{"id":"8f37d73e-057f-46ee-bc6a-4e84f6ede2ae","arxiv_id":"1908.01133","paper_version":1,"verdict":"UNVERDICTED","confidence":"UNKNOWN","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AI-assisted art is best understood as human-human collaboration mediated by machines that act as surrogates for the toolmakers' agency.","lead":"This paper argues that when artists use AI tools, they are not truly collaborating with the machine itself, but with the human designers, developers, and data curators whose choices are embedded in the tool. The authors introduce the term 'machinic surrogate' to describe AI systems that act as proxies for their creators' creative agency in human-human collaborations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never bridges its own admission that artistic collaboration is 'not directly applicable' to machines, yet the central claim depends on exactly that application.","rationale":"The reader's weakest assumption identifies exactly this gap: the transfer of third-identity collaboration to human-machine interaction is asserted, not argued. I agree, and I frame it as an internal coherence problem rather than an external verification problem. The paper itself flags that the model is 'not directly applicable' and then proceeds as if it were, without an account of how a machinic surrogate can sustain mutual recognition or emergent identity. The three cases are described in terms of influence and constraint, which support a weaker claim about mediation and distributed agency, but not the strong claim that users collaborate with toolmakers through the tool. This matters because the paper's novelty lies in that strong claim. The paper has real strengths: it draws on concrete recent projects, acknowledges toolmakers' agency, and proposes a useful vocabulary. The concern is that the vocabulary outruns the demonstration. I recommend CONDITIONAL: the paper should be accepted only if the authors either (a) provide a substantive theory of how identity can be instantiated in a machinic surrogate and show that the third-identity features appear in at least one case, or (b) revise the central claim to the weaker, well-supported thesis that AI tools mediate and distribute human agencies in creative work.","tokens_in":9032,"tokens_out":7017,"duration_ms":72214,"concrete_test":"Using the descriptions in the three case studies, code each project for the five features of third-identity collaboration: (1) mutual acknowledgment between parties, (2) shared exploration over time, (3) complementary contribution, (4) erosion of initial identity diversity, (5) emergence of a new identity. If none of the three projects exhibits all five features, the extension of the collaboration model to machinic surrogates is unsupported; if at least one does, the concern is mitigated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—co-creation with AI is collaboration between human agents mediated by 'machinic surrogates'—rests on applying the 'third identity' theory of artistic collaboration (Green; Lévi-Strauss) to human-machine interaction. In the section 'Human-machinic surrogate collaboration', the authors write that artistic collaboration 'is not directly applicable' to human agents and CC tools, then attempt to bridge this by giving the tool a 'derived identity' from its human inspirers. This is the load-bearing step: the paper never specifies how a derived identity can participate in the reciprocal, identity-eroding process that makes collaboration yield a third identity. Embodying choices and constraints is not the same as bearing an identity capable of mutual recognition. The three case studies show toolmakers' influence (DeepCloud's confined design space, My Artificial Muse's pre-chosen model chain, Learning to See's data feedback loop), but not a new collaborative identity emerging between user and toolmakers. Thus the conclusion overstates what the evidence and argument support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that co-creation with AI-enabled tools in art and design is better understood not as a partnership between human and machine agents but as a form of human-human collaboration mediated by technical artifacts. It introduces the term \"machinic surrogate\" to describe how algorithms and datasets carry the decisions, skills, and biases of toolmakers into a creative interaction. The argument proceeds by reviewing computational creativity and co-creation discourses, drawing on late-twentieth-century theories of artistic collaboration (Green, Lévi-Strauss), and presenting three case studies—DeepCloud, My Artificial Muse, and Learning to See (Hello, World!)—as instances of machinic surrogacy. A fourth example, the Obvious team's Portrait of Edmond Belamy, is used to trace a chain of surrogacies among developers, artists, curators, and audiences. The conclusion extends the frame to future AI systems and notes the limitation that the human contribution may decay in algorithms that generate algorithms.","tokens_in":9259,"tokens_out":3905,"duration_ms":41890,"significance":"If the central claim holds, the paper makes a useful critical intervention in the HCI and computational creativity literatures: it shifts attention from the supposed autonomy of creative machines to the socio-technical networks of human agency embedded in them. The three case studies are concrete, relevant, and clearly described, and the proposed term \"machinic surrogate\" gives designers and researchers a vocabulary for discussing authorship and agency that avoids both naive tool-instrumentalism and exaggerated machine-autonomy claims. The paper also offers a valuable discussion of roles (toolmaker, artist-user, curator, audience) and of the authorship and ownership debates surrounding AI art. Its main weakness is that the central theoretical step—applying the artistic-collaboration model to human-machine interaction—is asserted rather than fully argued, and the case studies are selected to fit the frame without any negative or contrasting cases. The contribution is therefore significant as a conceptual provocation, but its general conclusion is not yet supported as stated.","major_comments":[{"comment":"The paper's central claim is that co-creation with AI is human-human collaboration mediated by \"machinic surrogates,\" but this section explicitly states that artistic collaboration \"is not directly applicable\" to human agents and CC tools, and the subsequent bridge—asserting that tools carry a \"derived identity\" from their toolmakers—is not developed into a mechanism. Specifically, the manuscript does not specify how a derived identity can participate in the reciprocal, identity-eroding process by which Green and Lévi-Strauss define the emergence of a \"third identity,\" nor does it offer observable criteria for when a tool's embodied choices amount to such an identity rather than to mere constraint. Because the conclusion that \"co-creation is a special case of collaboration\" rests on this step, the argument is under-specified at its load-bearing point.","section":"Human-machinic surrogate collaboration"},{"comment":"All three case studies (DeepCloud, My Artificial Muse, and Learning to See) are presented as instances of machinic surrogacy, but no negative or contrasting case is considered, such as a generative system whose outputs are not plausibly traceable to toolmakers' decisions or an interaction where no human-human mediation occurs. The selection therefore shows that the frame can be applied to these examples, but not that it explains the range of human-machine co-creation; the paper should state the selection criteria and discuss at least one disconfirming case to support the general claim made in the conclusion.","section":"Machinic Surrogacy in Practice"},{"comment":"The manuscript acknowledges that outputs are \"often not predictable by the human partner\" (Conclusion) but does not analyze whether this unpredictability constitutes a separate agency that would break the surrogacy relation. The strong claim in the conclusion that \"it is the agency of the authors and toolmakers which is crystallized in the tool\" is not defended against the alternative explanation that a trained model's emergent behavior contributes something not reducible to the toolmakers' intentions; the paper should specify a criterion for distinguishing surrogate agency from emergent machine behavior and apply that criterion to the case studies.","section":"Authorship and Ownership"}],"minor_comments":[{"comment":"There is a grammatical typo in the sentence beginning \"These tools are the outcomes of joint efforts by an assembly of human agents, constitutes of\"—\"constitutes\" should read \"consisting\" or \"composed.\"","section":"Computational Creativity"},{"comment":"In the Obvious paragraph, the phrase \"various hum in action\" appears to be a typographical error; it likely should read \"various humans in action.\"","section":"Humans' Role in Collaboration with Machinic Surrogates"},{"comment":"The Sketch RNN discussion contains the phrase \"the audiences who were interactions through an online game\" which should be corrected to something like \"audiences who interacted through an online game.\"","section":"Machinic Surrogacy in Practice"},{"comment":"Reference [10] is incomplete (missing author names and article title), [25] misspells \"Kingma\" as \"Knigma,\" and [31] lists the authors as \"D. E. David Ha\" though the correct form is \"David Ha and Douglas Eck.\"","section":"References"},{"comment":"The term \"machinic surrogate\" is central to the paper but is never given an explicit one-sentence definition; adding a concise definition near its first use would improve clarity.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies on references [3] and [8], both authored by the third author, for the foundational premise that technologies enact human intent. This is not by itself grounds for rejection, but the authors should disclose this relationship and ideally strengthen the premise with additional independent support. The paper is a conceptual and interpretive contribution; if the journal's expectations include empirical validation or systematic case selection, the editor should be aware that the case studies are illustrative rather than a representative sample."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on 1908.01133. The genuinely new thing is the term 'machinic surrogate' and the framing that AI-enabled creative tools are not autonomous partners but proxies carrying the decisions and identity of their makers, so 'co-creation' is actually human-human collaboration routed through an artifact. That reframe is useful. It gives a vocabulary for the invisible human labor behind tools like DeepCloud, My Artificial Muse, and Learning to See, and the case descriptions are concrete and readable. The role taxonomy (toolmaker, user, curator, audience) and the Belamy/Barrat chain of surrogacies are the strongest sections; they make a real point that off-the-shelf generative tools are not authorless.\n\nThe soft spot is exactly where the stress-test puts it. The paper imports the 'third identity' theory of artistic collaboration from Green and Levi-Strauss to argue that co-creation with a tool is collaboration with toolmakers. It even admits the theory is 'not directly applicable' to human-machine relationships, then bridges that gap by giving the tool a 'derived identity' from its makers. But nothing shows how a derived identity can enter the reciprocal, identity-eroding exchange that yields a third artistic identity. The case studies show users facing the maker's choices - confined design space, preselected model chains, curated datasets. They do not show a new collaborative identity emerging between user and toolmakers. That is a load-bearing gap, not a nitpick, and the conclusion overstates by presenting these cases as confirming the full frame.\n\nThe self-citation point is minor. The foundational claim that technologies enact human intent comes from Cardoso Llach's prior work; leaning on a co-author's earlier book for a supporting premise is normal academic practice, and the synthetic argument here is new. It is not circular in any damaging sense.\n\nFor whom: people working on co-creation, computational creativity, and STS-of-AI who want a vocabulary for distributed authorship. It deserves a serious referee. My recommendation: send it to peer review, with a clear request that the authors either substantiate the third-identity derivation - what would count as a machinic third identity, and how would you recognize one in practice - or drop it and keep the more defensible claim that tools are surrogates for maker intent.","headline":"A useful 'machinic surrogates' vocabulary for distributed authorship in AI art, but the load-bearing claim that third-identity collaboration applies to human-tool interaction is asserted, not shown.","tokens_in":9730,"tokens_out":2315,"would_cite":true,"duration_ms":24232,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that 'co-creation' with AI is actually collaboration between human agents carried out through tools called 'machinic surrogates.'","keywords":["machinic surrogate","computational creativity","co-creation","human-machine interaction","AI-enabled creative tools","authorship in AI art","artistic collaboration","machine learning art"],"falsifier":"Conduct a controlled comparison of two versions of the same generative design tool, identical in interface, whose models are trained by different toolmakers on differently curated datasets; if the output distributions do not differ in ways traceable to those curation choices, the claim that users are collaborating with the toolmakers loses its observable support.","tokens_in":8868,"feed_emoji":"🎨","tokens_out":8166,"duration_ms":77815,"temperature":0.7,"pith_summary":"This paper argues that the familiar picture of human-AI co-creation—two partners, one human and one machine, jointly making art—is the wrong picture. Drawing on late-twentieth-century theories of artistic collaboration, it proposes that an AI tool acts as a 'machinic surrogate' carrying the identity and decisions of its human makers into the creative encounter. What looks like human-machine collaboration is therefore actually human-human collaboration, routed through algorithms and data. The authors support this with three recent generative-art projects and with a chain of toolmakers behind a well-known auctioned GAN portrait. The stakes: if true, it changes how we assign authorship, credit, and responsibility in computational art and design.","feed_headline":"AI co-creation is human collaboration via a surrogate","feed_subtitle":"Generative tools carry their makers' choices, so 'machine-made' art is a human exchange.","key_machinery":"The central object is the 'machinic surrogate': an AI/ML tool that acts as a proxy for its human makers, transmitting their identity and decisions into the creative process while producing novel, sometimes unpredictable outcomes. The paper adapts the late-twentieth-century theory that artistic collaboration forges a 'third identity' that exceeds the individual artists, and applies it to human-tool relations by locating the tool's identity not in the machine or algorithm but in its human makers. It also situates AI-enabled tools on a spectrum between automation and autonomy, and distinguishes roles—toolmaker, user, curator, audience—that human agents adopt across a project's life cycle. The three case studies show variations of surrogacy: a surrogate for absent toolmakers, a surrogate for a present audience, and a chain of surrogacies behind a widely publicized AI portrait.","core_discovery":"The paper's central claim is that applications of AI and machine learning in creative practice are best understood as a form of collaboration between human agents mediated by technical artifacts such as algorithms and data. It introduces the term 'machinic surrogate' to describe an AI-enabled tool that carries the identity of its human makers—called the 'human inspirers'—into a creative encounter, so that a user who appears to co-create with a machine is actually interfacing with the toolmakers through the surrogate. The machine's agency is derived, not autonomous: it is a re-creation of its makers' skills, decisions, and biases, situated in the algorithm and its hardware with all their limitations. If this is right, claims of autonomous machine creativity are incomplete, and authorship of computational artworks should be distributed across the human agents whose choices are inscribed in the tool.","pith_inferences":["A testable design consequence follows: co-creative tools that make the toolmakers' choices visible—datasets, objectives, model architectures—would make the human collaboration legible, while black-box interfaces would obscure it.","The surrogacy account has a natural limiting case: if algorithms begin designing algorithms, each iteration dilutes the original human agency, so the concept would eventually stop describing the creative relationship.","If courts or markets ever assign authorship to AI-generated art, the surrogacy view suggests the relevant human contributions include the toolmaking chain, which could reshape licensing and provenance practices.","One could probe the claim empirically: measuring whether users and audiences can detect different toolmakers' choices in generated outputs would test how strongly the surrogate actually transmits identity."],"forward_implications":["If the paper is right, the phrase 'human-machine co-creation' is misleading: the machine's contributions are the residue of human choices, so the active partners are all human.","Authorship of an AI artwork should be distributed along the toolmaking chain—from the developers of the base model to the artists who retrain or modify it—rather than credited to the machine or to the end user alone.","Users of an AI tool are working within a design space set by absent toolmakers; their freedom is real but bounded by decisions about architecture, data, and objectives made before the encounter.","The same artifact can switch roles: it can act as a surrogate for absent toolmakers, or, as in an installation retrained by live audience input, as a surrogate for the present audience.","Judging an AI artwork's quality therefore also means judging a human chain of curatorial and evaluative choices, not just the output of a probabilistic model."],"supporting_citations":[{"why":"It supplies the historical frames of computers as 'perfect slaves' or 'collaborative partners' and the argument that these technologies are infrastructural, designed artifacts enacting human intent.","marker":"[3]"},{"why":"It establishes that machine-learning tools carry the agency of their human makers, which is the origin of the surrogate's derived identity.","marker":"[8]"},{"why":"It supplies the third-identity account of artistic collaboration that the paper adapts to human-tool relationships.","marker":"[14]"},{"why":"It provides the account of an emerging identity in collaboration that lets the paper treat surrogacy as a form of collaboration.","marker":"[17]"},{"why":"It is the first case study, showing how toolmaker choices about architecture, dataset, and hyperparameters bound a user's generative design space.","marker":"[19]"},{"why":"It documents the second case study, in which absent toolmakers' choices shape a performance where audiences select inputs for a pose-to-image model.","marker":"[22]"},{"why":"It documents the third case study, an installation in which the tool acts as a surrogate for the present audience through real-time retraining.","marker":"[24]"},{"why":"It marks the creation of the GAN architecture, used as the example of developers acting as toolmakers and authors.","marker":"[27]"},{"why":"It demonstrates the chain of toolmakers behind a widely publicized AI portrait, as one developer's code is modified and retrained by others.","marker":"[30]"}],"fun_headline_variants":["AI art tools are surrogates for their human makers","Machine creativity is human collaboration via a surrogate","AI surrogates reveal the human behind the machine","When AI creates, it carries a human signature","Co-creation with AI is really co-creation with its makers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if a tool cannot inherit and transmit its makers' identity strongly enough for a genuinely new collaborative identity to emerge between a user and that inherited presence.","fun_headline_variants_meta":{"raw":{"variants":["AI art tools are surrogates for their human makers","Machine creativity is human collaboration via a surrogate","AI surrogates reveal the human behind the machine","When AI creates, it carries a human signature","Co-creation with AI is really co-creation with its makers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2478,"prompt_tokens":901,"completion_tokens":1577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":1500}},"tokens_in":517,"tokens_out":1577,"duration_ms":16797,"temperature":1.0,"reasoning_tokens":1500,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:22:01.056066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a controlled comparison of two versions of the same generative design tool, identical in interface, whose models are trained by different toolmakers on differently curated datasets; if the output distributions do not differ in ways traceable to those curation choices, the claim that users are collaborating with the toolmakers loses its observable support.","supporting_citations":[{"cited_title":"Empirically studying participatory sense-making in abstract drawing with a co-creative cognitive agent,","cited_arxiv_id":null,"evidence_quote":"It supplies the historical frames of computers as 'perfect slaves' or 'collaborative partners' and the argument that these technologies are infrastructural, designed artifacts enacting human intent."}],"review_version":1}