The Art of Subtraction: How AI Writing Mirrors Found Poetry

by

Author: Denis Avetisyan

A new framework reinterprets AI-assisted writing not as generation, but as a curated process of ‘subtractive authorship’ akin to blackout poetry.

The practice of blackout poetry, as exemplified by Macchini’s work in 2026, demonstrates how constraint-the deliberate obscuring of existing text-can reveal latent meaning and forge new narratives from the remnants of the old, highlighting the inherent ephemerality of language itself.
The practice of blackout poetry, as exemplified by Macchini’s work in 2026, demonstrates how constraint-the deliberate obscuring of existing text-can reveal latent meaning and forge new narratives from the remnants of the old, highlighting the inherent ephemerality of language itself.

This review proposes reframing human-AI collaboration in writing as a form of found-text creativity, leveraging activity traces to highlight authorial contribution and address concerns about content gatekeeping.

While increasing collaboration between humans and artificial intelligence promises new creative avenues, it simultaneously raises concerns about authorship and trust. This paper, ‘Human-AI Interaction Traces as Blackout Poetry: Reframing AI-Supported Writing as Found-Text Creativity’, proposes reframing the record of human-AI interaction-typically viewed as an audit trail-as a form of ‘subtractive authorship’ akin to blackout poetry, where AI-generated text serves as found material curated by the writer. By treating these interaction traces as expressive artifacts, we argue that designers can better highlight creative contribution and foster reader appreciation. Could this aesthetic approach to interaction design reshape perceptions of AI-assisted writing and unlock more fruitful collaborative practices?

The Shifting Sands of Authorship: Transparency and Perception

The landscape of content creation is undergoing a swift transformation, fueled by increasingly sophisticated artificial intelligence tools capable of generating text, images, and even complex narratives. This proliferation of AI-supported writing, however, introduces a critical concern regarding transparency. While these technologies offer unprecedented opportunities for efficiency and scalability, a lack of clarity about their involvement raises questions about authorship, originality, and potential bias. Determining the appropriate level of disclosure – and ensuring that audiences are aware when AI has contributed to a piece of content – is proving to be a complex challenge, demanding careful consideration from creators, publishers, and technology developers alike as the lines between human and machine contributions become increasingly blurred.

Studies reveal a counterintuitive truth regarding AI-assisted content: explicitly acknowledging its involvement frequently diminishes reader perception. While audiences may intellectually appreciate the benefits of AI in streamlining content creation, disclosures often trigger skepticism and a perceived reduction in quality or originality. This isn’t necessarily a rejection of AI itself, but rather a cognitive bias where human authorship is still strongly associated with trustworthiness and creativity. The effect appears to stem from a subconscious expectation that genuinely insightful or compelling work originates from human intellect, and the revelation of AI assistance can subtly undermine that expectation, even if the content is objectively well-written. Consequently, content creators face a delicate balancing act: leveraging AI’s capabilities while mitigating the potential for negative reader response linked to its disclosure.

The increasing integration of artificial intelligence into content creation presents a curious challenge to establishing and maintaining reader trust. While the benefits of AI-supported writing – increased efficiency, novel perspectives, and broader accessibility – are readily apparent, openly acknowledging its involvement can paradoxically diminish positive reception. Studies indicate a tendency for audiences to view AI-generated content with skepticism, even when the quality is comparable to human-authored work. This suggests that simply disclosing AI’s role isn’t enough; the method of disclosure – its framing, transparency, and emphasis – is critical. Navigating this tension requires careful consideration of how to reveal AI’s contribution without triggering negative biases, demanding a nuanced approach to authorship and credibility in the age of intelligent machines.

The Evolving Role of the Writer: From Originator to Curator

The increasing prevalence of artificial intelligence in content creation is fundamentally altering the traditional concept of authorship. Historically, authorship has been defined by the singular creation of original content; however, with AI capable of generating text, images, and other media, the writer’s role is evolving toward curation. This involves selecting, refining, and assembling AI-generated outputs into a cohesive and purposeful whole. Rather than originating all content, the author now exercises agency through critical evaluation and skillful integration of machine-generated material, effectively transitioning from a sole creator to a curator of content produced via algorithmic processes. This shift necessitates a re-evaluation of creative ownership and the skills required for effective content production in an AI-integrated landscape.

Content gatekeeping, increasingly vital for writers utilizing artificial intelligence, involves the deliberate selection and rejection of AI-generated text segments. This skill extends beyond simple editing; it requires evaluating outputs for factual accuracy, stylistic consistency, and alignment with intended meaning and audience. Writers employing content gatekeeping techniques assess multiple AI outputs, identifying and integrating only the portions that meet specified criteria, and discarding the rest. This process isn’t limited to correcting errors; it encompasses choices regarding tone, perspective, and the overall narrative structure, effectively shaping the AI’s raw output into a cohesive and purposeful final product. Proficiency in content gatekeeping demands critical judgment and a clear authorial voice to distinguish valuable content from unusable material.

The workflow of AI-assisted writing increasingly centers on Generative Substrates – the underlying AI models and data sets – which produce Intermediate AI Output. This output isn’t considered a finished product, but rather a foundational layer of text requiring subsequent authorial intervention. Writers then engage in a process of selection, revision, and integration, treating the AI-generated content as raw material. This ‘found material’ approach emphasizes that the final text is not solely the product of the AI, but the result of a curated refinement process applied to the initial AI output, forming the basis for a new, author-driven composition.

Our research explores a workflow where AI-generated text is intentionally treated as a source of ‘found material’ by writers, rather than original composition. This reframing serves a dual purpose: to actively mitigate inherent biases present in large language models – which are often trained on datasets reflecting societal inequalities – and to re-establish human creative agency in the writing process. By selectively adopting, revising, and arranging AI output, writers function as curators, applying critical judgment and stylistic control. This process emphasizes the human contribution not as initial creation, but as informed selection and refinement, thus highlighting the value of human oversight and editorial decision-making in shaping the final text.

Tracing the Echo: Visualizing AI’s Contribution

Activity Traces function as a granular record of interactions between a writer and an AI writing assistant. These traces detail specific actions such as text generated by the AI, edits made by the writer, acceptance or rejection of AI suggestions, and the temporal order of these events. Rather than a simple binary indication of AI use, activity traces capture the iterative nature of collaborative writing, documenting the continuous back-and-forth between human and machine. The resulting data allows for reconstruction of the writing process, revealing how the AI’s contributions were integrated, modified, or ultimately discarded by the writer, and providing insight into the evolving authorship throughout the document’s creation.

Transparency tools like HaLLMark and DraftMark utilize visualization methods to display the specific contributions of artificial intelligence during content creation. These systems function by analyzing writing processes and highlighting text generated, suggested, or edited by AI algorithms. Visualization techniques employed include color-coding, layering, and interactive displays that differentiate between human and AI contributions. The output is not simply a binary indication of AI usage, but a granular representation of how AI influenced the text at the sentence or phrase level, allowing for detailed assessment of the AI’s role in shaping the final product.

Current transparency tools differentiate themselves from simple AI detection by providing granular data on the specific edits and suggestions contributed by AI during the writing process. Rather than merely indicating AI involvement, systems like HaLLMark and DraftMark visually represent the provenance of individual text segments, highlighting additions, deletions, and modifications made by the AI, as well as the writer’s acceptance or rejection of those changes. This detailed tracing allows users to understand how the AI influenced the composition, identifying instances where it significantly restructured arguments, altered phrasing, or provided substantive content, and thereby enabling a nuanced evaluation of the final text’s authorship and originality.

Transparency tools utilizing activity traces and visualization methods provide detailed records of interactions between a writer and AI, enabling assessment of AI’s influence on generated text. This level of visibility allows writers to critically evaluate suggestions, refine their own contributions, and maintain authorial control. Simultaneously, readers can utilize these traces to gauge the extent of AI assistance, informing their evaluation of the text’s originality, accuracy, and overall quality. The ability to differentiate between human and AI contributions fosters accountability and supports informed judgments regarding the authenticity and reliability of AI-assisted content.

The Art of Absence: A New Aesthetic of Subtraction

Blackout poetry, a striking example of subtractive authorship within the broader realm of found art, reveals a creative process centered on deliberate removal. Practitioners take existing texts – books, newspapers, documents – and redact words, phrases, and even entire lines, obscuring the original content to reveal a new poem formed from what remains. This isn’t simply about what is written, but what is intentionally unwritten; the negative space becomes as crucial as the visible text. By carefully choosing what to erase, the poet sculpts meaning from pre-existing language, transforming the source material into something uniquely their own. The resulting poems often possess a fragmented, haunting quality, inviting readers to contemplate the interplay between presence and absence, and the power of suggestion within a limited textual field.

Contemporary writing practices are increasingly mirroring the process of subtractive creation, particularly as authors integrate artificial intelligence into their workflows. Rather than composing from a blank page, many writers now begin with text generated by AI, then meticulously refine and curate the output, deleting extraneous phrases and honing the remaining content. This approach positions the author not as the sole originator of the text, but as a discerning editor, shaping and focusing the AI’s expansive generation into a coherent and meaningful piece. The skill lies not in producing content ex nihilo, but in the artful selection and removal, transforming a broad, sometimes unwieldy, draft into a polished and purposeful expression – a process akin to sculpting meaning from a block of raw data.

The practice of deliberately restricting creative choices – embracing limitation – can unlock a distinctive aesthetic for writers. Rather than striving for boundless expansion, this approach centers on the power of what remains after careful curation and removal. By consciously paring down initial material, be it pre-existing text or AI-generated content, authors cultivate a focused voice and a heightened sense of intentionality. This subtractive process isn’t about deprivation; it’s about refining meaning through precise selection, allowing the essential elements to resonate more powerfully and forging a unique artistic signature born from thoughtful constraint. The resulting work isn’t simply shorter, but demonstrably more impactful due to the considered absence of extraneous detail.

The evolving relationship between authorship and artificial intelligence suggests a shift not toward replacement, but toward augmentation of creative processes. Rather than generating content ex nihilo, AI increasingly functions as a powerful tool for refinement and curation, presenting writers with expansive drafts to sculpt and shape. This collaborative dynamic reframes the creative act; the artist’s skill lies not solely in origination, but in discerning what to preserve within the generated material, establishing a unique aesthetic voice through selective subtraction. Consequently, AI isn’t diminishing human creativity, but rather providing a novel toolkit, prompting artists to explore new avenues of expression through a process of informed curation and deliberate reduction.

The study reframes human-AI collaboration not as seamless creation, but as a process of careful selection-a subtractive authorship mirroring the art of blackout poetry. This echoes a fundamental truth about all complex systems: they rarely spring forth fully formed. Rather, they evolve through refinement, pruning away the superfluous to reveal underlying structure. As Henri Poincaré observed, “Mathematics is the art of giving reasons.” This principle applies equally to writing; the author, even when aided by AI, provides the crucial reasoning-the judgment-that shapes raw potential into meaningful expression. The transient nature of ‘uptime’-of a polished, final text-becomes a fleeting moment in the inevitable cycle of decay and revision, a harmony achieved through deliberate subtraction.

What Lies Beneath?

The framing of human-AI collaboration as ‘subtractive authorship’ offers a temporary reprieve from the anxieties surrounding content generation. Systems learn to age gracefully, and this work suggests a means of accepting the inevitable presence of machine-generated text-not as a replacement for human creativity, but as raw material. However, this reframing merely shifts the locus of concern. The question of ‘gatekeeping’-who decides what remains, what is obscured-becomes paramount. The traces of interaction, the very act of subtraction, become the new artifact, demanding scrutiny.

Future investigations might consider the longevity of such reframings. Will the analogy to found art and blackout poetry withstand repeated exposure, or will it too become a transparent mechanism, revealing the underlying tensions it sought to mitigate? The field could also benefit from exploring the aesthetic qualities of these interaction traces themselves – not as evidence of authorship, but as emergent forms worthy of independent consideration.

Sometimes observing the process is better than trying to speed it up. The value may not lie in optimizing the collaboration, but in understanding how these systems – and their human counterparts – inevitably decay, adapt, and leave their mark on the resulting text. The traces are not a bug; they are the story.

Original article: https://arxiv.org/pdf/2604.09605.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-14 18:12