.png)
This essay enters the strange territory where AI, art, and human feeling collide, asking what changes when language moves us but no living voice seems to speak through it.
An AI cannot grieve. Grant that objection in its strongest form. Assume there is nothing inside the system that remembers a dead person’s handwriting, hesitates before deleting a phone number, or discovers months after a funeral that some ordinary household task has become unbearable. Assume, in other words, that there is no inner experience behind its language. This is not a scientific conclusion about machine consciousness. It is a premise for a narrower question: if we grant it, what follows?
Consider two invented sentences: The kettle clicked off. There was still water for two. Perhaps nothing happens. Perhaps the small mechanical sound opens onto a kitchen in which an ordinary calculation has outlived its purpose. The sentences do not specify a death. They do not even name loneliness. Yet a reader can supply an absence, hear the disproportion between a finished action and an unfinished attachment, and find a private memory reorganized by that modest surplus of water.
Now suppose a generative system supplied those sentences. Which part of the possible encounter has thereby become impossible? The system need not have boiled a kettle. It need not have lost anyone. But why should the absence of experience at the point of production entail its absence at the point of reception? Grief is not a substance that must first be deposited in a poem before it can be released into someone else.
The rejection of AI art on grounds of absent consciousness makes an unsupported crossing: from a claim about the producer to a verdict about everything its products can become. The easiest defense makes a different crossing. Because an output can move someone, it declares the output equivalent in every important respect to human work. One position abolishes the encounter; the other allows the encounter to abolish the conditions of its making. Both compress several kinds of significance into one certificate marked “real art.”
That certificate should be withdrawn. An artifact can organize a genuine aesthetic experience without originating in a corresponding experience of its immediate generator. It does not follow that the artifact expresses an actual person’s grief, constitutes testimony, embodies comparable achievement, or was produced justly. Aesthetic force and expressive origin are different forms of significance. Neither automatically cancels the other. Generative AI matters here as a philosophical stress test, not as a candidate for honorary personhood.
The necessary distinction is simple: production → artifact → reception. Production brings a configuration into being. The artifact is that available configuration: these words in this order, this distribution of light, this phrase entering after that silence. Reception is the situated activity through which its possibilities become experienced. The arrow does not represent a fluid called meaning passing intact between containers. It marks relations among different activities and conditions.
A poem on an unopened page does not cease to be an artwork. Its history and organization persist, along with possibilities no reader may ever realize. Nor does an unsuccessful performance cease to be art because its audience was bored. The distinction concerns aesthetic actualization, not a rule for making objects disappear. Making produces an artifact; an encounter gives its qualities an experienced life. The consciousness of the author is not the necessary location of that life.
The distinction is functional rather than a rigid chronology. A writer rereading a sentence is already its recipient; a composer listening to a developing phrase moves between making and hearing. Reception can send production back to work. Nor is “the audience” necessarily a solitary consumer. Performers, collaborators and successive readers can participate in different realizations. To distinguish production from reception is not to deny their traffic. It is to stop mistaking their connection for identity.
There is no warrant, either, for imagining synthetic production as humanity’s total evacuation. In “On the Dangers of Stochastic Parrots” (2021), Emily M. Bender and her coauthors describe language models trained on text produced within human social practices. Such systems do not acquire language from an uninhabited universe. Their human antecedents matter. So do choices about prompts, selection, revision and presentation. The absence of a single experiencing speaker inside the generator is not the absence of people from its causal history. ([Bender et al.][1])
But distributed human causation does not conveniently add up to a person who underwent the experience represented. A corpus has no collective bereavement that can be attributed to every generated elegy. Conversely, a person who selects and transforms generated material may make a consequential artistic intervention. These possibilities require attention to what was actually done. Neither “AI made it” nor “a human prompted it” describes enough to settle authorship, much less every question of value.
A familiar escape from this difficulty is to observe that novelists invent people and actors portray emotions they are not presently feeling. That defeats autobiographical literalism, but not the stronger objection. An invented speaker can still be deliberately imagined by someone who understands the act of speaking. Unknown authorship is likewise different from absent authorship. A poem without a name attached may remain someone’s address. The challenge posed by generation cannot be answered merely by reminding us that fiction is fictional.
It can, however, force a reconsideration of where artistic experience has been located. W. K. Wimsatt and Monroe C. Beardsley’s essays collected in The Verbal Icon (1954) attack two forms of misplaced critical authority. The “intentional fallacy” makes a maker’s intended result the standard of the poem’s success. They do not deny a designing intelligence caused the poem. They deny that recovering its intention supplies the governing criterion. Their “affective fallacy” rejects the corresponding confusion between the poem and its psychological effects: what it does to a particular reader cannot simply substitute for an account of what it is. ([Internet Archive][2])
The New Critical defense of a publicly examinable object remains indispensable. “It moved me” cannot settle a disagreement about how a stanza works. Yet, as an account of aesthetic life, the relocation stops too early. A pattern can persist without anyone perceiving it; suspense cannot occur without expectation. The poem’s organization and the activity it organizes are distinguishable. Protecting the first does not require treating the second as an embarrassment to criticism.
Roland Barthes’s “The Death of the Author” attacks another sovereignty: the use of an originating person to close down the plurality of a text. In Richard Howard’s translation, “the unity of a text is not in its origin, it is in its destination.” That destination is not straightforwardly the individual psychology of whoever happens to read; Barthes’s reader is a largely abstract gathering place for cultural language. His argument helps remove the author’s final veto over meaning. It does not abolish production, and it does not by itself explain the particular person whose attention changes while reading. ([Penn Writing Center][3])
John Dewey supplies a more experiential emphasis—and a warning against using him too cleanly. In Art as Experience (1934), he distinguishes the art product from “the actual work of art,” which is “what the product does with and in experience.” The distinction relocates the question. Instead of asking where the completed artistic substance is stored, it asks how an encounter develops. A painting’s arrangement is not interchangeable with the experience of looking at it, even though the experience depends upon what the arrangement makes available.
But Dewey refuses a final separation of making and appreciation. He imagines an object admired as finely wrought and then discovered to be an accidental natural product; the discovery alters appreciative perception even though the thing confronting the viewer has not physically changed. Production history can therefore enter the experience itself. Dewey helps relocate aesthetic life toward encounter without licensing indifference to provenance. ([Art as Experience][4])
Louise Rosenblatt gives the relation a particularly useful name: transaction. In Writing and Reading: The Transactional Theory (1988), meaning is not something already finished inside either the text or the reader. It happens through the encounter of a particular person, a particular text and a particular situation. The reader brings linguistic habits, memories and expectations; the text activates and constrains them. Reading adjusts itself as later words revise what earlier words seemed to promise. Her distinction between aesthetic and efferent reading concerns attention: living through the qualities of an experience, or concentrating on what can be carried away and used afterward. It is not a division between inherently artistic books and other books. Her reader inhabits both a shared language and an individual history. Nor does transaction authorize arbitrary interpretation. Textual evidence and shared criteria still matter. ([Rosenblatt][5])
Transaction is more demanding than the sentimental claim that everyone has a personal truth. An encounter involves something that can resist you. The reader who anticipates reassurance may meet a rhythm that refuses to settle, an image that complicates the familiar mood, a final sentence that makes the preceding tenderness look cruel. Participation includes being corrected. The artifact is not raw material over which an audience exercises unlimited sovereignty. Nor does a transaction require agreement. A reader can resist an invitation to mourn, find the expected tenderness oppressive, or judge a supposedly tragic detail merely convenient. Such resistance still answers to an arrangement. The right to respond differently does not confer a right to describe the artifact inaccurately.
Wolfgang Iser’s reader-response account in “The Reading Process: A Phenomenological Approach” similarly distinguishes an artistic pole, the author’s text, from an aesthetic pole, its realization by a reader. Readers connect what is given, anticipate what is coming and work across what remains unwritten. The written material limits those activities. Iser and Rosenblatt do not thereby establish that human authors are dispensable. Applying their accounts to generated text requires a further argument: can a differently produced artifact still supply structured opportunities for such realization? ([Iser][6])
Nothing in the description of that encounter makes corresponding experience in the generator a necessary condition. A recipient must be capable of experience; the artifact must offer qualities and relations to attention; the circumstances must allow that attention to develop. But it does not follow that the process producing the artifact must itself have lived what becomes vivid in reception. To deny this possibility requires an additional premise, not another declaration that the machine feels nothing.
Consider the stronger necessity claim through a thought experiment. Two different production processes yield the same words, presented in the same way to a reader who knows neither history. To say that the synthetic version cannot occasion a corresponding felt response is to give an inaccessible fact about production a power unexplained by anything available in the encounter. This does not make the histories interchangeable. One artifact might be testimony and the other fabrication; discovering that difference could matter enormously. But a difference in what the artifacts are does not by itself establish that the reader encountering the synthetic version felt nothing or attended to no form. These judgments require different grounds.
But if experience is where aesthetic life becomes actual, why stop at poems? Sunsets, birdsong, stains and accidental noises can receive aesthetic attention. A piece of terrible news can produce grief more reliably than a sonnet. Has the artwork become whatever manages to cause a feeling?
First, an emotion is not necessarily an aesthetic event. Fear at an approaching vehicle concerns an urgent practical danger; attention to the pacing, sound and spatial arrangement of a frightening scene concerns how an experience is organized. These activities can overlap, and neither pleasure nor disinterest is a compulsory test. What matters here is that aesthetic attention engages qualities and relations in their experienced unfolding. “I felt something” is too indiscriminate to describe it.
Second, aesthetic experience is not by itself a sufficient definition of an artwork. A sunset can be aesthetically compelling without becoming a painting. An object’s place in an artistic practice, its history of making or selection, and the kind of attention for which it is presented may all matter to its classification. An account of where aesthetic possibilities become actual does not automatically answer every question about which objects deserve which names. To insist otherwise would confuse the event with the category, repeating the very mistake this argument opposes.
That concession does not return authority to the maker’s consciousness. It changes the claim being tested. Perhaps an artwork must have an appropriate relation to intentional activity or an artistic practice. It still does not follow that its maker must have felt the emotion its recipient experiences. Nor does classifying an artifact as art guarantee that anyone will find it aesthetically compelling. Admission to the category and life within an encounter are different achievements.
The artifact’s role now becomes decisive. Change the invented sentence from “There was still water for two” to “There was barely water for two.” The situation tilts from persistence toward insufficiency. This does not command a particular feeling: there might be drought, poverty, comedy or mere inconvenience. It changes the available paths. The reader brings associations, but the words help determine which associations can be sustained. An interpretation of abundance would now face resistance it did not face before. Neither sentence specifies death. Someone may simply have left the room. A responsible interpretation must acknowledge that openness, just as a critic should distinguish what a work makes possible from what it establishes. This is where craft remains consequential even when production is partly automated. Differences of timing, syntax, relation and omission can change what attention discovers. A transactional account does not depreciate the artifact. It explains why its particulars matter without imagining that they experience themselves.
We keep looking for the ghost in the machine because we have failed to notice that the ghost is in the reader.
The empirical question is whether synthetic artifacts can, under actual conditions, support more than this conceptual possibility. Brian Porter and Edouard Machery’s 2024 study in Scientific Reports tested readers’ responses to human-written and AI-generated English-language poetry. In one experiment, 1,634 United States adults made 16,340 source judgments and identified provenance with 46.6 percent accuracy, against a 50 percent chance baseline. A separate experiment with 696 participants assessed ten poems across fourteen rating dimensions while actual source was crossed with human, AI or undisclosed source framing. AI-generated poems received favorable evaluations, in some analyses exceeding human poems on several dimensions, while beliefs about human authorship also improved some judgments. These results concern a restricted set of poems and immediate nonexpert assessments. They do not establish superior poetry in general, expert endorsement or lasting transformation. ([Scientific Reports][7])
The counterevidence is important precisely because the argument does not require a machine victory. In a study in Computers in Human Behavior (2021), Nils Köbis and Luca D. Mossink compared human poems with GPT-2 continuations of human-written opening lines. Their 384 preference judges chose human poems in 64.9 percent of 3,840 paired comparisons. Human selection also made generated outputs harder to identify, underscoring how much curation can matter. Disclosing the source did not produce a statistically significant additional preference for human work in that experiment. Different systems, corpora and comparison tasks yielded different results; these studies are not successive rounds of one poetry championship. ([Köbis and Mossink][8])
What follows is narrower than the promotional claim and more disruptive than the categorical dismissal. Favorable aesthetic judgments of generated verse are possible, but neither preference nor a rating of “moving” measures the whole of an artistic encounter. A numerical assessment cannot stand in for the poem someone returns to after ten years. Yet a critic cannot preserve the necessity of an originating emotional life simply by announcing that every favorable response to synthetic work must be mistaken. That would make the proposed condition immune to the very experience it claims to explain. These are not tests of machine consciousness, or of production free from every human decision. They test encounters with particular outputs. Those limits block claims of equivalence; they do not restore a prohibition on reception.
An easy thrill and a difficult recognition are both events, not thereby equal accomplishments. A work that disturbs, withholds or exhausts may offer more than one that immediately pleases. A critic defending it must show how its particulars sustain that claim. Positive ratings can establish a reception worth explaining; they cannot confer immunity from judgment. A reader’s report of feeling deserves to be taken seriously, but it does not automatically establish the quality of its occasion. Aesthetic experience remains subject to criticism precisely because it involves more than registering that a stimulus worked.
Those studies establish a reception problem. They do not establish an addresser.
Steven Knapp and Walter Benn Michaels, returning to their wave-poem thought experiment in a 2023 Critical Inquiry forum, distinguish words used by an agent from marks that merely resemble such an utterance. Imagine surf leaving what looks like a poem in the sand. If the marks are accidental, what actual meaning has been communicated? They acknowledge that language models are not random surf: their outputs reproduce patterns of human writing. Their central challenge concerns an act, not merely a hidden feeling. Identical words can be praise or irony depending on what someone is doing with them. A reader’s attribution of an imaginary speaker does not create an actual speaker who performed that act. ([Knapp and Michaels][9])
This objection identifies a genuine loss. An apology produced without an apologizing agent may calm its recipient without anyone having apologized. A text resembling a witness’s account may evoke compassion without anyone having witnessed what it describes. No intensity of response can manufacture the missing action. If the value sought is another person’s acknowledgment, testimony or deliberate revelation, the substitution may fail completely. Being affected is not the same as having been addressed.
But that is a reason to distinguish kinds of meaning, not to declare every other significance void. An intelligible configuration of language, a warranted account of a speaker’s action, and a reader’s recognition of something in their own life are different achievements. The first can organize the third without establishing the second. The interpretation that “someone is confessing to me” may be false while the attention to shame, evasion or self-exposure occasioned by the text remains real. The loss of an addresser limits the encounter; it need not reduce the encounter to nothing.
An informed reader need not restore the absent speaker through self-deception. It is possible to attend to an openly generated lament as a configuration that makes sorrow perceptible, without supposing that sorrow is being confessed. Here “expressive” has two meanings that should not be smuggled into one another. A passage can have a sorrowful character; it can also be someone’s expression of sorrow. The second describes a relation to an agent that the first need not establish. Losing that relation can be a serious deprivation. But judging how a passage gives form to sorrow is not answered by establishing that no one suffered while generating it.
The reverse complication is equally important. Suppose a bereaved person chooses generated sentences because they say what that person wishes to communicate, revises them, and offers them as an avowed response to an actual loss. There is now an intentional act whose significance cannot be settled by identifying which words the software supplied. This does not automatically establish skill, sincerity, or entitlement to credit for every feature. It shows why the actual use matters. A human can take responsibility for generated language; a machine label cannot tell us whether that has happened. The relevant distinction runs through practices, not neatly between two species of artifact.
The demand for human expression is strongest when it names what it values. To encounter a person’s attempt to give form to something they barely understand can matter beyond the resulting pattern’s immediate effects. An awkward elegy might be precious as an act of attention, fidelity or struggle. A technically smoother substitute might lack precisely what its reader seeks. This preference requires no fantasy that machines must be secretly conscious. It requires only the recognition that artistic significance can include a relation to someone’s activity, rather than an interchangeable supply of emotional results.
Once that distinction is granted, provenance returns inside aesthetic experience, not outside it as an administrative label. Dewey had already built the trouble into his own account: learning that an admired object was accidental could alter appreciation without altering the object’s physical arrangement. Contemporary labeling experiments give that difficulty a more controlled form.
In “Humans versus AI” (2023), Lucas Bellaiche and colleagues presented selected AI-generated images while randomly assigning human or AI authorship labels. Images labeled human received more favorable judgments on several dimensions, including liking, beauty and profundity. The manipulation changed attributed origin, not the image being viewed. This does not show that viewers were foolish. It suggests that viewing can include an understanding of what sort of act the image represents. The visible arrangement does not exhaust the circumstances of its reception. ([Bellaiche et al.][10])
Nor do labels exert a uniform power. Nicolas E. Neef and colleagues’ study of German online participants, published online in AI & Society in 2024, found no significant difference in aesthetic value between separate groups seeing AI-generated images labeled entirely human or entirely AI. Human-label advantages were more apparent in mixed-label comparisons. Together with the contrary poetry disclosure result, this cautions against a simple law of automatic devaluation. What people are comparing and what they think they are being asked to judge matter. ([Neef et al.][11])
Music allows a different provenance test: whether an attributed source changes the listener’s willingness to construct an imagined world. Sarah H. Wu and Kevin J. Holmes’s 2026 study, “Is there a ‘mind’ behind the music?”, assigned human-composed and AI-generated piano clips either a human or an AI label. In its main experiment with 300 participants, listeners reported imagining stories on an average of 50 percent of AI-labeled trials, compared with 57 percent of human-labeled trials. The actual-source effect on this measure was not statistically significant. The finding is neither equivalence nor extinction: attribution altered one measured activity without eliminating it. These were self-reports about short clips, not an inventory of everything music can mean. Still, the design captures an important distinction. A source label can change the likelihood of imaginative participation while leaving participation available. ([Wu and Holmes][12])
Imagine a piano phrase approaching a resting place. One note hangs after the others have died away; then the phrase begins again. For one listener, that return might resemble reluctance; for another, insistence. No bereaved composer must be inferred for the movement to become an image of refusing to leave. But a listener who hears a deliberate human gesture may inhabit the phrase differently from one who hears a generated sequence. Reception survives as a possibility, not as an invariant. The machine’s missing inwardness and the listener’s available inwardness do not cancel each other out.
Blind encounters can isolate an effect; blindness is not consequently the ideal condition of criticism. Knowing that an image records an event rather than invents it may change what there is to attend to. Knowing that a passage is a deliberate quotation can change its force. Where such facts matter, the more informed experience is not automatically less authentic because it differs from the first. The history of an artifact can become part of what is perceptible in it.
The labeling experiments do not, however, establish what happens when an already invested reader discovers a mistaken attribution. Learning a source afterward is a different situation. Imagine returning to the kettle sentences after learning that the supposed elegist and the death behind the poem never existed. The sense of communion might collapse. The language might become a specimen of manipulated intimacy. It might remain moving in another way. None of these outcomes should be guaranteed in advance.
What cannot be changed retrospectively is that an earlier experience occurred. Its interpretation may have been mistaken; its beliefs may have been false; its emotional force may now be unavailable. Genuine does not mean informed, justified or immune to revision. Yet discovering an absent mourner does not prove that the reader’s sorrow was absent. It tells us that the sorrow did not have the relationship to an originating person that the reader had supposed.
Baudrillard names a further disturbance—not as an oracle of artificial intelligence, but as an analyst of what happens when signs loosen from an independently authoritative original. In “Simulacra and Simulations,” the hyperreal is not simply a convincing fake: models and circulating signs help establish what will count as the real rather than merely copying a reality already secured elsewhere. Applied to synthetic expression, this vocabulary describes a possibility: recognizable signs of intimacy or profundity can operate without a corresponding intimate or profound experience in their immediate generator. This is an extension of his framework, not an empirical finding about every generated work. ([Baudrillard][13])
A vacant chair, the stopped watch, the second cup: grief has legible devices. In a poem, they can organize an encounter with loss. They can also advertise “loss” so efficiently that recognition arrives before attention has done much work. A generated artifact may assemble such signals without any mourner behind them. The unsettling consequence is not that mourning becomes unreal. It is that the appearance of mourning and the experience of encountering it can function without reciprocal testimony.
The danger is therefore more interesting than “nothing is real.” Simulated signs can occasion genuine responses. There may be no rememberer in the generating process, yet an artifact can call up an actual memory. There may be no desire seeking expression, yet a musical phrase can become inseparable from someone’s desire. The simulation belongs to one relation—the signs’ apparent connection to a producing interior—and reality to another—their occurrence within a recipient’s life.
Calling both relations fake obscures the event. Calling both authentic invents an origin.
A further cultural risk follows, though it should not be mistaken for an established condition. Highly recognizable signs of depth could become the standard against which depth itself is judged. An untidy grief that will not arrange itself into a beautiful ending might then appear less persuasive than a fluent synthetic elegy. This is not an inevitable property of AI: a human poem can be conventional, and generated material need not be reassuring or easy. The concern is a criterion of recognition that mistakes familiarity with the signs of experience for knowledge of another life.
The ethical consequence is not permission. A work’s power cannot settle whether those involved in its production were treated fairly, whether attribution was deceptive, or what obligations attach to its sources. Where consent is owed, applause cannot supply it. Conversely, condemnation of a production practice does not establish that no aesthetic experience occurred. Ethical knowledge can transform attention and judgment; aesthetic efficacy cannot grant ethical acquittal. These categories interact because they are different, not because one is a euphemism for the other.
A critic or editor can therefore value human work for reasons a reception experiment does not measure: sustaining a practice, recognizing achievement, encountering other people, preserving occasions for deliberate expression. There is no need to discredit every synthetic encounter to defend those commitments. Equally, acknowledging that generated work can be aesthetically forceful does not oblige anyone to treat it as interchangeable with human testimony or to welcome its production under any conditions.
The useful judgment has two movements. Attend to what the artifact enables: what its form makes perceptible, what its repetitions disclose, where it resists easy recognition. Then ask what sort of act and history the encounter involves, whose agency deserves acknowledgment, and which claims about its origin are true. Neither inquiry should be used to counterfeit the result of the other. “Can AI make real art?” is too compressed a question to perform this work. Before answering it, criticism must decide where art happens and which of its meanings is at stake.
Even the little kitchen in these pages did not arrive without circumstances. It was placed in an argument about grief. Its framing helped make an absence available; the word “still” supplied an invitation, not an order. The reader did not make everything, and the artifact did not make its own experience. The ghost in the reader is no more sovereign than the author it has displaced.
What remains is a relation without a guaranteed reciprocity. An artifact may fail to bring us into contact with another person and still alter our contact with ourselves. Something important may be missing. Something important may nevertheless happen. The machine does not have to become human for this to be possible. The unsettling question is what humans will make of an encounter in which the feeling is theirs, the form has power, and there may be no one on the other side.
[1]: https://s10251.pcdn.co/pdf/2021-bender-parrots.pdf "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"
[2]: https://archive.org/stream/in.ernet.dli.2015.101371/2015.101371.The-Verbal-Icon_djvu.txt "The Verbal Icon"
[3]: https://writing.upenn.edu/~taransky/Barthes.pdf "The Death of the Author"
[4]: https://www.uio.no/studier/emner/uv/uv/UV9406/dewey-john-%281987%29.-art-as-experience.pdf "Art as Experience"
[5]: https://files.eric.ed.gov/fulltext/ED292062.pdf "Writing and Reading: The Transactional Theory"
[6]: https://www.philol.msu.ru/~discours/images/stories/article1.pdf "The Reading Process: A Phenomenological Approach"
[7]: https://www.nature.com/articles/s41598-024-76900-1 "AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably"
[8]: https://arxiv.org/pdf/2005.09980 "Artificial intelligence versus Maya Angelou"
[9]: https://critinq.wordpress.com/2023/06/30/here-is-a-wave-poem-that-i-wrote-i-hope-you-like-it/ "Here Is a Wave Poem that I Wrote . . . I Hope You Like It!"
[10]: https://link.springer.com/article/10.1186/s41235-023-00499-6 "Humans versus AI"
[11]: https://link.springer.com/article/10.1007/s00146-024-02020-z "Drawing the full picture on diverging findings"
[12]: https://link.springer.com/article/10.1186/s41235-026-00715-z "Is there a ‘mind’ behind the music?"
[13]: https://web.stanford.edu/class/history34q/readings/Baudrillard/Baudrillard_Simulacra.html "Simulacra and Simulations"