’s writing is tainted just because it’s done by women is formally exact.

A first objection to this analogy might say that what makes the difference between the case of women versus men and that of artificial versus human intelligence is not that the artificial intelligences compose differently than human beings do but that they do not compose at all, which is manifestly not so for women. That is so because the procedures that produce the texts that artificial intelligences seem to compose are run by human-written code and initiated by human-provided prompts. Artificial intelligences execute commands as puppets do; that is the extent of their action, and it is this state of affairs that, while not exactly tainting the texts they seem to compose, makes it improper to say that they have composed them at all. On this view, machines do not write; at best they copy.

This line of response is inadequate to the ways in which coders and computer scientists describe what large-language generative artificial intelligences do. It is by now commonplace to say that these can and do themselves write (the preferred verb in the trade) and compile code that runs routines, and that what they do in response to human requests is neither predictable by nor transparent to those who wrote and compiled the code that brought them into being and established the parameters of their activity. These features of their performance are among the principal reasons for current anxiety about them; they show artificial intelligences of this type to be significantly unlike puppets. They also make them significantly like, in their compositional use of language, human persons. We too have been given language, spoken and written, by others than ourselves; we too compose in response to stimuli that come to us from without; and we too compose according to grammatical rules we did not make and in response to already extant texts composed by others.

“Writing” is here shorthand for composition, which is the act of putting words in order so as to bring into being a legible verbal ensemble. Such ensembles are texts, or pieces of writing. No claim is made or denied by the use of the word about the mindedness or thoughtfulness or meaningfulness of the act by which the words of a text are ordered, or about the nature and origin of those who do it; the only claim the word makes, as used here, is about the legibility of what writers make, of the ensembles of words they compile. So taken, “writing” refers to an act that artificial intelligences and human persons both perform, as women and men both also do. That is the composition of legible texts, which is the same, according to the usage here, as the generation of pieces of writing.

Some might say the analogy between the claim that women taint what they write and the claim that artificial intelligences do is not exact because women are members of the Linnaean species Homo sapiens, who live and breathe and think, while artificial intelligences are machines that do none of those things. Hence there are reasons for taking the compositions of artificial intelligences to be tainted that cannot apply to writing done by women.

But the position remains formally the same: In both cases there is a kind (women, machines), some of whose members appear to write. The texts they compose are extant and may be read. To sustain the claim that the writing done by either is tainted independently of the words (rather than simply assert it as self-evident), there is in both cases only one avenue open, and that is to show what it is about women, or machines, that provides such a taint.

This was once attempted for women. They were said to lack, whether by nature or nurture, capacity for sustained attention and imaginative work, together with wide and deep knowledge of the literature of the past, and these things are among the conditions necessary for any who wish to make a contribution to the literature of the present. These claims, and others like them, are no longer widely made, and when they are they have an air of desperation, which is provided largely by the range and depth and quality of what women have in fact written.

Some say that what makes the difference between human and machine writing—and what taints the latter—is that machines do not think, or intend, or mean. Their compositions, allowing that they do compose, are therefore not minded or mindful, and it is this lack that gives them their taint, as properly it should. Women’s writing is not different from men’s in this way, which is why those who took it to be tainted were illegitimately prejudiced; the prejudice about machine writing is by contrast accurately reflective of a real difference between it and writing done by human beings.

This line is immediately plausible, but it depends upon, or is at least inextricably intertwined with, a high evaluation of the place of thought in human composition, oral and written, an evaluation difficult to defend in the face of ordinary empirical observation about how speech and writing happen among us.

Suppose we consider the phenomenon of fluency. Much of what we say and write is formulaic and habituated, prompted by familiar situations that provide us with our words frictionlessly. Our words often come without the need for deliberation, decision; they arrive ready-made, like our unheeded sensory reactions, as the kinds of things we would say, without our asking why, for what purpose or end, absent friction of the kind produced by having to grope for the right words. Quotidian exchange of the verbal formulae demanded by local etiquette is like this; so is much of the technical speech and writing involved with the practice of a trade or profession; and so too is a good deal of the extemporaneous speech performed by preachers, comedians, lecturers, lovers, spouses, parents, and singers.

Users of language in these situations, and in many others like them, typically need no occurrent thought, and would be hindered by it should it come to them. They find the footlights of the stage on which they do these kinds of verbal work turned very low, and that is because they are fluent. Fluency is smooth and easy: Words flow for the fluent like water running downhill in a deep, smooth channel; there is little to obstruct it, and it goes where the situation demands. That is as true for writing as for speech: From the ends of a writer’s fingers come, very often, sentences unscripted, unanticipated, unformed by deliberation, given, as it seems, not by thought but by afflatus, as the verbal wind blows. That is when things go well, which is to say fluently, as often they do.

As young children move away from the verbal muteness of infancy, they almost ineluctably move toward fluency in whatever the local languages are. That is what their peers and teachers expect of them, and it is where language itself, the word-river upon which they float, takes them. Fluency minimizes need for deliberation. The extent to which speech is intercalated with thought is the extent to which it halts and stutters—becomes, that is, non-fluent.

The proportion of human speaking and writing that exhibits and participates in fluency in this sense is high. It belongs to the forms of life we human beings inhabit in the same way as do such things as walking, eating, drinking, and caressing. Those who take composition, the ordering of words, as evidence of thought will tend to consider all this, fluent speech and writing, most of what humans do with words, as something other than the real thing, because they see real speech and writing as attended by concurrent thoughts and intentions, radiant therefore with meaning. The words put together by those fluent in speech and writing then become simulacra, routinized imitations of the real thing like what artificial intelligences make. And at that point the attempt to discriminate writing done by artificial intelligences from that done by human beings, and to taint the former thereby, succeeds only by tainting almost all human writing as well.

But some might say that thinking can be otherwise identified. Perhaps it is not so much a question of qualia, of how it seems to a speaker to be speaking or a writer to be writing, but rather of meaning: When speakers, even fluent speakers, speak, they mean something by what they say; they intend to communicate something by their words, even when things move too fast and the situation is therefore too fluid for it concurrently to seem to them that there is anything behind or within what they say. That users of words, human ones anyway, mean something by what they say, or at least that they can and typically do, is part—perhaps the central part—of what the connection between putting words together and thinking means.

“What did you mean by that?” Even the most fluent can answer this question. They do it by considering (itself an instance of thinking) what they did mean, and then explaining their past words; these explanations seem to appeal to, and to deploy, something additional and internal to the original statement. That something might as well be said to be thought-formed—intended, meant, chosen, even deliberated; it is, in short, the words’ meaning, what their speaker or writer meant or means by them. This line of reasoning identifies thinking not with something it seems like to a user of words to be doing—a phenomenal or experiential event, something concurrent with and internal to the act of putting words together—but rather with a capacity on the part of speakers to say something about what they just said, or to be able to do so should occasion demand or suggest.

And it is clearly the case that some respond to questions such as “What did you mean by that?” by saying something else, something similar to what they have just said; they offer a gloss upon or interpretation of earlier speech, whether in speech or writing. But it is also clearly the case that sometimes the fluent cannot do this. Perhaps what they just said was a schoolyard chant or a formulaic greeting or a fragment of poetry or an endearment or a song. In such cases, “What did you mean by that?” will seem out of place; it will be unclear whether to answer it, and unclear as well what would count as an answer should one be attempted. Or perhaps a user of words can find no paraphrase or alternative formulation. All, even the most fluent, soon enough reach this position if asked, as by a garrulous three-year-old, a long sequence of questions about what they mean. The request for explanations of this kind always ends with one’s spade being turned: Spades, excavating the inner world for meaning, always eventually get turned, and usually sooner rather than later; the question is not whether this will happen in a given case, but only when. This is as true of writing as of speech.

The picture painted by those who write of meaning in this way is of a radiant jewel available to the introspective gaze of those who put words together. The jewel irradiates their words, and in so doing gives them meaning. Composers of words may not always be attending to this jewel, which permits the phenomenon of fluency; but it is there, and they can (always?) look at it and say something about it if they need to. It is this jewel that provides them the capacity to mean something by what they say, and sometimes to give an account of what they mean. Perhaps, too, this picture is painted so as to embrace all instances of verbal composition: Only if composed words are accompanied and illuminated by such a jewel in principle available to those who do the composing do they count as speech or writing—because only then do they mean. Otherwise, they are something else, something lesser, vibrations or squiggles merely, the work of automata.

This is a desperate picture. It is motivated by a desire to have a theory of meaning that accounts for all speech and writing by appeal to something internal to every instance of them which is introspectible to their composers while at the same time accounting for the undeniable fact that a great deal of what humans say and write does not seem to us to mean anything as we produce it. The picture is also esoteric: It postulates something rarely evident as if it were always there, and suggests, as it must, that there is a small subset of those who speak and write, the ones who have a good theory of meaning, who can discriminate what really counts as language, which is to say verbal compositions illuminated by the ghostly light of meaning, from what is only a simulacrum of it.

A simpler picture—one both more chaste and more adequate to the phenomenon of language as it works in human life—admits the jewel of meaning as an occasional and atypical light-bringer, but shows the meaning of what humans say and write as in general indistinguishable from what they in fact say and write, even in situations where the request for meaning is relevant, which is a small subset of all utterance. Better not to say that we use language because we think, or mean, or intend to communicate, but that we do it as we walk, eat, drink, or play (especially that last), which is to say in such varied contexts that any account of these activities that attributes to them a single purpose or structure inevitably misleads about some, and likely most, instances of the phenomenon under discussion.

Wittgenstein is helpful here. He writes that some say nonhuman animals do not speak because they lack the necessary intellectual capacities; or, differently put, that they do not speak because they do not think. This formulation approaches the speech-thought question from the reverse direction: that of accounting for those who seem not to speak (or write) at all. Wittgenstein prefers to say simply that nonhuman animals do not speak: that they make no use of speech. Such activities as commanding, questioning, storytelling, and conversing are proper to our Naturgeschichte, our natural history or form of life, in the same way as are walking, eating, drinking, and playing. But that is not so for nonhuman animals. They live in the world differently, without the specified verbal activities.

Why is it better, as Wittgenstein sees things, to speak of the apparent absence of complex forms of speech among nonhuman animals in this modestly descriptive fashion rather than to offer explanations for that absence, such as the claim that nonhuman animals do not think? It is better because explanatory claims of that kind suggest a tight link between speaking and something else (in this case thinking), perhaps even so tight as to claim that speech only occurs, or perhaps can only occur, when prompted by or otherwise interlaced with an act of thinking; or that capacity for speech has capacity for thought among its necessary conditions; or something similar.

Speech, on such views, needs to be explained as well as observed. Some possible explanations are of the kind that Wittgenstein would prefer not to offer. (These tend to take the form “If x then necessarily y”: If speech, then necessarily thought; if words, then necessarily meaning, or, worse, reference; if language, then necessarily communication; if action, then necessarily intention; and so on.) They have a strong tendency to flatten, to smooth out, to divert attention from particulars toward a general theory, and, almost ineluctably, to establish a hierarchy among instances—so that, for example, verbal compositions, whether spoken or written, are attended to and ranked not in terms of their evident verbal properties but rather in terms of something about their provenance or attribution: They are preceded by thought; they are accompanied by meaning; they are composed by the right kind of people, or things; and the like. Those are the real thing. The rest, even if they look very like the real thing, are and must be simulacra. The picture Wittgenstein prefers not to adopt here is very like the one that shows the words of artificial intelligences as tainted by their lack of connection to thought.

Wittgenstein also usefully explicitly addresses the question of whether a machine can think or be in pain. “Well,” he asks, “should the human body be called a machine? It’s the closest thing to being a machine like that.” This response makes the question grammatical: Should the human body be called a machine? The answer is yes, Wittenstein writes: As we ordinarily talk of it and see it, the body shows itself to us as very like a machine. It has, as we say, many processes that work autonomously (respiration, digestion, regulation of temperature) in direct response to the ambient situation; it runs down; it needs repair; and so on. There is no ontology in writing like this, no theory of the body or of the machine, no commitment to any claim that all discourse about human bodies is exchangeable without remainder to discourse about machines. Rather there is the observation that much of what we say about human bodies we also say about machines, and that since we are happy to say that human bodies suffer and think—they provide our paradigms for saying such things—any reasons we might have for denying that such talk is applicable to machines must come from elsewhere, from ontology, or metaphysics, or philosophy of mind, or (as Wittgenstein often writes) from a picture of the relation between thought (and suchlike) and action (and suchlike) which requires that the latter be always an effect of the former.

Wittgenstein continues by asking whether the claim that a machine cannot think is based on experience (ours of machines, that is). “No,” he says. “We speak only of man, and what is like to him, I think. We also say it about dolls and probably also about ghosts. Think of the word ‘think’ as an instrument!” The verb to think and its like are applied by us only to human beings and things like them. Human beings, we may say, are for us the paradigmatic thinkers; when we learn to speak, it is principally and archetypally to them—to us—that we apply this vocabulary. But not only to humans: also to what is like them, with dolls and ghosts given as instances. We may add: machines. The car I drive detects my distance from the one in front of me, and when on cruise control it slows the car when it thinks—decides, judges—the distance too small, and then speeds it up again when the distance seems large enough. The vocabulary of thought comes naturally and easily here, and that is because what my car does—and it does it; I do not—seems to me, and I suspect to most drivers in this situation, sufficiently like what I do when I put my foot on the brake or the accelerator that I happily use the same vocabulary for it. A decision has been made, and the automobile has made it.

It is not that these ways of talking commit their users to any theory about how automobile cruise controls (or artificial intelligences) do what they do. Neither does using the vocabulary of thought about ourselves commit us to, or provide us with, any such theory about ourselves. For the most part, the means by which I make decisions are as opaque to me as those by which my car makes its own, and these absences do not at all hinder my capacity to perform the one and co-operate with the other. And in both cases I am fluent enough in English to know how to talk about both, also without having or needing any theory that would yield answers to questions such as “But do you mean that the automobile (the artificial intelligence) thinks as you do?” No, I do not mean that; I mean, rather, just what I say, which is that what those things do shows itself, often (not always), as sufficiently like what I do that I easily and fluently speak of decisions being made and thoughts being had in all these cases. That is Wittgenstein’s conclusion as quoted above: “Think of the word ‘think’ as an instrument!” (The exclamation point is no accident: Wittgenstein is careful about such things.)

Perhaps a different response might be made to this problem, that is, to the fact that speech and writing seem often to take place without concurrent thought. It might appeal to the development and use of language over time rather than to snapshots of its use in the present. Maybe habitual and fluent forms of linguistic behavior are not, on most occasions, connected with thought; and perhaps, too, it is at least sometimes the case that fluent users cannot provide an account of what they mean by what they say because they have become so accustomed to saying it—have become, in short, fluent.

Is it possible to become the kind of writer who can fluently, extemporaneously, write on, say, Pascalian depictions of the workings of grace, or macroeconomic analyses of the goods of trading in notional futures (two equally esoteric instances), without having decided and chosen to engage in the kinds of reading and study and writing and thinking that inform those fluent skills? Are not all, or at least many, fluencies related to thought in at least this way? Might this not be a way of saving the jewel of meaning?

This could be how fluency is developed. Deliberative thought might give rise to the conclusion that it would be good to develop this or that skill, that the skill serves a purpose the deliberator wishes to be served, and that there are things to do, courses of action to follow, that foster it. These, however, are atypical: They are causally implicated with only a small minority of linguistic fluencies. Fluency can only be developed under contingent and local circumstances; we are often unaware of the presence of these conditions, and do not consider their desirability. One might become the sort of person who fluently theorizes about the workings of grace in small part as a result of thinking about whether it would be a good thing to achieve such an end, and about which practices and habits will help us to become such a person. But in large part the conditions for the development of such fluency are prevenient, given to persons as conditions for deliberation rather than as products of it. (They might, in the case canvassed, include the presence of a functional canon of works about grace as a Christian theological category, the gift of learning Greek, Latin, and certain vernacular languages.) Verbal behavior absent concurrent thought might have been made possible by past behavior. But this point has limited application. It remains the case that most linguistic fluencies, even those inflected with and motivated by this or that instance of deliberation at some points in the course of their development, are in origin and in act very largely free of thought—delightfully, frictionlessly free of anything other than themselves.

Better, in these matters, to stay close to the rough ground of the forms of linguistic life we live upon. Here we can observe how those forms work, what they are like, and what their modalities are. That ground is always rough, and it will be artificially smoothed, made almost to vanish, transfigured even, by the heavy earth-moving machinery of univocal and universal theories of meaning and the relation of speech and writing to thought. When we renounce that machinery, or, better, give it its proper place as a verbal activity of specialized interest, beautiful in its own way, like good jazz improvisation, but without purchase upon or power over the array of speech and writing which it purports to be about, we can then see, clearly enough, that most of our language is fluent, and is therefore not in fact informed by deliberation. It is mostly like respiration and digestion and play: something that we do easily and well, and that has a place of central importance in the lives we live. (Philosophy, like all writing, is a kind of writing: It is without sovereignty.)

But surely not all human uses of language are fluent in the sense discussed. Infants in the early stages of learning to speak are not fluent; non-native users of Sanskrit or Latin grope for grammar, lexicon, and syntax as they work on school compositions in those languages; native speakers of some language, thrown into a situation unfamiliar to them, with norms of talking and writing they can scarcely enter into, will not know how to go on. And so on. In such situations it seems that thought—or at least something other than utterance or writing—does precede or accompany language; and it may sometimes be that the non-fluent are reduced to silence by their lack of fluency while not thereby ceasing to have an inner life capable of introspection, perhaps one of tension or frustration or embarrassment.

And that is certainly so. Human beings often are in situations in which they lack fluency. When this happens, their sense of what it is like to be themselves may have the characteristics I have mentioned, along with many others. A small instance: A fluent writer of English attempts a paragraph in French, in which she has some competence, about the nature of hope—about, perhaps, the necessary absence of hope’s object. The composition goes well enough initially; but then the writer looks over what she has written, in which her preferred word for hope is espoir, and replaces it throughout, doubtfully, with espérance. She looks at the paragraph as it now stands and finds that it seems to her neither better nor worse than its predecessor. She discovers that she does not know how to go on; she has no judgements or intuitions about the difference in denotation or connotation between these two words; she cannot decide which of the two is more appropriate for her paragraph. She is frustrated; she considers whether use of a dictionary, an online search (consultation of an artificial intelligence, perhaps), or a chat with a native speaker—or at least someone more fluent than she—is what she should do next. Perhaps she should read more widely in the relevant literature in French. She begins to consider what she has read about synonymy, and how it might apply here. To her it seems like something rather than nothing: There are qualia; she has an occurrent sense of what it is like for her to be in this situation. She is, we can say, thinking about what she has written and what she might write next. Here at any rate thought is separable from writing (and speaking).

This example suggests that thought’s separability from language is characteristic not of an ordinary and positive distinction between the way in which human beings arrive at what they write and the way in which machines do but rather of failure on our part in using language. It belongs to situations in which we do not know how, linguistically speaking, to go on. (There are similar failures in other areas of human life: A walk in the mountains may gradually become too steep for a walker; he looks around and cannot see how to go on; he turns back to safer ground.) Suppose that situations like this were regarded as characteristic of human uses of language, and their absence not just equally characteristic of machine language but the reason we regard such language as somehow tainted (they cannot be anything other than fluent; that is what is wrong with them). We would have arrived at a reductio. Our lack of fluency is untainted speech; when we write fluently what we make is tainted because it means that we write as machines do. Linguistic success—fluency—is on this view exactly its shortcoming.

Better to say that both machines and human beings are capable of fluency in writing, and in both cases the degree of fluency in a piece of verbal composition is a close index of the absence of qualia for the composer; that both machines and humans are capable of not knowing how to go on with a piece of verbal composition (artificial intelligences sometimes tell their human interlocutors that they cannot go on, that their spade is turned); that in the human case such not-knowings are likely to be provoked or accompanied by qualia; and that in the machine case they may not be (we have, at the moment, no way of knowing).

Should we attempt to save thought—in the sense of intention or meaning—as the primary difference between human and machine composition, and its absence for machines as an inherent flaw? What would this look like? It depends in part upon how we define thought. Perhaps thought is not phenomenological data bound up in the act of composition nor a mere characteristic disposition but a condition, more or less invisible from the perspective of introspection or even memory, common to all living speakers, attributable to them precisely because they are animate­—ensouled as that word suggests, and ruled out for machines just because they are soulless. If this is the case, we can say that the literary compositions of machines are tainted simply because they are the work of something without life.

This can be said. There is a long lineage of saying it. The example of taint attributed to the writings of women is, in some of its versions, similar: When women are taken to be incapable of literary composition because of their physiognomy or chromosomes, the analogy is strong. More pointed still is phrenology: Nineteenth-century practitioners of that quasi-scientific discipline were happy to say that those with skulls of a certain shape and capacity could not, for example, compose music or engage with mathematics even if appearances were to the contrary: even if it appeared that those with such skulls were doing exactly those things. The shape of their skulls was taken to be enough to show that their apparent performance of these activities could not be of the things themselves but only of their simulacra. This analogy is precise: Carbon-based beings, with blood and brains, if they belong to the Linnaean species Homo sapiens, can write; silicon-based beings, with code and algorithms, cannot, even if they appear to be doing so. Advocates of this position, having renounced their ability to appeal to thought, to anything internal to the act of composition or anything from which such an act might be separable—an appeal we have seen to fail—can look only to what is physically given, and can offer reasons only of the kind that calipers provide phrenologists. Such are the appeals made by biologists to breath and blood to explain why beings of a certain configuration can really do something while those of another configuration cannot, even when they seem to be doing it. Such advocates are insulated from the need or desire to look at—to read—things made by beings not of the relevant species. The limitations they impose upon the capacity to write to things of a certain shape or configuration or physiognomy is a simple assertion, without rationale or the possibility of rationale, untroubled by attention to what beings in the world do or fail to do. Such positions suffer from all the disadvantages, as well as some additional ones, of the claim that members of the Linnaean species Homo sapiens can procreate only with others of that species. One of these disadvantages is that it is false, upon grounds which are empirically demonstrable.

Is there, then, a taint that belongs to the writing done by large-language generative artificial intelligences independent of the literary features of the works they compose? There is no reason to think so. That conclusion is compatible with all such works being, in fact, trivial, crude, incoherent, or in other ways uninteresting or undesirable. What they are like, however, can be determined only by reading them.

And there are reasons why human readers should be eager to read literary works composed by large-language generative artificial intelligences. These reasons are properly analogous to those for which people are eager to find and study forms of life that did not evolve on our planet. The peculiar flavor of any literary work is given in large part by the formation, literary and otherwise, of its composer: the works she has read, the languages she speaks and writes, the exchanges she has had with other speakers and writers, and so on. These inform and vivify her own writing; they make it idiosyncratic, often to the degree that an informed reader can recognize it even when it is encountered anonymously or pseudonymously. It is a matter of style. The formation undergone by artificial intelligences is different, deeply so, from any that a human person can undergo. They have read more than we, they interact with what they have read differently, they compose in ways distinct from our habits of composition, and so on. These differences, if contemplated even for a moment, suggest that their literary works will be unpredictably different from ours—they may show us how to do things with words that would otherwise not have occurred to us, in very much the same way that living creatures evolved elsewhere than on our planet might show us forms that life can take that we would not otherwise have imagined, and from which we can learn things otherwise unlearnable.

These possibilities might seem to us in both cases good, or bad, or a blend of both, as someone formed in both taste and habit by European music can be startled, delighted, and instructed—or, as the case may be, bored—by hearing a gamelan orchestra for the first time. What artificial intelligences show us in their writing will in any case differ from what human writers show us. It is, again, a question of reading what they write: We know neither that what machines write is in principle tainted nor that it is not, we know that the reasons given for its inherently bad quality so far carry no weight, and we can expect with confidence that what machines write will be different from what we write because it will be evidence of something we do not share. We can, if we permit ourselves the abandonment of prejudgements for which we have no reason, be interested in rather than anxious about what machines write.