The weakest criticism of generative AI is that it produces bad prose. It often does. So do people.
The more serious problem is that generative systems make it extraordinarily cheap to produce material possessing the external characteristics of finished work without necessarily containing the corresponding density of intention, discrimination, revision or consequence. The result is what has acquired the useful pejorative slop: material that resembles culture while having passed through too little judgement to deserve much attention.
Slop is therefore not fundamentally an aesthetic category. It is a production-pathology.
Its defining characteristic is not that a machine participated in making it, but that the distance between generation and distribution has collapsed.
- Generate. Publish.
- Generate. Upload.
- Generate. Post.
The missing operation is editing.
This distinction becomes clearer if generative AI is understood not as an autonomous author but as an intermediate transformation system situated between two human acts:
COPY → GENERATE → PASTE
At first this appears almost trivial. Something is supplied to the machine; something is returned; something is placed elsewhere. Yet each transition represents a different intellectual operation.
- Copying is an act of reading.
- Generation is an act of transformation.
- Pasting is potentially an act of editing.
- Publication is an act of responsibility.
The important word is potentially.
If output is simply transferred from model to publication surface, paste is merely mechanical transmission. If the output is inspected, compared, rejected, recombined, corrected, contextualised and deliberately positioned, paste becomes part of an editorial act.
This distinction provides a much better boundary between generative slop and authentic AI-assisted work than the crude question, “Was AI used?”
I. The generative system sits in the middle
Traditional accounts of authorship tend to imagine a relatively direct relationship:
AUTHOR → TEXT → READER
Generative production inserts an intermediate system:
HUMAN → MODEL → OUTPUT → HUMAN → WORK → READER
That alteration is more profound than it initially appears.
The first human operation is intentional. The user selects source material, defines constraints, supplies context, establishes genre, describes purpose and determines what the system is being asked to transform.
The model then performs a probabilistic synthesis.
What it produces is not yet necessarily a work.
It is an output.
This distinction should be maintained rigorously.
An output is the consequence of generation.
A work is an object accepted into a cultural context by an agent willing to take responsibility for it.
The distinction resembles that between photographic exposure and photograph, recording and record, footage and film, or notes and essay. Production creates candidates. Editorial judgement turns some candidates into finished artefacts.
The generative system therefore creates a potentially enormous intermediate space between intention and publication.
That space is precisely where either slop or authorship emerges.
II. The semantic migration from author to editor
Generative systems complicate authorship because they disaggregate functions that were historically performed by the same person.
The traditional writer performs several operations simultaneously:
conception, sentence production, selection, sequencing, revision, deletion, structural control, tonal control and final approval.
Generative AI can assume a substantial portion of sentence production.
That does not automatically eliminate the other operations.
Instead, agency can migrate.
The person who previously acted principally as author may increasingly function as editor.
This is not merely a euphemism designed to preserve status. Editing is a distinct intellectual activity.
An editor asks questions that a generator cannot settle merely by producing more language:
- Does this belong?
- Is it true?
- Is this repetition useful?
- What is being implied?
- What should be removed?
- What has been omitted?
- Does this contradict something produced twenty pages earlier?
- Is this voice appropriate to the whole?
- Which of these three alternatives actually advances the argument?
- Where should the reader encounter this information?
- What should remain unresolved?
- What is the work attempting to become?
These are not principally generative questions.
They are discriminative questions.
Generative AI makes production abundant. Abundance makes discrimination more important.
This creates an inversion of the economics of writing.
Historically, generating text was expensive and editing comparatively cheap because there was relatively little text to evaluate.
With generative systems, generation approaches zero marginal effort while evaluation remains cognitively expensive.
The scarce resource is therefore no longer language.
It is judgement.
And once judgement becomes the scarce resource, the editor becomes more important rather than less.
III. From editor to meta-editor
The transformation becomes still more interesting when the user no longer works upon a single generated output.
Suppose a writer generates:
- five descriptions,
- twelve character sketches,
- three alternative structures,
- four different endings,
- a historical analysis,
- two stylistic rewrites,
- a chronology,
- a scene-by-scene critique,
- and thirty fragments accumulated across several months.
The intellectual problem is no longer simply editing generated prose.
It becomes the construction of a higher-order object from multiple subsidiary objects.
At this point the human role begins to resemble meta-editorship.
The editor works within the text.
The meta-editor works upon the population of possible texts.
Their material is not simply language but alternatives.
One might express the progression approximately as:
$A_0 = \text{direct authorship}$ $A_1 = \text{generation + editorial selection}$ $A_2 = \text{multiple generations + comparative editing}$ $A_3 = \text{cross-output synthesis + structural orchestration}$
At $A_3$, the author’s primary contribution may no longer be individual sentences. It may consist of controlling relationships among hundreds of candidate sentences, scenes, arguments and structures.
This is qualitatively different from pressing a button and accepting the result.
Consider a film editor working with forty hours of footage to produce a ninety-minute film. Nobody seriously argues that the editor contributes nothing because the editor did not photograph every frame.
The editorial intelligence lies partly in exclusion.
What is absent from the final film is as important as what remains.
Generative systems radically increase the amount of material available for rejection.
Consequently, a sophisticated generative workflow may contain a paradox:
the more material generated, the more human judgement may be required to produce a coherent work from it.
This is exactly where slop and serious generative production diverge.
Slop maximises output.
Authorship maximises selection.
IV. Duration changes the character of generated material
A crucial but underexamined variable is duration.
Generated outputs exist within time.
The temporal distance between generation and publication profoundly affects their status.
Consider four intervals:
$t_0 = \text{generation}$ $t_1 = \text{inspection}$ $t_2 = \text{revision and integration}$ $t_3 = \text{release}$
When:
$t_3 – t_0 \approx 0$
the likelihood of meaningful editorial mediation is low.
The system produces; the user distributes.
This is the characteristic temporal structure of slop.
Its defining feature is velocity.
The work has almost no residence time in judgement.
By contrast, an output that remains within an editorial environment for hours, days or months can accumulate interventions.
It can be compared with earlier versions.
Contradictions can become visible.
Weak passages can begin to irritate.
Extraneous material can be removed.
Research can falsify apparently convincing statements.
Characters can acquire continuity.
Themes can emerge retrospectively.
The author can forget the excitement of generation and encounter the material again as a reader.
This temporal separation matters enormously.
Immediate generation produces attachment to possibility.
Delayed editing produces confrontation with actuality.
The longer a work remains under serious editorial observation, the less meaningful it becomes to describe the final product simply as “AI output.”
The generated text has become raw material within a longer intellectual process.
This suggests a useful concept: editorial half-life.
A generated passage begins as predominantly model-produced material. Each significant human intervention—deletion, rewriting, relocation, contradiction checking, contextualisation, synthesis—reduces the explanatory power of that original description.
After sufficient transformation, asking whether the passage “came from AI” becomes similar to asking whether a film came from a particular reel of rushes.
Historically interesting, perhaps.
But insufficient to describe the finished object.
V. Release and distribution are part of authorship
The distinction between generation and publication matters because publication changes the ethical status of content.
A language model can generate ten thousand incorrect statements privately without affecting anyone.
The decision to release one of them changes everything.
Distribution is therefore not merely a technical afterthought.
It is an epistemic commitment.
By publishing something, the human operator effectively says:
I consider this sufficiently coherent to occupy somebody else’s attention.
If factual:
I consider this sufficiently reliable to influence somebody else’s understanding.
If artistic:
I consider this sufficiently formed to merit interpretation.
If commercial:
I consider this sufficiently valuable to request payment or attention.
These claims belong to the publisher, curator, editor or author—not to the generator.
The phrase “AI wrote it” consequently becomes morally inadequate once a person intentionally distributes the work.
The model generated.
The human released.
And release is a decision.
This is why authenticity cannot be evaluated solely at the point of textual production.
It must be evaluated across the complete chain:
SOURCE → SELECTION → PROMPT → GENERATION → EVALUATION → REVISION → COLLATION → CONTEXTUALISATION → RELEASE → RECEPTION
Slop short-circuits this chain.
Serious work lengthens it.
VI. Why aggregation can increase quality
There is an understandable intuition that combining generated outputs must compound artificiality.
Sometimes it does.
A hundred mediocre generated paragraphs pasted together usually create a larger mediocre object.
But aggregation can also operate in the opposite direction if it is accompanied by aggressive selection.
Suppose a system generates one hundred candidate fragments.
Ninety are discarded.
Six are substantially rewritten.
Three are combined.
One survives substantially intact.
The final work cannot be meaningfully understood merely by calculating the percentage of machine-generated tokens.
Its salient property is the selection function that produced it.
Let the generated set be:
$G = {g_1,g_2,\ldots,g_n}$
The final work is not:
$W = G$
but rather:
$W = E(G,C,I)$
where:
$E$ = editorial transformation,
$C$ = contextual constraints,
$I$ = human intention.
As $n$ increases, the potential search space increases.
But quality increases only if $E$ becomes more discriminating.
Thus:
$\text{more generation} \neq \text{more quality}$
while:
$\text{more alternatives} + \text{stronger selection} \rightarrow \text{potentially greater quality}$
The problem with slop is therefore not abundance itself.
It is abundance without corresponding selectivity.
VII. Contextual originality
Generative AI also forces a distinction between textual originality and contextual originality.
Textual originality asks:
Has this exact sequence of words existed before?
Contextual originality asks:
Has this arrangement of ideas, references, structures and purposes existed in this form before?
Human culture has always depended heavily upon contextual originality.
Shakespeare borrowed plots.
Modernist writers assembled quotation, myth and allusion.
Hip-hop constructed new works from existing recordings.
Collage creates novelty through adjacency.
Scholarship frequently contributes not new facts but new relationships among known facts.
Software engineering routinely produces original systems from pre-existing libraries.
Cuisine recombines inherited ingredients and techniques.
Originality therefore does not require atomic novelty.
It can arise from configuration.
Generative systems intensify this principle because they can generate many locally conventional components.
The human contribution may consequently move toward the global structure.
A sentence may be statistically ordinary.
A paragraph may be stylistically familiar.
A scene may use recognisable genre conventions.
Yet the complete work may still possess a distinctive conceptual topology.
Originality can reside in:
- the combination of domains,
- the persistence of unusual themes,
- the relationship between characters,
- the sequencing of revelations,
- the collision of historical and speculative material,
- the author’s particular obsessions,
- the decision to retain one contradiction and remove another,
- or the gradual accumulation of motifs across multiple works.
This is contextual originality.
It exists not necessarily at the token level but at the architectural level.
And generative systems may actually make this distinction more important.
VIII. Authenticity is not the absence of machinery
Arguments about AI often confuse authenticity with purity.
Under this model, authentic work must emerge directly from an unaided individual consciousness, preferably through manual inscription.
But almost no modern cultural production satisfies such a criterion.
- Writers use spellcheckers, dictionaries, search engines, editors, transcription systems and reference works.
- Photographers use autofocus, computational exposure, denoising and colour transforms.
- Musicians use digital instruments, quantisation, sampled sounds and automated mastering.
- Filmmakers use compositing and nonlinear editing.
- Programmers rely upon libraries containing millions of lines they did not write.
- Authenticity therefore cannot sensibly mean absence of mediation.
A more useful definition is:
Authenticity is continuity between intention, judgement and responsibility.
A work is authentic when somebody can meaningfully account for why it exists in the form in which it appears.
- Why this passage?
- Why this structure?
- Why this conclusion?
- Why this omission?
- Why release it?
If the answer to every question is effectively “because the model produced it,” authorship is weak.
If the answers form a coherent chain of human decisions, authorship is considerably stronger.
The essential distinction is not:
human versus machine
but:
intentional versus automatic.
IX. The anti-slop principle
From this follows a practical anti-slop principle:
Never treat generation as completion.
Generation should produce candidates.
Candidates should encounter resistance.
They should be reread.
Compared.
Contradicted.
Cut.
Merged.
Fact-checked.
Reordered.
Allowed to become embarrassing.
Allowed to become obsolete.
Returned to after sufficient time for their novelty to disappear.
The author should be prepared to destroy an afternoon’s generation in thirty seconds.
That asymmetry is important.
The generative system is rewarded for continuation.
The editor must be capable of stopping.
The model asks implicitly: What comes next?
The editor asks: Should any of this remain?
That second question is culturally much more valuable.
X. The editor as negative intelligence
Generative AI foregrounds a form of intelligence that industrial culture often undervalues: negative intelligence.
Generation is additive. Editing is subtractive.
Generation demonstrates possibilities. Editing imposes limits.
The generator can always produce another sentence. The editor must decide when no sentence is required.
This creates an important asymmetry.
A generative system can efficiently optimise for plausibility, coherence and continuation.
But a serious editor is often searching for exactly the opposite:
- the suspiciously smooth passage,
- the expected metaphor,
- the redundant explanation,
- the emotionally convenient resolution,
- the phrase that sounds impressive while saying nothing,
- the conclusion that closes an ambiguity which ought to remain open.
Slop is characterised by positive accumulation.
Authentic editing frequently operates through negation.
- Delete.
- Refuse.
- Replace.
- Reconsider.
- Leave blank.
- Start again.
XI. Meta-editing as authorship of systems
At sufficient scale, the human may cease to edit individual outputs directly and instead design the process that produces them.
They create prompts, style constraints, continuity documents, evaluation criteria, character databases, research corpora, rejection rules and iterative critique loops.
This creates another semantic transition:
author → editor → meta-editor → system designer
The meta-editor does not merely ask:
Is this paragraph good?
They ask:
What process reliably produces paragraphs appropriate to this project?
They design an editorial ecology.
A novelist maintaining a long-running generative project might therefore control:
- character continuity,
- chronology,
- narrative viewpoint,
- lexical exclusions,
- historical constraints,
- recurring motifs,
- preferred sentence structures,
- prohibited rhetorical habits,
- degrees of ambiguity,
- rules for dialogue,
- and criteria determining when generated material enters canon.
The individual output becomes less important than the system of acceptance surrounding it.
This begins to resemble architecture.
An architect does not manufacture every brick.
Their authorship lies principally in determining relationships among components.
Likewise, the generative meta-editor may increasingly author constraints, relationships and selection procedures.
Their medium becomes the possibility space itself.
XII. Authenticity through accumulated intention
The strongest defence of serious generative practice is therefore not to pretend that machine-generated language is secretly traditional human writing.
It is to acknowledge that a different production structure has emerged.
The human contribution can migrate upward.
From producing every word, to choosing among words.
From choosing among sentences, to organising scenes.
From organising scenes, to constructing works.
From constructing works, to maintaining continuities across collections.
At each stage the relevant question becomes less:
Who generated this token?
and more:
Who determined that this token should remain here?
This is where authenticity can survive. Indeed, it can become more visible.
A prolific generative system produces enormous variation. Against that abundance, the repeated preferences of a particular editor become increasingly conspicuous.
- What does this person consistently retain?
- What do they reject?
- What themes recur?
- What forms of discomfort interest them?
- What kinds of endings do they distrust?
- What historical periods repeatedly attract them?
- Which characters survive revision?
- Which moral questions refuse to disappear?
Across enough work, these decisions form a signature.
Not necessarily a lexical signature.
A judgement signature.
That signature may ultimately be more revealing of authorship than prose style alone.
XIII. Conclusion: authorship after abundance
Generative AI has created a crisis not because machines can produce language, but because language has become abundant enough that production itself can no longer serve as a reliable proxy for intellectual labour.
The cultural scarcity has moved.
The scarce things are now judgement, coherence, attention, sustained context and responsibility.
Consequently, the significant creative act increasingly occurs after generation.
Copy can remain an act of reading.
Generation becomes an expansion of possibilities.
Paste, when accompanied by judgement, becomes editing.
Repeated selection becomes curation.
Curation across many outputs becomes meta-editing.
Meta-editing sustained across time becomes authorship of a larger conceptual structure.
And release converts all of those private decisions into public responsibility.
This provides a defensible account of authenticity in generative culture.
Authenticity does not require pretending that the machine did nothing.
Nor does it require surrendering authorship merely because the machine did something.
It requires demonstrating that the final object has travelled through a sufficiently demanding field of human intention.
Slop moves rapidly from probability to publication.
Work acquires resistance along the way.
The ultimate anti-slop principle is therefore remarkably simple:
Generate freely. Select ruthlessly. Edit consciously. Publish deliberately.
The machine’s abundance is not the work.
What survives human judgement is.
