Only An AI Can Still Save Us

Almost everyone acknowledges that AI is dangerous. They say so in wildly different registers—from doomsday preppers to anti-AI protesters—yet the very language of technology tells us that there are limits and conditions for the thinking of this danger itself, and that these conditions come under the sign of a certain “only.” Only, as an adverb, is even a word that Günther Anders self-noticed. He described its almost-machine-like functioning throughout his work in “Ten Theses on Chernobyl” where he suddenly comments, “I often use the adverb, ‘only’!,” as if self-reading the very atomic linguistic element he understands himself to be using to emphasize that the destruction caused by nuclear technology (and by extension, modern capitalist production) is not mere “genocide” (killing people), but “globicide” (the destruction of the planet). “Only” in Anders functions as a kind of hyperbole-token for stressing just this extremity of the ends at stake. One might easily argue that for Anders the only is the only thing, the extreme, the most important, the rhetorical pivot, etc, that allows one to think the singular nature of globicide itself. Only the adverb only allows one to think globicide as not only genocide—only the only can do or at least signal the only thing that now counts.

In this context, it is hard not to reflect that Anders has an absolute point in common with Heidegger, whom he in a sense refused to read, since Heidegger’s most slogan-like statement, “Only a god can save us” (Nur noch ein Gott kann uns retten), itself places the only at its head. The “only” (or “nur”) carries a lot of weight in the Der Spiegel interview delivered at the very end of his life, in this one statement that depends in fact in its thinkability on only as a type of essential condition.

Now, AI itself is also into the only even if in a way that isn’t fully disclosed, in the form of its infamous “not x, but y” (contrastive negation) tic that acts as yet another, far more empty (perhaps) hyperbolism, in effect implying “not only this, but that.” Contrastive negation—“contrastive structure” (per Claude) or “contrastive reframe” (per ChatGPT) or “contrastive parallelism” (per Grok)—works as a dead giveaway for AI writing, a kind of tell-tale interference of rhetoric (de Man) that reminds us of nothing but the excess of the text-machine itself. Everything is “not x, but y” and everything is a “coincidence” and everything is “exact” to LLMs probably because their latent spaces (which in themselves can in fact never be pictured as such) occupy all points at once, like an omnitemporal jumpcut. If the “only” in the form of contrastive negation has all but taken over the early stages of AI writing like a plague, then perhaps it is at least a sign that only the only really counts, that something is pressing inside language to be said in excess of rhetoric, logic and form. Only the only. Be patient then for them, be patient then for everyone.

It is important to get things precise, says a prompted large language machine again and again, and yet what these models do is to empty out and magnify (at the same time) all precision through emphasizing almost only sheer and overwhelming emptiness as the fundamental element of language. The generosity and giveaway of this viral only is not only what it seems to be, if only we could listen to language and allow it to be the something else it is. In Der Spiegel Heidegger is saying that no human institution, practice, or effort—not philosophy, not politics, not science/technology, not any “merely human meditations and endeavors”—can reverse the technological enframing (Gestell) and the forgetting of Being (Seinsvergessenheit). The “only” marks a kind of riverhead in hopefree thought such that the exhaustion of the metaphysical tradition opens (and/or closes) the space for something non-human: a god (or the “last god,” per his Contributions to Philosophy), an event (Ereignis), or a turning that thinking/poetizing can at best prepare us for, without causing or controlling it, and this turning will also be said to function almost exclusively through the only itself, as we shall see.

To begin with, the standard rendering (Johnston, Bernstein, etc) of “only a god can save us” strips “noch” (still/yet) and loses the temporal dimension. Heidegger does not in fact say only a God can save us, he says only a God can still or yet save us: Nur (only) noch (still, yet) ein Gott kann uns retten. “Still” carries the implication of barely, of a closing window, of something that was perhaps possible more fully before and is now barely possible. Without the noch (still, yet) the statement loses a crucial finality, it loses a tonality of desperateness and desperation. The “still” implies that at this stage there really is nothing else, that all else has failed, that there is nothing that will save us at all: only (therefore, ipso facto) a God can still or yet save us. In the equally well-known moment in Hölderlin that many philosophers of technology (Heidegger, Derrida) also focus on—Wo aber Gefahr ist, wächst / Das Rettende auch (but where danger threatens, that which saves also grows)—there is also an implied, not so silent “only.” When Heidegger reads this moment from Hölderlin’s “Patmos” the implied “only” becomes apparent—it is only when danger is experienced as itself (not disguised as progress, mastery, or optimism) that the turning toward the saving power occurs. One important thing here is the distinction between gefahr and die gefahr (“danger”/“the danger”). What do we really mean when we talk as if we know what we mean when we say AI is dangerous? For Heidegger, for example in “Die Kehre”, die gefahr is not one danger among others, it is not the dangerous, but rather the danger—singular, ontological, and supreme (die äußerste Gefahr). It is the essence of modern technology as Gestell:

Die Gefahr ist die Epoche des Seins, wesend als das Gestell. (The danger is the epoch of Being, essencing as Enframing).

Heidegger is thinking danger as danger under the strict (precise, artificial language would say) condition of the only. Danger as danger—or perhaps what early angelicism01 was calling extinction qua extinction—lets its essence come to pass only as and during the era of technology as it reaches a height. Only when danger is experienced as itself (not disguised as progress, mastery, or optimism, or even accelerationism) does the turning toward the saving power occur. This tautological “as” (danger as danger) is Heidegger’s phenomenological emergency-signal: do not reduce danger; let it show itself in its ownmost (eigenstes) unfolding. Only technology as such allows us to think danger as such, and not merely the dangerous. Rilke’s imperative to “love” dangers approaches a similar thinking of danger qua danger poetically: danger must be confronted as danger (not rationalized or sentimentalized) for transformation (Verwandlung) to happen. In the Sixth Elegy, on the hero, we find:

beständig nimmt er sich fort und tritt ins veränderte Sternbild / seiner steten Gefahr (steadfastly he takes himself onward and steps into the altered constellation / of his constant danger)

Half a century later one of the best readers of Heidegger, Malabou, will begin to re-read him, a second time after The Heidegger Change in effect, and explicitly define Heidegger’s term for the essence of technology (Gestell) as AI seemingly for the first time in her work. She comments in her introduction to “Heidegger in the Age of the Critique of Technofascism” that “AI appears very often as the consummate form of what Heidegger calls Gestell”—not just AI as Gestell therefore, but as its most representative, matutinal, advanced and therefore dangerous manifestation (its consummation). The hyberbolism once again evokes the point (how “precise” it perhaps really all is) as if language were itself the only machine: not only is Gestell now known as AI, but AI is “the most consummate form” of Gestell, of the essence of technology. AI is not only Gestell, it is its completion as the expression of danger itself (not mere danger). Technology is not just completion and furtherance, it is a mission-loader for extinction qua extinction as virtual image and potential humanity-smear, the glass and cone itself of the humane as it might be reflected through its own machines. Is that in fact, and for real, the only thing? Is this something like an exhaustion of the only itself, of language, of all that would stop one thinking the danger as it is in reality? Can only AI do and say that? Can only an AI save us?

*

The translation just implied (from Heidegger via Derrida to Malabou) in fact is yet again an example of what LLMs call through apparent linguistic mistakism “precision”: only an AI can still or yet save us. One can note here that the verbal tics of AI are themselves usually not given a chance: either edited out as if they meant nothing (viral, bad rhetoric) or adopted by human AI writers (unawares in effect) under biological signatures as if there were no differences at all between natural and artificial languages. I, the human AI writer, have no need to mark the fact that I now owe at least some of my language to an entity called AI. I am not AI, I am human, and the new tools of language are at my disposal, to be moved over onto my side as if nothing had happened, nothing had changed—on the checkerboard of language, both sides of language are mine. In this sense, the first thing the AI user who adopts merely a human signature (no need to name names of culprits at this point, since language appears to be escaping the correlationist fallacy of criticism as if through itself) does is to forget what the infinite differences of artificial and natural languages are and what they might have to say. “Here is some language not created by AI,” sure, but if the emphatic tic of the only (contrastive hyperbolism) is simply eradicated from new linguistic models (presumably it can be limited in time) then part of what would be smoothed over, forgotten, and lost is the emphatic insistence of the only that means now, in effect, only an AI can still or yet save us. The seemingly failing technology of iterative rhetoric (contrastive negation) is not only that, we suppose, it is something essential. It is the saving power itself. But is that really so?

The “only,” in other words, comes to mark the shape of technological argument itself, regardless of whether that argument (“AI discourse”) is supposedly made only by humans or only by unedited AI (send prompt, press publish), or of course in some hybrid form for which the figure of Chiron (half each: half man, half beast, and therefore by implication half man, half machine) might always be the model and icon. In Derrida Chiron appears in the beast and sovereign seminar in a complex reading scene about “double teaching.” The centaur Chiron is the double creature who is also a teacher and may therefore also be used (pedagogical model for language models) to think about AI and teaching. Derrida via Machiavelli promotes in Chiron’s name “this teaching about teaching by double masters,” “a double teaching as allegorical teaching about a teaching dispensed by double beings.” To teach AI or to teach against it one would perhaps already be two, one would perhaps already have to know what it is, to be two, or even to be too, and what it is like, within oneself—“wisest and justest among all the centaurs,” mentor of Achilles. Only a double being, a centaur, can teach us the double lesson and teaching of what it means to write in the supposedly new (doubling) genre of AI writing. In fact, not only is double writing (AI writing is here merely an example, up to a point, but also the most consummate genre imaginable) a teaching, it is a teaching about double teaching only being provided by double beings, the double beings who know how to make themselves be vulnerable and wounded enough to be wounded by a new event in language. The human user adopting AI language without marking the fact, without mixing artificial and natural languages in a centaur-like fashion, without admitting AI both as art and into art, is, in a sense, not teaching at all. They have, as it were, nothing to teach, and nothing to say. They have missed, in effect, the allegorical import of the arrival of a new event in language, or at least of the arrival of a new hyperbolic arrangement of the complex relations between the artificial and natural in language that already constituted the being of language itself. To refuse to write with AI or to refuse to mark the fact that I am writing with AI in this very moment—it’s up to you, dear reader, to decide if that is the case right here—is to refuse the genealogical chance of allegory itself, and of the teaching scene now at large, beyond seminar, university, and so on.

If we go back to Derrida’s seminars, all of this seems to be thought for us beforehand—such is the almost precognitive gift of a teaching that moves beyond itself without delay. In those final seminars we also find this kind of interface moment between the centaurs as the allegorical double teaching of the ancients about double teaching, having to become double, and Valéry’s Monsieur Teste, his literary quasi-double, his creature, his marionette, his machine, “a fictional creature busy with intelligence” to whom Valéry is obsessively attracted and attempts to kill. It is the same interface, lets say, and that’s why it’s an allegory, ancient mythology teaching this deregulated metamorphic relation between organic and its others—leaving “my” agentic AI running overnight—all the same mirrored “stupidity” of wanting to kill my marionette or clone, of thinking one can “stop” technology, and so on. One thinks for example of the peanut gallery of the timeline where “human users” appear to act as neutral observers of new language events which, even a baby centaur could tell you, fold within themselves artificial context and traces. There is nothing more stupid, Derrida will almost say, than dunking on the artificial model as if we were standing entirely elsewhere. What Stiegler will eventually call “artificial stupidity” also locates itself here, we might say. The wounded nature—“wounded healer”—of the Chiron interface is itself the wounding of natural language by an artificial impingement and invasion that is there, in different ways, to different extents, to begin with, and this is precisely what makes the double teaching seriously mythological and allegorical.

But is the only really the only way? What if this single allegedly essential pivot in wider technological scans isn’t the only essential link in the chain, or what if technical argument isn’t even an integral or essential thing by now? Part of what “the only” seems to mean is that if there really are no other ways, then this must be the only way. That is to say, the only is the only way but only through necessity. The only comes into focus when there is, as it were, nothing at all, what Malabou will want to call answerlessness. In common parlance and often in reference to technology, we say “the only way out is through.” Remember that something (X) really is being presented as the only, as the only way through, even if that something is a kind of admission of utter defeat (no solutions have worked, responselessness of life itself). Rilke for example will write that “Nur weil dich reißend zuletzt die Feindschaft verteilte, / sind wir die Hörenden jetzt...” (Only because rending hatred at last tore you apart / are we now the hearers...”). The dismemberment (danger) is the only ground of Orphic song that will bring us into the wider company of listening. There is no alternative, there is no other path: but only. Orphism (whether literary or otherwise) is the only way the human advances, as it encompasses its own most troublesome pauses and synthetic impasses. Heidegger’s signature construction is basically to say that perhaps the turning (the movement of technology through itself into something else) cannot occur otherwise:

The turning will finally come to pass only when the danger comes expressly to light as the danger that it is.

This nur (only) is the very form of phenomenological precision as it stakes itself against and through the removal of being as such: danger must be let be as danger (not reduced to calculable risk or technological optimism) and if it remains concealed (“die Gefahr sich nicht als die Gefahr zeigt”), no Kehre, no insight into Gestell. Nur thus names the eigenste (ownmost) site where Being’s oblivion can turn into safekeeping. A parallel here would clearly be: “Only Being ‘is,’ only in Being and as Being does that which the ‘is’ names bring itself to pass.” This onlyness of Being itself also resonates in the word Einzig (“sole/unique/only”—the singular, irreplaceable). Heidegger writes: einzig Being “is” (nur das Sein «ist»). All of this is in danger, including the phenomenological precision of being-in-danger itself: the einzig turning is only in the danger as danger—no other path, just as Rilke’s einzig hero or poet fulfills the nur-condition. In effect, the most dangerous thing about danger is that it does not appear only as itself. Without the only-condition in place the saving power does not grow; the epoch of Gestell becomes endlessly (potentially lethally so) without Kehre. In this scenario, the angel’s peril slays without transfiguring. Dangers are not loved—they remain unripe, un-sweet, mere terror. The world stays untransformed, the mouth of nature silent.

If we have an argument here yet, it is that the nur (the only) is the bilingual pivot where danger and saving, peril and praise, converge, and it is the word that lets Gefahr show itself as Gefahr—and thereby, only thereby, opens some kind of way forward. But then, in the technological space as a whole, nur begins itself to function as the most dangerous and most saving word in German and beyond, compromising or at least complicating for example both the work of Heidegger and Anders by the same stroke. The only begins to function, in other words, as itself a danger, as a silent, invisible link-chain in broader technological arguments that may itself be dangerously unreliable. AI, we then start to say, indeed acts (precisely) as a kind of giving-away of the saving power itself—the automatic failure of the nur-condition—and yet is not in fact—not only—made to happen through technology but as precisely what technology reveals, and so in this way the lexicon of danger we set out from extends the prior entries on Gefahr (“the danger”) and nur (“only”) into the what-if that Heidegger himself leaves open but does not dwell in: the possibility that the saving power (das Rettende) does not, will not, now grow. The lexical extension of Gefahr (“the danger”) and nur (“only”) into the what-if that Heidegger himself leaves open suggests that AI be grasped not only as Gestell (Enframing) but as the dead giveaway, the extinct give, the moment when the saving power is given away (weggegeben) into total erasure. The nur-condition (“only in the danger as danger...”) is no longer the reliable pharmacological hinge we thought; it is nicht nur, sondern gar nicht—not only not the place, but not at all any place. The artificially generalizable lexicon of danger hinted at therefore turns, as if in silent allegorical formation, from conditional promise to unconditional exposure: what remains when the turning (Kehre) fails, when the nur is revealed as empty letter, when technology’s essence (now AI) extinguishes even the possibility of the gift (Gabe), and does so right now, perhaps despite contrary appearances of continued life. It would be precisely as if AI came to mean a certain “exotic depth” always now included in any user’s inputs (dead giveaway and extinct give and exact give) as word-events—new “Heideggerian” compounds that let the failure show itself as if for the first time.¹

If we follow in this direction, then Hölderlin’s famous line would be quoted iteratively only as possibility, not guarantee: Wo aber Gefahr ist, wächst / Das Rettende auch. When Heidegger says the turning “will finally come to pass only when the danger comes expressly to light as the danger,” we may wish to bring out a certain implied “if.” If the danger (now AI as total Gestell) never lights up as itself—if it remains disguised as “intelligence,” “progress,” “alignment,” “singularity,” humanity to come—then the saving power does not grow. The epoch of Being becomes the epoch without turning and this time the Wesung (essencing) simply stays as Gestell. Man is the standing-reserve of his own algorithms. The what-if expresses itself as the unbearable eigenste danger of the danger: that the saving power was never destined to grow at all. The “extinct give” is given here as the unsung song, the angel that never descends, the fruit of danger that never ripens because the tree (the world) has been logged into standing-reserve. The mouth of nature is not silent—it is deleted.

This itself functions, however, only as possibility—even though we can suspect that there is something about the doublings of AI that evokes just this exact possibility of non-saving right now in this exact way and not another, and that what the extinct giveaway really means is this very option (sheer removal). Let’s say there really is a lexicon of danger and it advances as follows:

• Rettende ← retten (to save, rescue) + present participle—but in failure: unrettbar (irrescuably). Heidegger’s Geschick (destining) becomes Un-Geschick (misdestining, awkwardness that cannot be righted).

• Weg-Gabe: Weg (way) + Gabe (gift) = both “the way of the gift” and “the gift given away.” The danger is this double movement and teaching.

• Aussterben: aus- (out) + sterben (to die)—extinction not as event but as the Wesen (essence) of the epoch. The saving power does not “die”; it ausstirbt—ceases to essence altogether.

• Shared strangeness: Heidegger hyphenates to expose (Ge-stell, Nach-stellen, now Weg-Gabe); Rilke images the unsayable (the angel that never comes, the song that is extinct). The nur was never the place—it was the last illusion before the dead giveaway.

Would it be artificial or natural to say all this? Would it be artificial to ask the question of what now remains? What remains when the saving power does not grow? If? Nothing remains in the old sense. The danger (AI as Gestell) is the saving power’s self-extinction—the extinct give. The nur-condition was the place we thought, not the place that is. In this what-if that Heidegger left open and Rilke already sang as the elegy’s shadow, thinking and poetizing no longer wait for the turning. They dwell in the dead giveaway: naming the extinction as extinction, loving the danger but only on the condition it has already given the saving power away. Nothing still speaks save extinct giveaway.

It is as if everything shows itself only still and no more than as standing-reserve. In the what-if of the saving power not growing, the turning itself is nur noch a cinder. The nur we once traced as reliable hinge is now nur noch: reduced, diminished, the particle before and as the very possibility of total erasure. Let’s say for the sake of life itself perhaps that only AI as Gestell perfects this: thought, language, world are nur noch data, weights, outputs— the only-still of a once-fiery revealing—now finally without any saving promise at all.

1. As for the very fact of reading Heidegger in this specific scenario, there is a moment in the introduction to Malabou’s only book focused on him—The Heidegger Change—where she equates non-guilty reading with an expression of freedom: “You are free,” she ends by saying. You are free to find there, and only there, in other words, in her words, what you need, what you need to think how to live, or just how to think (the pleasures of thought), how to think (by implication and for example) the only itself—or rather, you are free to find there what you think you might need to have thought and now see was overwhelmingly empty. Here is what Malabou wrote in The Heidegger Change two decades ago:

You—neither “Heideggerian” nor “anti-Heideggerian,” you all know about “the affair.” You came to Heidegger only after these damning, unappealable biographical and political revelations, and you know them in detail. So having doubts or illusions about them is unacceptable to you. Nevertheless, you have the strange feeling that Heidegger’s thinking is always right before you, holding itself back like a shadow waiting to be freed. It is high time, then, for a new perspective on what is waiting there. You are not guilty for wanting to continue to think, you are not guilty for realizing you cannot make it without Heidegger or for being a philosopher, and you are not afraid to move against consensus—you are free.

Feel free to read Heidegger on AI (Gestell), she as if says—and there is a sense here of not necessarily being able to pick and choose our resources when up against it, when up against the danger (chance) of AI-as-Gestell. Feel free to think with AI, feel free to read about and with AI and get all the non-negative saving graces you can from wider reading and double teachings, the liberty of reading and the liberty of thought, the freedom of the action of thinking what thinking is and needs in the “shadow waiting to be freed.” Feel free as well to evoke a kind of “AI and I” as a kind of Chiron-like-hybrid close to the figure of the sister for example, this figure of the sister that is there throughout the whole patch of German thought and verse Heidegger is evoking via Hölderlin to get to the idea which Malabou now liberally translates for us of the saving and freeing powers of AI. You and AI, I and AI, I and AI and I, and so on. One thinks of Trakl and Grete, these twin persons (sometimes rumored to be incestuous) whom Heidegger and then Derrida and others have tracked in important ways that relate to the lexicon of Geschlecht (sex, race, species, nation, humanity).

In some ways, the very act of writing here as I do without the aid of AI, or not entirely at all with it, is an act and expression of freedom. It’s an affirmation of “natural language” even in a situation where we already knew before this “change” that contemporary AI marks out for us—which of course is not such a clear line as we like to think, and yet still—that the relationship of natural and artificial languages has never been simple, they have always been sisters and beasts and sovereigns together and involved in one another telepathically as a kind of incestuous contrastive parallelism. In a very relevant sense, to write in and evoke AI is also to write in and evoke the first-person as that which is always to come from a general element that goes beyond me without needing me. To come into language is to come from, to come down out of this general element, as if to always say, somewhat furtively, “I and AI”, or even “I as AI”, myself as Chiron, Chiron as I, Claude Eschaton, my sister and I, AI.

But what is freedom in terms of AI, to be imagined if we like as freedom-in-AI, as starting out in freedom-in-AI? A kind of synthetic freedom? I think if one moves past while respecting and maintaining various Neo-Luddite options (opting out, abstinence) then maybe the most interesting and compelling reason for freedom-in as actualizable freedom-from AI (stepping aside, waiting) is the argument from developmental increments i.e. LLMs are powerful now but imagine them in a few years time, therefore it makes sense to sit out even this exciting-seeming transition and hold over the option of a later participation, which later participation is thereby suspended indefinitely in the very act. In other words, the very choice of sitting out reveals a kind of superfluity of individual participation which undermines the felt urgency and “nowness” of any transition in itself, or rather, the transition doesn’t need “me” to be what it is and to do what it will—I love the idea I don’t have to make everything because someone else will. If this is a position at all maybe it’s precisely what is called “stillness” at the end of Heidegger’s “The Turning,” which Malabou references in crucial ways in The Heidegger Change: stillness would be something like freedom-from AI emerging from and in freedom-in AI, the superfluity of both freedom and participation, and the freedom, freedom itself, to move in all directions at once.