The Green Desert Effect

01 · AbstractA forest that is not a forest

AI-generated marketing copy is everywhere, and it looks good. Like it should work. It's grammatically clean, structurally on-pattern, and dependably on-topic. It can be produced at near-zero cost in any volume a business asks for. And here is the puzzle: in controlled head-to-head tests, AI-written persuasion performs about as well as human-written, sometimes better.

Bai et al. (2025), in three pre-registered experiments with 4,829 participants, found LLM-generated policy messages about as persuasive as ones written by laypeople; Salvi et al. (2025) found that a personalized GPT-4 out-persuaded human debaters. The copy is not, in isolation, weak.

And yet the channels feel dead. The dominant explanation in the industry, AI copy doesn't convert, sits awkwardly against that experimental record. The contradiction resolves the moment the unit of analysis changes. The failure is not in any single piece of copy; it is in what happens when an entire market generates from the same few models, tuned toward the same center, at the same time.

The individual message persuades. The aggregate flattens into sameness, and sameness is precisely what the reader's brain is built to discard. Your first instinct might be to call this a writing-quality problem and try to coach the model into better prose. That instinct is wrong.

The problem runs deeper than word choice. It sits at the meeting point of two systems: how the human brain decides what is worth attending to, and how AI language models decide what to say. To see why they pull against each other, start above ground, with something you can see from a car window.

If you drive past a commercial tree farm, from the road it looks like a forest. The trees are roughly the same height, planted in straight enough lines to look orderly, the right shade of green. Walk into it and the illusion drops.

The soil is acidic and compacted. Nothing lives in the understory. No bird species nests there. The water cycle has broken because the canopy is the wrong shape and the root systems are the wrong age. From a satellite it looks like a forest.

But when you walk into it, it feels different. It feels wrong. Forest-shaped, not forest.

Ecologists call this the Green Desert Effect.

And this is exactly what's been happening in marketing. Marketing has been clearcutting and replanting "monoculture tree farms" in the human attention ecosystem for roughly a century. Funnel optimization, A/B testing, SEO content factories, and templated email sequences produced this same kind of forest-shaped copy at industrial scale long before any large language model existed.

AI did not cause the problem. AI revealed it, by making the failure mode universal, cheap, and impossible to miss.

In this essay we diagnose the failure mode (Green Desert Marketing) at the architectural level, ground it in primary cognitive-neuroscience and AI-systems literature, and propose what a regenerative alternative would have to look like. The argument is that the predominant failure of AI-generated copy is mechanical, not aesthetic.

It is a function of how the reader's predictive brain works, and of how alignment-tuned generative models, by construction, produce signals the brain has evolved to deprioritize. We will define it operationally, show how it can be measured, ground it in primary literature, and trace it through the nine functional roles any persuasive text has to perform on a reader.

Then we will propose coppicing, the 5,000-year-old regenerative-forestry practice, as the operational alternative for writing that produces a real living ecosystem rather than depleting the very soil in which it grows.

02 · The readerThe predictive mind

The claim in this section is narrow and well-attested. The brain is a prediction model. It is constantly forecasting incoming input, and it updates its internal model only when the difference between prediction and actual signal is large enough, and relevant enough, to be worth the metabolic cost of updating.

The empirical foundation runs through several converging research programs rather than a single citation. Rao and Ballard (1999) first showed that the extra-classical receptive-field effects in early visual cortex are well-modeled as predictive coding, with feedforward signals carrying prediction error and feedback signals carrying predictions.

Friston (2010) extended this into the free-energy principle, a theoretical synthesis arguing that minimizing prediction error over time is a unifying principle for self-organizing biological systems. Summerfield and de Lange (2014) and de Lange, Heilbron, and Kok (2018) reviewed the neural and computational evidence for expectation effects on perceptual decision making and the microcircuitry that implements them.

Clark's Surfing Uncertainty (2016) and Hohwy's The Predictive Mind (2013) provide book-length syntheses. None of these is load-bearing alone; together they constitute a working model that maps onto a real and measurable behavior pattern in human cognition.

Two related findings sharpen the picture. Memory, contrary to common assumption, is forward-facing. Schacter, Addis, and Buckner (2007) argued that the evolutionary function of episodic memory is not preservation of the past but recombination of stored elements into simulations of future scenarios the organism can pre-evaluate.

The brain stores in order to simulate. Schultz, Dayan, and Montague (1997) showed that midbrain dopamine neurons encode a reward prediction error signal: they fire not in response to reward as such but in response to the difference between expected and received reward.

Prediction error has a neurochemical signature; the brain is built around the discrepancy, not around the absolute value.

Attention, on this picture, is allocated rather than given. Global workspace theory (Baars 2005; Dehaene 2014) describes conscious access as a bottleneck. Many parallel processes run unconsciously; only signals that exceed an ignition threshold reach the broadcast space where they become available to evaluation, memory, and decision.

The threshold is metabolically expensive to clear, and most signals do not clear it.

Predictive processing and global workspace theory are distinct frameworks. We use them as complementary rather than competing. Predictive processing describes how the brain weighs incoming signal by precision and how it updates internal models when discrepancy is informative. Global workspace theory describes which signals, having survived that weighting, reach conscious access.

The two operate at different levels of analysis. For the argument here, predictive processing supplies the mechanism that determines what gets a chance at the workspace, and global workspace theory supplies the mechanism that determines what reaches the place where persuasion can actually happen.

A reader will reach an obvious objection. If the brain rewards surprise, then a hallucinating language model should be persuasive, because it produces surprising output constantly. The objection misreads what kind of surprise matters. The brain does not reward random surprise. It rewards productive surprise: the kind that lets the predictive model update toward a state in which future predictions will be better.

Predictive processing models call this minimizing long-term, not immediate, free energy. A signal that is novel but not connected to anything the model is trying to predict gets coded as noise, filtered, and forgotten. The criterion is not novelty alone. Four conditions have to hold at once: the signal must be surprising, new, true, and relevant. Novelty without truth produces a hallucination the reader will not act on.

Truth without relevance produces a fact the reader will not remember. Relevance without surprise produces a confirmation the model already had. The combination is what gets through.

A line that clears all four looks nothing like the average. "Do you make these mistakes in English?", an ad that ran for decades, is surprising (you assume you do not), relevant (you care how you sound), true (everyone makes some), and new in the gap it opens (which mistakes?). The flat version, "Improve your English with our course," is true and relevant and predicts itself word for word, so the reader's model files it without a flicker.

There is a second constraint on the surprise, and it is a matter of dose. Prediction error does not buy attention without limit; the relationship is an inverted-U. The psychology of aesthetics has long described this shape, at least since Berlyne (1971): arousal and reward rise with novelty, complexity, and incongruity up to a peak, then fall as the signal becomes too much to resolve.

Advertising research found the same curve directly. Meyers-Levy and Tybout (1989) showed that moderately incongruent products were evaluated more favorably than products that were either fully expected or wildly incongruent. This corrects a tempting misreading of everything above. The goal is not to maximize surprise. A signal that cannot be resolved against the reader's model is not productive surprise; it is noise, and it is discarded as fast as the bland signal is. The target is a band: surprising enough to mark the signal, resolvable enough that the resolution succeeds.

AI-mean copy fails by sitting below the band; copy that tries too hard fails by overshooting it.

Below it sits "we deliver innovative solutions that drive results," where the reader's model predicted every word and none of it lands. Above it sits the headline so strenuously novel it cannot be parsed, the marketing equivalent of a pun that needs a footnote. The reader bounces off both.

The villain is the center of the distribution, not order or pattern as such.

One honest complication belongs here, because it cuts against the grain of the argument. Fluency, the felt ease of processing, is itself persuasive. Familiar, predictable, easy-to-read material is judged more likeable (the mere-exposure effect, which Bornstein (1989) meta-analyzed across two decades of studies) and judged more true (the illusory-truth effect, which Dechêne et al. (2010) pooled to an effect near d = 0.5).

A flattened, on-pattern sentence is not inert. It buys liking and credibility. The resolution is not to deny this but to separate the axes. Prediction error governs one set of outcomes (attention, encoding, durable belief-update): the things that decide whether a message is noticed and whether it changes what the reader does.

Fluency governs another: liking and felt-truth, which decide only whether a message that is processed goes down easily. These are two scales, not two ends of one. That the first decides behavior and the second only ease of reading is our synthesis (inferential); the effects on each axis, taken singly, are well attested.

Two things keep the concession from sinking the argument. In a saturated market every competitor banks the same fluency gains, so they confer no relative advantage; they raise the floor for everyone and move no one ahead. The symmetry is the point rather than a problem: when every competitor also sits at the averaged center, the prediction-error penalty is a flat tax too, and the field settles into a uniformly low-engagement baseline that the marked voice, and only the marked voice, can break. That is what the later prescription is for.

And the felt-truth that fluency buys is the cheap, decaying kind: repeated against a reader's accumulating experience of the source, it erodes rather than compounds, for reasons taken up in Truth across time.

Green-desert copy can score on the fluency axis and still fail the prediction-error one, which is exactly the failure marketers describe: copy that reads fine, tests fine, and moves nothing.

A word on the limits of this analogy.¹

03 · The machineAI is built for the opposite

Generative language models are also prediction systems. They predict the same way the brain does, in the sense that both produce expectations about what comes next given what came before. They predict in the opposite direction, in the sense that what counts as a good prediction inside the model is, by design, the most likely continuation rather than the most informative one.

The post-training step we now call alignment sharpens that opposition until it becomes the dominant feature of deployed model output.

A transformer language model is trained on the next-token prediction objective: produce the most probable continuation given prior context (Vaswani et al. 2017). At inference time, sampling strategies such as temperature, top-p, and top-k inject variance into the output, because greedy decoding from the maximum-likelihood next token feels mechanical to human readers.

The base transformer architecture is not, on its own, the source of homogenization. A base model sampled at high temperature with no alignment training can be strange, idiosyncratic, even disturbing. It is the post-training that does the flattening.

Modern deployed language models are post-trained via reinforcement learning from human feedback and related alignment techniques (Christiano et al. 2017; Ouyang et al. 2022; Bai et al. 2022). The procedure is roughly the following. A pretrained model produces candidate responses to a prompt; human raters rank them on helpfulness, harmlessness, and honesty; a reward model is trained to predict the rankings; the base model is then fine-tuned with reinforcement learning to maximize the reward model's output.

The result is a model whose outputs are regressed toward what human raters approve of, which in practice means polite, helpful, frictionless, un-opinionated, hedged at the edges, and structurally familiar.

This is the central mechanism. Alignment is functionally a regression-to-the-acceptable-mean filter applied on top of an architecture capable of producing wild output. Base models are not green deserts. Base models are stranger forests. Alignment is the mechanization of the commercial tree-farm metaphor: the wild edges are pruned because some rater flagged them as unhelpful or off-tone or weird, and the model is trained to produce more of what gets approved.

Over millions of preference comparisons the surface of the output distribution smooths. Zhang et al. (2025) locate the mechanism in the preference data itself: a typicality bias, raters systematically favoring familiar, fluent, modal text, that concentrates probability mass on the most expected response. Bender et al. (2021) supply the deeper half of the grounding: the system stitches together forms from its training distribution without communicative intent, so nothing pulls output away from the rated center, because there is no one behind the words with a stake in being understood. The output distribution is not a copy of the training distribution; it is a flattened image of the rater consensus over the training distribution.

That flattened image is what most readers now encounter when they read AI-generated copy.

This distinction matters because a common rebuttal misses the mechanism. Just increase temperature is not a full answer. Sampling increases variance inside an aligned preference manifold; it does not erase the manifold. The model can sound more varied while still converging on the same ratified center, the one with a balanced tone and a low-risk claim and an empathy addressed to no one in particular.

In persuasion terms, higher-temperature decoding produces many flavors of acceptable sameness. The variance is on the surface; the regression-to-rater-consensus operates underneath the sampler and survives it.

The flattening is not a worry about the future; it has been measured in present human writing, and the measurement isolates alignment as the cause. Padmakumar and He (2024) had people write with three kinds of help and found that writing with a feedback-tuned model significantly reduced the diversity of what different writers produced, while the un-tuned base model did not.

The homogenizing agent is alignment specifically, not language modeling in general, which is exactly the distinction this section has been drawing. Doshi and Hauser (2024), in Science Advances, found the same shape at the scale of a population: access to generative AI made individual stories more novel and more useful, and made the body of stories as a whole more similar to one another.

The average writer gains; the commons flattens. This is the mechanism the opening promised. The copy is individually competent (Bai et al. 2025; Salvi et al. 2025) and collectively homogenizing (Padmakumar and He 2024; Doshi and Hauser 2024), and it is the homogeneity, not any single sentence, that the reader's predictive brain meets.

The same pull operates inside the training loop itself. Shumailov et al. (2024), publishing in Nature, showed that language models trained on data recursively generated by previous language models degrade systematically across generations, losing the long tails of the original distribution and converging on a narrower output center.

This is direct evidence that the averaging effect is self-reinforcing across generations, and a structural parallel for the homogenization of culture under AI saturation. It is not direct proof of why an individual reader fails to engage with an individual AI-generated piece; that link is the predictive-processing analogy from the previous section.

Model collapse is the macro-ecological signature; predictive processing supplies the per-reader mechanism. The diagnosis rests on both.

Whether deployed language models also lack something deeper, a persistent world model or causal understanding, is contested. Choi (2023) and LeCun (2022) argue that current LLMs rely on shallow statistical heuristics rather than grounded causal reasoning. Benchmark studies have shown specific planning and causal-reasoning failures consistent with that critique, though counter-arguments and capability surprises continue to appear.

We do not need to resolve the debate to support the diagnosis here. The weaker claim is sufficient: alignment-tuned LLMs are demonstrably trained toward an output center that is, on average, smoother and less surprising than equivalent human writing produced under real stakes.

A careful reader will object that some AI-generated copy does convert. This is true, and the mechanism-level exceptions are recognizable. Heavy human editing that injects novelty after generation moves the output back off the alignment center. Offer asymmetry strong enough that copy quality is rate-limiting only at the margin can carry the conversion regardless.

Low-stakes informational contexts permit conversion because nothing significant is being asked of the reader's predictive model. Schema-congruent contexts where matching the predicted form is itself the goal, such as transactional product descriptions or boilerplate confirmations, perform adequately because the reader is looking for the form, not for surprise.

The Green Desert Effect is a propensity claim, not an absolute. Where the surface conditions invalidate it, the failure mode does not appear. Where they hold, in high-stakes persuasion, attention-scarce contexts, and audiences with prior exposure to AI-generated language, it does. These are the individual-level exceptions.

The collective claim is separate and stronger: even where a single AI-generated piece would convert, a market saturated with near-identical AI-generated pieces erodes the attention any one of them can earn. The propensity is a property of the field, not only of the page.

The bridge to the rest of the argument is short. The mathematical-and-alignment center of any deployed LLM is, by construction, the kind of signal the brain has evolved to deprioritize. The next sections define the failure mode operationally, show how to measure it, and trace it through the nine functions any persuasive text has to perform.

04 · DiagnosisGreen Desert Marketing

Before the metaphor, the definition. We define Green Desert Marketing as the failure mode in which language meets every surface criterion of competence (grammar, structure, fluency, on-topic, on-brief) and still fails to produce model-updating engagement in its intended audience. The diagnosis is operational, not aesthetic. Three observable criteria identify it in practice.

First, low orienting response. Reader skim rates exceed engagement rates by a wide margin; eye-tracking and read-depth metrics show pattern-skim rather than pattern-attend. The reader's eyes pass over the surface without the saccade-and-fixate pattern that indicates the workspace has been engaged.

Second, coherence without novelty. The text is fluently parseable but produces no measurable update in reader belief, recall, or behavior versus a no-treatment baseline. The reader can read it; the reader cannot tell you what it said an hour later.

Third, surface match, mechanism miss. The text would, by every editorial rubric a copy chief might apply, pass review. The conversion does not move. Each of the three is measurable. Together they constitute the working definition.

It helps to have a specimen in view. "Unlock your team's full potential with an all-in-one platform built for the way you work." Grammatical, on-brief, inoffensive, and the ten-thousandth sentence of its kind a reader has skimmed. Nothing in it was chosen; it is the most expected string of words its category permits. That is green-desert copy, and most readers can no longer see it at all.

The phenomenon is a propensity, not an absolute, and it lives at the level of the field more than the single page: a lone green-desert piece can convert, but the averaged voice repeated across a saturated market cannot. The effect describes the dominant failure mode in attention-scarce, high-stakes persuasion contexts, where small differences in reader engagement compound into large differences in business outcome.

With the definition in hand, the metaphor becomes literal rather than rhetorical. The tree farm again, now exact: AI-generated marketing copy that meets the three criteria above is structurally the same thing. It is recognizable as writing. It is functionally absent as persuasion. The trees are planted in rows because that is how you maximize harvest yield per acre under a clearcut economic logic; the sentences are structured the way they are because that is how you maximize approval rate per token under an RLHF training logic.

In both cases the surface optimization produces something that, viewed from the place that matters, has lost the property the system was originally producing.

The failure is mechanical, not aesthetic. A well-edited piece of AI prose can be elegant and still produce no global-workspace ignition, no crossing of the threshold into conscious attention. The reader does not consciously reject it. The reader skips it. The skip is automatic, pre-conscious, and unrelated to whether the prose is good in the writing-workshop sense.

This is what makes the failure mode confusing for editors and frustrating for marketers: aesthetic competence does not save copy whose mathematical center is, by construction, what predictive systems have learned to deprioritize. The criterion the reader's brain is applying is not the criterion the editorial review is applying.

This is also what explains the felt sensation many readers describe when they say something is wrong but cannot say what. The sensation is coherence-checking failing in the background. The reader's brain is running parallel checks, the nine that the next section enumerates, and one or more is returning a non-match.

The check is run pre-consciously; the reader feels the result, not the operation. The reader cannot articulate what is wrong because the check happens below the level of conscious attention.

One link in this chain has to be marked honestly. That a homogenized field of copy depresses engagement across an entire market has not been measured directly (inferential); the studies that look like evidence for it mostly measure something narrower: a reader's reaction to detecting that a given piece is AI-generated.

What is established is the behavior the claim rests on. The orienting response habituates: a stimulus that repeats and conforms to expectation stops eliciting attention, and it dishabituates the moment a novel stimulus appears (Rankin et al. 2009), which is why a distinctive voice can recapture a reader the average voice has lost.

Advertising has measured the same decay as wear-out, where response rises with exposure to a point and then falls (Schmidt and Eisend 2015). The most familiar instance is banner blindness, where readers learned to route their eyes around a class of stimulus that had become predictable in form and position (Drèze and Hussherr 2003).

Homogenized AI copy is, to the predictive brain, what the banner ad became to the eye: a high-prior, low-surprise class of signal the orienting system learns to skip. The per-field measurement is still owed; the mechanism it would confirm is not in doubt.

AI did not invent green-desert copy; the funnel factories were producing it long before any transformer existed. What AI changed is scale: it raised production volume by a factor of a thousand and dropped the marginal cost to near zero. What was tolerable when copy was expensive is intolerable now that it is free, because supply outstrips any audience's capacity to consume it and the average quality of the signal collapses.

The phenomenon already has names. Marketers call it the sea of sameness; the critic Kyle Chayka (2024) calls the broader cultural version Filterworld, the flattening of taste under algorithmic recommendation. Those names describe the surface. The contribution here is to locate the cause one layer down, in the alignment training that produces the language, and to give the felt sameness a mechanism in the reader's predictive brain rather than leaving it as an aesthetic complaint.

05 · MeasurementHow you'd measure it

The previous section called the failure measurable and left it there. That word is load-bearing: it is the difference between a real diagnosis and an aesthetic complaint. So here is what you would actually measure.

The quantity is surprise. Any language model can score, word by word, how expected each word was given everything before it: a low score for the obvious next word, a high score for the one you did not see coming. Information theory calls this surprisal, and Shannon (1948) showed it is the basic unit of how much a message actually tells you. Green-desert prose scores low the whole way through, because by construction it is the most expected continuation at every step.

That score is not just ink on the page; it predicts what the reader's body does. People slow down and fixate on high-surprise words and skim past the low-surprise ones (Rayner 1998; Smith and Levy 2013), and the same score predicts the N400, the brain's electrical "that was unexpected" response (Michaelov et al. 2024). A flat, low-surprise text is, measurably, one the eyes skim and the brain forgets.

So each of the three symptoms gets a real instrument. Skim instead of read shows up in eye-tracking and in a flat surprise score. Fluent but forgettable shows up as a telltale pair: easy to read, yet barely distinguishable from every other piece in the set, a sameness one sentence can hide but a whole body of copy exposes (Zhu et al. 2018). Passes review but does not convert shows up as the gap between an editor's yes and the A/B result. Well-written prose even has an optimum: it spreads its surprises out evenly, never going flat and never going frantic (Levy and Jaeger 2007).

Two honest limits. Score surprise with a mid-sized model, not the largest: a frontier model knows too much to stand in for an ordinary reader (Oh and Schuler 2023). And this is surprise in the words, not the felt, emotional kind the later roles lean on. So the lab confirms the mechanism, while the field gives you three instruments you can actually run: surprise-score, sameness, and conversion. None of them reaches the field-level claim that a homogenized market depresses engagement at scale, which stays inferential until someone measures it directly.

06 · LineageRhetoric got there first

The mechanism in this essay is new only in its application to language models. The study of persuasion arrived at it long ago, under a different name. McQuarrie and Mick (1996), in the foundational modern treatment of advertising rhetoric, define a rhetorical figure as an artful deviation from expectation: an expression that departs from what the reader predicted, in a way that is not an error and that the reader can resolve.

Their warrant is the one used here. Deviation, in Berlyne's sense, is what attracts and arrests attention; a figure works because it violates a prediction the reader did not know they were making. The same shape turns up wherever a sudden, resolvable shift in expectation produces pleasure.

Edward de Bono located humor and creativity in this same mechanism, what he called the brain's asymmetry of patterns: the mind runs down a main track, is shifted onto a side track it never saw coming, and finds the new route obvious only in hindsight (de Bono, The Mechanism of Mind, 1969).

The joke and the metaphor are the same move; each pays off a prediction the reader did not know they had made.

This dissolves an objection the argument so far invites. If persuasion runs on deviation from expectation, what about rhyme, alliteration, parallelism, the figures that are more ordered, more regular, more patterned than ordinary prose? They look like the opposite of surprise. McQuarrie and Mick's framework answers cleanly: figures come in two kinds, and both are deviations.

Schemes (rhyme, antithesis, chiasmus) are excess regularity; tropes (metaphor, pun, irony) are irregularity. A rhyme is surprising precisely because prose does not rhyme; it violates the reader's expectation that the sounds of a sentence are unordered. Excess order and deficient order are two doors into the same room.

Both doors are familiar from ad copy. "Melts in your mouth, not in your hands" is a scheme: a tight antithesis, more ordered than ordinary speech, surprising because prose does not balance itself that way. "Red Bull gives you wings" is a trope: literally false, a metaphor the reader has to resolve. Set either beside "a refreshing beverage for active lifestyles" and the difference is not quality. It is that someone chose to say it that way.

What the predictive brain rewards is not disorder, it is marked departure from the expected, in either direction. A marked departure is a signature of intent: it tells the reader that a mind chose to say it this way, on purpose, rather than defaulting to what anyone would expect. Unmarked copy carries none, which is why it reads as coming from no one in particular. And for the aligned model, that reading is not a heuristic misfiring; it is accurate. The stochastic parrot earlier was the same observation from the machine's side: there is no intent behind the form for the marking to signal.

That is the precise sense in which AI-mean copy fails. It is regular, but unmarked, regular in the way the reader already predicted, neither the patterned excess of a scheme nor the semantic swerve of a trope. It sits at the expected center on both axes at once.

A sonnet is extremely regular and maximally deviant from prose; green-desert copy is mildly regular and maximally expected. The distinction between those two kinds of regularity is the whole argument.

The empirical record here is unusually strong, and on exactly the condition that matters. McQuarrie and Mick (2003) embedded figurative and literal ads in a mock magazine and tested recall under incidental exposure, where the reader was free to turn the page and was never told to attend.

Figures roughly doubled recall even then. That is the field condition the whole diagnosis depends on, a reader under no obligation to keep reading, and it is the condition under which the bland, expected, unmarked signal most reliably loses.

07 · FrameworkWhat persuasion actually is

Persuasion is what happens when a reader's prediction model updates and the perceived best course of action changes as a consequence. The update requires material that is surprising, new, true, and relevant; it must answer the nine background questions every reader's mind is running in parallel as it processes any persuasive text.

The catalog of persuasive techniques developed across the history of writing is, in aggregate, the menu of moves a writer can make to satisfy those questions. The nine functional roles are the categories under which those moves are sorted, organized by which question each move primarily answers.

Real persuasion is not a sequence of these moves but their convergence in a single reader: the moment all nine are answered at once. What makes that convergence hold is the subject the section closes on.

The historical anchor for this picture is Aronson's self-persuasion frame (Aronson 1972). The persuader supplies material; the update is intrinsic to the reader's brain working on that material. Petty and Cacioppo's elaboration likelihood model (1986) gives the most empirically supported decomposition of how the processing unfolds: a central route, active when the audience is motivated and the stakes are real, in which arguments are evaluated on substance; and a peripheral route, in which heuristic cues do the work.

This bears on the present argument in a specific way. The kind of update this essay is about, a durable change in what the reader believes and will do, is a central-route event: it happens only when the reader is motivated and able to elaborate.

Surprise that lands on the peripheral route buys a flicker of attention and little else. So productive surprise is not only resolvable and relevant, it does its lasting work on a reader willing to think, which is a further reason the bland signal, giving the central route nothing to work on, neither converts nor lasts.

Mercier and Sperber's argumentative theory of reasoning (2011) complicates any simple "facts update beliefs" picture: reasoning evolved primarily to produce and evaluate arguments in social contexts. The reader is not asking whether the claim is true in the abstract; the reader is asking whether the claim is true coming from this speaker, in this context, about this topic, given what the reader already believes the speaker is up to.

The six principles of influence (Cialdini 1984) remain the most-cited historical reference in modern persuasion writing. The nine functional roles are a structural departure, organized by function-in-the-reader rather than principle-of-influence, but the lineage is acknowledged. Where Cialdini's principles describe levers a persuader can pull, the roles describe questions a reader's brain is running.

The same technique may answer multiple questions; the same question may admit multiple techniques. The closest precedent for organizing by the reader rather than the technique comes from the direct-response tradition itself: Eugene Schwartz's Breakthrough Advertising (1966) sorts copy by the reader's stage of awareness, from unaware to most-aware, on the principle that what to say is fixed by the state the reader is already in.

The nine roles extend that instinct from one linear scale to nine simultaneous questions.

What follows is each of the nine roles, with the mechanism citation, the causal-chain pass tracing how RLHF-driven averaging degrades that specific function, the persuasion link, and an in-line confidence marker where the bridge from laboratory to field is inferential. One scope note carries through all nine.

Read each of the nine as a propensity of the averaged center, not a verdict on any single piece. Isolated, an AI-written line can clear any of these thresholds, which is exactly what the persuasion experiments found. The claim is directional: on each function the aligned default underperforms a marked, intentional alternative, and that gap widens as a field saturates and attention grows scarce, until the averaged voice is the one the reader has already learned to skip. A heavily edited passage can pass any of these roles. The default does not, and the default is what scales to fill a channel.

DISRUPT answers the reader's first question, why am I even reading this? Conscious access is rationed; only signals past an ignition threshold reach the workspace where evaluation and decision happen (Baars 2005; Dehaene 2014). Novelty and prediction-error spikes trigger ignition; the P3 oddball component indexes exactly this orienting to unexpected stimuli (Polich 2007).

The pull is motivational as well as neural: Loewenstein's (1994) information-gap account holds that curiosity is the felt response to a gap between what one knows and what one wants to know, sharpest when a prompt reveals a gap the reader did not know they had, which is what an opened loop or a withheld answer does.

Alignment-tuned output, regressed toward expected forms, generates lower prediction error in the reader. Against a field of near-identical openers, the averaged headline wins a shrinking share of the workspace, and the copy beneath a headline that lost the slot goes unread.

The marked alternative is a museum piece: Volkswagen's "Lemon." (DDB, 1960), one word under a photograph of the car, violating every prediction a car advertisement had ever installed. The reader is inside the ad before deciding to be.

IDENTIFY answers is this written for me? The cortical-midline self-referential network privileges processing of stimuli that match the reader's current self-model (Northoff et al. 2006; Metzinger 2003). Green-desert copy is calibrated to no individual reader's self-model. It lands as generic. It fails the self-relevance filter and is attenuated before deeper processing begins.

The marked version names the reader exactly. "To People Who Want to Write—but Can't Get Started" ran for decades because it is not the description of a market segment; it is the reader's own self-talk, returned verbatim. It passes the self-relevance filter because it is the self-model.

AGITATE answers how bad is my problem, really? Felt severity is constructed, not received. Interoceptive prediction error in the anterior insula integrates visceral and homeostatic signals into a feeling-state; the feeling is the discrepancy (Craig 2009; Barrett 2017). Green-desert copy describes problems in abstract, statistically common terms.

The insula generates no specific interoceptive prediction error in response. The reader does not feel the problem as worse than before, so there is no felt impetus to act. The mechanism is well-established; the bridge from insula activation in a laboratory paradigm to copy outcomes at scale is inferential.

The textbook specimen is Listerine's "Often a bridesmaid but never a bride" (1923), which converted a clinical abstraction into a specific, social, felt scene. It is also the cautionary half of the lesson: the campaign manufactured the severity it agitated, a coherence violation of exactly the kind Truth across time prices.

REFRAME answers am I thinking about this correctly? Conceptual blending is the operation here: schemas built from embodied source domains pre-structure what counts as a relevant fact, and reframing forces the brain to recompute the mapping between mental spaces (Lakoff and Johnson 1980; Fauconnier and Turner 2002; Gibbs 2006).

Averaged copy reproduces the reader's existing frame, because the existing frame is what the training distribution contains in aggregate. No reframe is offered. The reader's understanding does not change, and so the perceived best course of action does not change.

Avis's "When you're only No. 2, you try harder" is the move in its purest form: the same fact, second place, recomputed from liability into a guarantee of effort. Nothing about the product changed; the mapping did.

PROVE answers can I trust the claim? Credibility runs on two parallel systems. Source monitoring tags incoming claims with metadata about who said it and how confident they were (Hovland, Janis, and Kelley 1953; Pornpitakpan 2004). Processing fluency, the felt ease of comprehension, is itself read as a credibility cue (Schwarz 2004); warmth is assessed before competence (Fiske, Cuddy, and Glick 2007).

Green-desert copy is fluent, which helps fluency-based credibility, but it sources no one, which hurts source monitoring. The reader registers no specific person standing behind the claim, only an aggregated voice with no identifiable origin.

"At 60 miles an hour the loudest noise in this new Rolls-Royce comes from the electric clock" (Ogilvy, 1958) shows the role satisfied: a claim so specific it carries its own verification, with a named writer standing behind it.

ELEVATE answers what does my life look like after? Episodic future thinking uses the same medial-temporal and prefrontal network as episodic memory; concrete, sensorily anchored future scenes produce stronger valuation signals than abstract benefit claims (Schacter, Addis, and Buckner 2007; Peters and Büchel 2010; D'Argembeau et al. 2010).

The copywriting tradition reached the same point by craft: Eugene Schwartz's The Brilliance Breakthrough (1994) built its method on picture words, concrete terms that force a specific mental image, against the abstract connectors that leave the reader seeing nothing. Averaged output renders generic transformed-life scenarios that align to no reader's particular life.

The simulation is bland, the valuation signal is weak, the future does not feel real enough to want. The mechanism is established; the field bridge from simulation vividness to purchase intent is inferential.

Apple's "1,000 songs in your pocket" (2001) is the role satisfied in five words: a scene rather than a spec, your pocket, your songs, a number concrete enough to simulate. Picture words, exactly as Schwartz prescribed.

RESOLVE answers what's stopping me from saying yes? Loss aversion is the operative finding: anticipated losses are weighted more heavily than equivalent gains (Kahneman and Tversky 1979; Tom et al. 2007). Identity-protective cognition compounds the effect: self-affirmation procedures reopen the audience to disconfirming information by reducing the felt threat of being wrong (Cohen and Sherman 2014).

Green-desert copy handles objections in templated, predictable ways. No novel reframe of risk is offered. Identity threat remains unaddressed. The barrier between intent and action holds.

Domino's "30 minutes, or it's free" dissolved the operative objection, what if it arrives late and cold, by moving the loss onto the company's side of the ledger: loss aversion answered in kind rather than reassured away.

COMPEL answers why should I act right now? Subjective value computation in the ventromedial prefrontal cortex tracks immediate and delayed rewards on a common neural currency; urgency framing shifts weight onto the immediate side (Kable and Glimcher 2007; Schultz, Dayan, and Montague 1997).

This kind of copy uses urgency framing in templated ways the reader's predictive model has fully discounted. The value computation registers the cue, recognizes it as the usual move, and weights it barely. The field bridge from a laboratory choice paradigm to a real purchase context is inferential.

The urgency that still works is structural rather than claimed. The auction's "going, going, gone" cannot be discounted, because the deadline is a mechanism, not an assertion; "limited time offer" may be the most fully discounted phrase in commerce.

BOND answers do I like and trust who's speaking? Narrative transportation reduces counter-arguing and activates identification with the speaker (Green and Brock 2000). Source credibility decomposes into expertise and trustworthiness, with warmth assessed before competence (Hovland, Janis, and Kelley 1953; Fiske, Cuddy, and Glick 2007; Pornpitakpan 2004).

Green-desert copy has no first-person speaker the reader can transport into; warmth and particularity, the signatures of intent a reader leans on to feel a real person behind the words, are exactly the dimensions alignment training averages out. The relationship the reader is being invited into has no one on the other end of it.²

The marked alternative ran for 28 years: the Wall Street Journal's "two young men" letter (Conroy, 1975), which opens with a story, same college, same graduation, two different lives, and holds the reader inside a specific voice until the pitch arrives. Transportation does the bonding; the reader reaches the offer already inside the narrative.

Truth is the governing principle that holds the nine together. Persuasion is not a sequence of techniques deployed one after another. It is the progressive alignment of nine dimensions of truth in a single reader's experience. The decision happens at convergence: when all nine become simultaneously true.

Truth cannot be deployed tactically, because it is the precondition that makes any technique work; a technique applied without the underlying truth will fail the reader's coherence check on its own dimension and pull the other eight down with it.

08 · The long runTruth across time

The picture above is the idealized version. A single reader, a single piece of copy, all nine dimensions answered. In practice, short-term persuasion routinely succeeds without all nine being true. A reader can be tricked once. Tricked twice. Tricked five times. Hyped offers, manufactured urgency, fabricated proof, manipulated identity claims; these convert reliably enough to support a whole category of marketing built around them. The conversion rate is real. The coherence is not.

What changes is the second-order. Every coherence violation that survives the initial conversion gets written into the reader's predictive model as evidence about the speaker. Memory is forward-facing. Each subsequent encounter with the same source starts from a more skeptical prior. Trust is built additively and lost multiplicatively, so one violation can erase ten consistent touches, an asymmetry consistent with the negativity bias by which bad impressions weigh more than good (Baumeister et al. 2001).

The cheap felt-truth that fluency buys runs down on the same schedule: an easy, familiar line feels true in the moment, but that impression is written against the reader's growing memory of the source, so it fades where earned coherence compounds.

The reader who has been tricked five times is not merely gone. They become an active anti-signal in their network, warning others away.

The brands that compound do it the other way. Patagonia, Levi's, and In-N-Out hold communication, product, and identity consistent across thousands of contact points over decades. Each consistent touch deepens the predictive model the reader holds, and each is an investment in the only kind of attention that compounds: earned attention. Manipulative tactics work as one-shot extraction. Truth works as compounding return.

This is the point at which an evidence-based-marketing tradition would object, and the objection is worth meeting head-on. Byron Sharp and the Ehrenberg-Bass school (Sharp 2010) argue, from decades of purchase data, that brands grow less through persuasion than through mental availability and distinctive assets, the consistent, repeated codes (the color, the logo, the jingle) that make a brand easy to recognize and recall.

Repetition, on that account, not surprise, is what builds a brand. The two views reconcile, and the reconciliation sharpens both. A distinctive asset is itself a sustained deviation, not from the brand's own history, where it is rigidly consistent, but from the category: it is the thing that makes this brand unexpected against its competitors, held still long enough to be learned.

Novelty earns the encoding; consistency is what gets encoded. Prediction error is how a distinctive asset enters memory in the first place; mental availability is what it becomes once it is there.

This also meets the objection that the two mechanisms collide. If a fully-predicted signal is exactly what the orienting system skips, how does a familiar logo or jingle keep working on its thousandth exposure? Because the two run in different systems. Orienting decides what wins attention in the moment, and it habituates; retrieval decides what a learned cue fires at the point of decision, and it never had to clear the orienting threshold to do that. The asset earns its place through deviation once, then pays out through effortless recognition for years. The claim here and Sharp's are compatible rather than identical, operating in different systems and on different timescales; the reconciliation is our synthesis (inferential), not a position Sharp himself takes, since he de-emphasizes persuasion in favor of availability.

The conventional split between direct response and brand equity reads, on this view, as a false tradeoff. A piece of copy coherent across all nine dimensions converts and deepens at the same time. The dichotomy is an artifact of writing that paid the conversion bill by burning the brand bank.

The 9 Roles dissolve the split. Coherence in the unit of work and coherence across the arc of work are the same operation at two timescales.

Truth, then, is not a moral preference. It is a mechanical necessity for any business that intends to operate beyond a single purchase cycle. The Green Desert is one expression of this failure. Manipulative direct-response copy that converts once and decays is another.

Both fail the coherence check. AI made the failure universal and cheap, but the diagnosis is older than the technology. This is what we mean when we say marketing has been clearcutting the human attention ecosystem. Yield is being extracted from a system whose conditions for future yield are not being maintained.

The soil is dying because the practice is unsustainable, not because each individual harvest fails. There is another way.

09 · PrescriptionCoppicing as the regenerative alternative

Coppicing is a 5,000-year-old forestry practice (Rackham 1986). The practitioner cuts the tree near the base. The system regrows from the stool. Yield continues for centuries from the same root. The soil deepens rather than thinning. Biodiversity is preserved because the canopy is broken periodically and light reaches the understory.

The opposite of commercial clearcutting at every point: clearcutting takes and leaves a depleted system; coppicing takes and leaves the system more alive than it found it. The metaphor that names the company also names the writing practice the diagnosis implies. Take from the system. Leave the system alive.

The diagnosis in the prior sections does not just imply "write better." It implies a set of mechanical requirements for any writing system that operates at AI scale and still produces output the predictive brain registers as worth attending to. The four requirements below describe what such a system has to satisfy structurally.

They are stated as principles, not recipes. Each one is an infrastructure problem in its own right; together, they describe what regenerative writing at scale would have to look like before any specific implementation gets built.

A controlled vocabulary, not probability. Persuasion moves have to be specific, finite, and retrievable by function. A system without a controlled vocabulary has no choice but to freestyle, and freestyling at AI scale converges back to the training-distribution average. The vocabulary has to be grounded in primary sources: actual customer language, actual proven copy, actual technique catalogs maintained by people who have written and tested at volume.

A vocabulary inherited from the training distribution will simply regenerate it. This is not only a craft preference; it is the documented direction of repair. Retrieval-augmented generation, which grounds output in a curated external corpus rather than in the model's own parameters, produces text that is measurably more specific and more diverse (Lewis et al. 2020), and it is grounding in real rather than model-generated data that arrests distributional collapse. A controlled vocabulary is that corpus, built for persuasion.

Engineered variety. A system has to structurally allocate against three competing pressures at once: what is proven, what is undersampled, and what is underrepresented. Without an explicit budget that surfaces all three, every session collapses toward whichever is weighted highest, and the system produces a smaller and smaller fraction of its own surface area as it runs.

The variety lane is not optional. It is what prevents the system from becoming its own training distribution. The collapse it guards against is the typicality bias met in the diagnosis: alignment training pulls output toward the mode of the distribution.

It is also, encouragingly, reversible by architecture rather than by abandoning the model. Methods that elicit a distribution of candidates instead of the single most-likely one recover much of the diversity alignment removes, at no measurable cost to quality (Zhang et al. 2025; see also Meister et al. 2023 on diversity-preserving decoding).

Engineered variety is that finding applied as a standing budget rather than a one-off setting.

Output-layer pattern checking. A system has to check its output against the model's natural fingerprint before the output reaches a reader. The check has to be statistical, not editorial, because the patterns that make AI-generated prose register as AI-generated operate below the level of conscious review.

Sentence-length variance, threading density, banned-trope frequency, parallelism quotas. These are measurable. They are not what an editor sees. An editor catches the obvious; the measurable patterns are what survive an editor's pass and still tip the reader's coherence check. That AI-generated text carries a measurable statistical fingerprint is established; it occupies characteristic high-probability regions a detector can read (Mitchell et al. 2023).

Turning such detectors from a provenance test into a pre-publication gate is a reasonable engineering inference rather than a proven practice (inferential), but the signal they read is real.

A closed feedback loop from real outcomes. A system that does not measure what worked downstream cannot stay regenerative. Real conversion events have to flow back into which persuasion moves surface tomorrow, weighted by the trust the source deserves and the value the outcome represents.

Without that loop, even an initially-good system reverts to its prior. The proven pool stops being proven. The system becomes a polished version of what every other aligned model is producing. Optimizing content against real outcomes is mature engineering: contextual-bandit methods reweight what surfaces next from live engagement and reliably outperform static selection (Li et al. 2010).

But a closed loop is also precisely how a system homogenizes itself if left unconstrained; outcome-only optimization collapses toward a narrow winner and reproduces the green desert from the inside (Jiang et al. 2019). So the feedback loop and the variety budget are not two requirements but two halves of one control problem: the loop pulls toward what has already worked, the variety budget pulls back toward what has not yet been tried, and the system is only as regenerative as the balance held between them. That balance is engineered and maintained, never automatic. This is the standing explore-exploit tension, and naming it is part of the architecture rather than a hole in it: a system that claims to have dissolved the tension has almost certainly collapsed toward exploitation, which is the desert by another route.

The four requirements describe what a system has to do. Underneath them is one practice that describes what it has to receive: human signal at every step. Feedback on individual moves, drafts converted to live assets, annotations on output, corrections to errors that surface in production.

Every signal reweights what the system surfaces next session. Without human signal flowing into the loop, the loop is closed on the wrong measure.

These are not soft prescriptions; they are mechanical requirements, and taken together an infrastructure problem of real scope. Berry's agrarian essays, Bateson's Steps to an Ecology of Mind, and Kimmerer's Braiding Sweetgrass sit in the background as the broader intellectual tradition this argument draws on; the relevant volumes appear in the bibliography below.

Coppicing with AI is not anti-technology. It is anti-monoculture. It asks any system that writes at scale to preserve the conditions under which language can still cause update.

The Persuasion Taxonomy is the public reference work, the controlled vocabulary made openly citable. The application at coppica.com is where the four requirements have been built into an operational system. We have published the parts of this work that are gifts: the catalog, the 18 Theses, this essay.

The implementation stays in the working system, where it belongs. The alternative to the Green Desert is not better prose. It is the architecture, and the architecture has to be built.

10 · CodaClosing

The argument has a central claim and three subordinate ones. The central claim: most AI-generated marketing copy fails not because it is poorly written, and not because it cannot persuade in isolation, which it demonstrably can, but because alignment training, by design, pulls an entire market's output toward a shared center the predictive brain has learned to deprioritize.

The failure is collective before it is individual. The subordinate claims: the failure is mechanical rather than aesthetic and so cannot be fixed by editorial polish; the failure runs through nine specific functional roles each of which is degraded by the same averaging mechanism; and a regenerative alternative is available, and is an architecture rather than a style: specificity, embodied diction, identifiable voice, and responsive practice over time, built into the system that produces the writing.

The Persuasion Taxonomy is the working library this essay points to. A catalog of techniques that have, across centuries, actually moved prediction models in actual readers. Free to read, to cite, to teach, and to build on. The 9 Roles overview is the structural argument; the catalog entries are the techniques themselves, sorted by which of the nine reader-questions each one primarily answers.

We end where we began, on the limits. Predictive processing is one cognitive framework, not consensus. The specific claims marked inferential above are bridges from laboratory finding to field application that we believe are well-motivated but have not been demonstrated at field scale by ourselves or others.

The Green Desert Effect is the explanatory bridge between the AI-side mechanism and the reader-side mechanism, and as a unified diagnosis it is our claim rather than a finding we are reporting. Chief among the bridges still owed is the one from measured homogenization of output to depressed engagement across a whole field; we have rested it on convergent analogs in wear-out, habituation, and banner blindness, and marked it as inferential rather than shown.

What comes next is ours to do. We are going to keep testing this account of how persuasion moves a reader, validating the inferential bridges or correcting them, and folding what we learn back into the system that applies them. The honest measure of the diagnosis is not whether it reads well but whether it makes the writing work better. That is the test we are running, and we will report what it shows.

Footnotes

1. Predictive processing and the free-energy principle are real research programs with substantial influence in cognitive science. They are not consensus. Critics including Klein (2018) and Bruineberg, Kiverstein, and Rietveld (2018) raise serious challenges to whether the framework explains as much as its proponents claim. Klein's "dark room" problem is the sharpest form: a literal reading of surprise-minimization predicts that an organism should seek out a dark, silent room and stay there, which organisms conspicuously do not, so the principle needs auxiliary assumptions before it predicts any behavior at all. We use predictive processing as a model that maps onto a real failure mode in AI-generated content, not as settled neuroscience. The analogy itself is our claim. The neuroscience grounds the analogy; it does not prove it. To be explicit about epistemic status: Established, the brain produces predictions about incoming input and updates internal models against discrepancy. Inferential, alignment-tuned generative AI's regression-to-acceptability behavior maps onto what predictive processing models would treat as low-precision, low-novelty signal. Speculative, we propose The Green Desert Effect as the explanatory bridge between these two and as the unified diagnosis of a class of writing failures the field is currently noticing piecemeal. ↩

2. The much-cited Kosfeld et al. (2005) finding on intranasal oxytocin and trust has had mixed replication and is not load-bearing for the argument here. The more recent careful update by Declerck et al. (2020) is cited instead. ↩

Bibliography

Aronson, E. (1972). The Social Animal. W. H. Freeman. https://archive.org/details/socialanimal0000aron

Baars, B. J. (2005). Global workspace theory of consciousness: toward a cognitive neuroscience of human experience. Progress in Brain Research, 150, 45–53. https://doi.org/10.1016/S0079-6123(05)50004-9

Bai, Y., Jones, A., Ndousse, K., et al. (2022). Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv:2204.05862. https://arxiv.org/abs/2204.05862

Bai, H., Voelkel, J. G., Muldowney, S., Eichstaedt, J. C., and Willer, R. (2025). LLM-generated messages can persuade humans on policy issues. Nature Communications, 16, 6037. https://doi.org/10.1038/s41467-025-61345-5

Barrett, L. F. (2017). How Emotions Are Made: The Secret Life of the Brain. Houghton Mifflin Harcourt. https://archive.org/details/howemotionsarema0000barr

Bateson, G. (1972). Steps to an Ecology of Mind. University of Chicago Press. https://archive.org/details/stepstoecologyof0000bate

Baumeister, R. F., Bratslavsky, E., Finkenauer, C., and Vohs, K. D. (2001). Bad is stronger than good. Review of General Psychology, 5(4), 323–370. https://doi.org/10.1037/1089-2680.5.4.323

Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). On the dangers of stochastic parrots: can language models be too big? Proceedings of FAccT 2021, 610–623. https://doi.org/10.1145/3442188.3445922

Berlyne, D. E. (1971). Aesthetics and Psychobiology. Appleton-Century-Crofts.

Berry, W. (1977). The Unsettling of America: Culture and Agriculture. Sierra Club Books.

Bornstein, R. F. (1989). Exposure and affect: overview and meta-analysis of research, 1968–1987. Psychological Bulletin, 106(2), 265–289. https://doi.org/10.1037/0033-2909.106.2.265

Bruineberg, J., Kiverstein, J., and Rietveld, E. (2018). The anticipating brain is not a scientist: the free-energy principle from an ecological-enactive perspective. Synthese, 195(6), 2417–2444. https://doi.org/10.1007/s11229-016-1239-1

Chayka, K. (2024). Filterworld: How Algorithms Flattened Culture. Doubleday. https://www.kylechayka.com/filterworld

Choi, Y. (2023). Why AI is incredibly smart and shockingly stupid. TED Talk. https://www.youtube.com/watch?v=SvBR0OGT5VI

Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg, S., and Amodei, D. (2017). Deep reinforcement learning from human preferences. arXiv:1706.03741. https://arxiv.org/abs/1706.03741

Cialdini, R. B. (1984). Influence: The Psychology of Persuasion. HarperCollins. https://www.harpercollins.ca/9780062937650/influence-new-and-expanded/

Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780190217013.001.0001

Cohen, G. L., and Sherman, D. K. (2014). The psychology of change: self-affirmation and social psychological intervention. Annual Review of Psychology, 65, 333–371. https://doi.org/10.1146/annurev-psych-010213-115137

Craig, A. D. (2009). How do you feel, now? The anterior insula and human awareness. Nature Reviews Neuroscience, 10(1), 59–70. https://doi.org/10.1038/nrn2555

D'Argembeau, A., Stawarczyk, D., Majerus, S., Collette, F., Van der Linden, M., Feyers, D., Maquet, P., and Salmon, E. (2010). The neural basis of personal goal processing when envisioning future events. Journal of Cognitive Neuroscience, 22(8), 1701–1713. https://doi.org/10.1162/jocn.2009.21314

de Bono, E. (1969). The Mechanism of Mind. Jonathan Cape.

de Lange, F. P., Heilbron, M., and Kok, P. (2018). How do expectations shape perception? Trends in Cognitive Sciences, 22(9), 764–779. https://doi.org/10.1016/j.tics.2018.06.002

Dechêne, A., Stahl, C., Hansen, J., and Wänke, M. (2010). The truth about the truth: a meta-analytic review of the truth effect. Personality and Social Psychology Review, 14(2), 238–257. https://doi.org/10.1177/1088868309352251

Declerck, C. H., Boone, C., Pauwels, L., Vogt, B., and Fehr, E. (2020). A registered replication study on oxytocin and trust. Nature Human Behaviour, 4(6), 646–655. https://doi.org/10.1038/s41562-020-0878-x

Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. Viking. https://www.penguinrandomhouse.com/books/308282/consciousness-and-the-brain-by-stanislas-dehaene/

Doshi, A. R., and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290

Drèze, X., and Hussherr, F.-X. (2003). Internet advertising: is anybody watching? Journal of Interactive Marketing, 17(4), 8–23. https://doi.org/10.1002/dir.10063

Fauconnier, G., and Turner, M. (2002). The Way We Think: Conceptual Blending and the Mind's Hidden Complexities. Basic Books. https://www.basicbooks.com/titles/gilles-fauconnier/the-way-we-think/9780465087860/

Fiske, S. T., Cuddy, A. J. C., and Glick, P. (2007). Universal dimensions of social cognition: warmth and competence. Trends in Cognitive Sciences, 11(2), 77–83. https://doi.org/10.1016/j.tics.2006.11.005

Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Gibbs, R. W. (2006). Embodiment and Cognitive Science. Cambridge University Press. https://doi.org/10.1017/CBO9780511805844

Green, M. C., and Brock, T. C. (2000). The role of transportation in the persuasiveness of public narratives. Journal of Personality and Social Psychology, 79(5), 701–721. https://doi.org/10.1037/0022-3514.79.5.701

Hohwy, J. (2013). The Predictive Mind. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199682737.001.0001

Hovland, C. I., Janis, I. L., and Kelley, H. H. (1953). Communication and Persuasion. Yale University Press. https://archive.org/details/communicationper0000unse

Jiang, R., Chiappa, S., Lattimore, T., György, A., and Kohli, P. (2019). Degenerate feedback loops in recommender systems. Proceedings of AIES 2019. https://www.aies-conference.com/2019/wp-content/papers/main/AIES-19_paper_187.pdf

Kable, J. W., and Glimcher, P. W. (2007). The neural correlates of subjective value during intertemporal choice. Nature Neuroscience, 10(12), 1625–1633. https://doi.org/10.1038/nn2007

Kahneman, D., and Tversky, A. (1979). Prospect theory: an analysis of decision under risk. Econometrica, 47(2), 263–291. https://doi.org/10.2307/1914185

Kimmerer, R. W. (2013). Braiding Sweetgrass. Milkweed Editions.

Klein, C. (2018). What do predictive coders want? Synthese, 195(6), 2541–2557. https://doi.org/10.1007/s11229-016-1250-6

Kosfeld, M., Heinrichs, M., Zak, P. J., Fischbacher, U., and Fehr, E. (2005). Oxytocin increases trust in humans. Nature, 435(7042), 673–676. https://doi.org/10.1038/nature03701

Lakoff, G., and Johnson, M. (1980). Metaphors We Live By. University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/M/bo3637992.html

LeCun, Y. (2022). A path towards autonomous machine intelligence. OpenReview. https://openreview.net/forum?id=BZ5a1r-kVsf

Levy, R., and Jaeger, T. F. (2007). Speakers optimize information density through syntactic reduction. Advances in Neural Information Processing Systems, 19, 849–856. https://proceedings.neurips.cc/paper/2006/hash/c6a01432c8138d46ba39957a8250e027-Abstract.html

Lewis, P., Perez, E., Piktus, A., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474. https://arxiv.org/abs/2005.11401

Li, L., Chu, W., Langford, J., and Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. Proceedings of WWW 2010, 661–670. https://doi.org/10.1145/1772690.1772758

Loewenstein, G. (1994). The psychology of curiosity: a review and reinterpretation. Psychological Bulletin, 116(1), 75–98. https://doi.org/10.1037/0033-2909.116.1.75

McQuarrie, E. F., and Mick, D. G. (1996). Figures of rhetoric in advertising language. Journal of Consumer Research, 22(4), 424–438. https://doi.org/10.1086/209459

McQuarrie, E. F., and Mick, D. G. (2003). Visual and verbal rhetorical figures under directed processing versus incidental exposure to advertising. Journal of Consumer Research, 29(4), 579–587. https://doi.org/10.1086/346252

Meister, C., Pimentel, T., Wiher, G., and Cotterell, R. (2023). Locally typical sampling. Transactions of the Association for Computational Linguistics, 11, 102–121. https://doi.org/10.1162/tacl_a_00536

Mercier, H., and Sperber, D. (2011). Why do humans reason? Arguments for an argumentative theory. Behavioral and Brain Sciences, 34(2), 57–74. https://doi.org/10.1017/S0140525X10000968

Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity. MIT Press. https://mitpress.mit.edu/9780262633086/being-no-one/

Meyers-Levy, J., and Tybout, A. M. (1989). Schema congruity as a basis for product evaluation. Journal of Consumer Research, 16(1), 39–54. https://doi.org/10.1086/209192

Michaelov, J. A., Bardolph, M. D., Van Petten, C. K., Bergen, B. K., and Coulson, S. (2024). Strong prediction: language model surprisal explains multiple N400 effects. Neurobiology of Language, 5(1), 107–135. https://doi.org/10.1162/nol_a_00105

Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D., and Finn, C. (2023). DetectGPT: zero-shot machine-generated text detection using probability curvature. Proceedings of ICML 2023. https://arxiv.org/abs/2301.11305

Northoff, G., Heinzel, A., de Greck, M., Bermpohl, F., Dobrowolny, H., and Panksepp, J. (2006). Self-referential processing in our brain: a meta-analysis of imaging studies on the self. NeuroImage, 31(1), 440–457. https://doi.org/10.1016/j.neuroimage.2005.12.002

Oh, B.-D., and Schuler, W. (2023). Why does surprisal from larger transformer-based language models provide a poorer fit to human reading times? Transactions of the Association for Computational Linguistics, 11, 336–350. https://doi.org/10.1162/tacl_a_00548

Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155. https://arxiv.org/abs/2203.02155

Padmakumar, V., and He, H. (2024). Does writing with language models reduce content diversity? Proceedings of ICLR 2024. https://arxiv.org/abs/2309.05196

Peters, J., and Büchel, C. (2010). Episodic future thinking reduces reward delay discounting through an enhancement of prefrontal-mediotemporal interactions. Neuron, 66(1), 138–148. https://doi.org/10.1016/j.neuron.2010.03.026

Petty, R. E., and Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. Advances in Experimental Social Psychology, 19, 123–205. https://doi.org/10.1016/S0065-2601(08)60214-2

Polich, J. (2007). Updating P300: an integrative theory of P3a and P3b. Clinical Neurophysiology, 118(10), 2128–2148. https://doi.org/10.1016/j.clinph.2007.04.019

Pornpitakpan, C. (2004). The persuasiveness of source credibility: a critical review of five decades' evidence. Journal of Applied Social Psychology, 34(2), 243–281. https://doi.org/10.1111/j.1559-1816.2004.tb02547.x

Rackham, O. (1986). The History of the Countryside. J. M. Dent. https://archive.org/details/historyofcountry0000rack

Rankin, C. H., Abrams, T., Barry, R. J., et al. (2009). Habituation revisited: an updated and revised description of the behavioral characteristics of habituation. Neurobiology of Learning and Memory, 92(2), 135–138. https://doi.org/10.1016/j.nlm.2008.09.012

Rao, R. P. N., and Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580

Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124(3), 372–422. https://doi.org/10.1037/0033-2909.124.3.372

Salvi, F., Horta Ribeiro, M., Gallotti, R., and West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour, 9(8), 1645–1653. https://doi.org/10.1038/s41562-025-02194-6

Schacter, D. L., Addis, D. R., and Buckner, R. L. (2007). Remembering the past to imagine the future: the prospective brain. Nature Reviews Neuroscience, 8(9), 657–661. https://doi.org/10.1038/nrn2213

Schmidt, S., and Eisend, M. (2015). Advertising repetition: a meta-analysis on effective frequency in advertising. Journal of Advertising, 44(4), 415–428. https://doi.org/10.1080/00913367.2015.1018460

Schultz, W., Dayan, P., and Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593

Schwartz, E. M. (1966). Breakthrough Advertising. Prentice-Hall.

Schwartz, E. M. (1994). The Brilliance Breakthrough: How to Talk and Write So That People Will Never Forget You. Instant Learning.

Schwarz, N. (2004). Metacognitive experiences in consumer judgment and decision making. Journal of Consumer Psychology, 14(4), 332–348. https://doi.org/10.1207/s15327663jcp1404_2

Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x

Sharp, B. (2010). How Brands Grow: What Marketers Don't Know. Oxford University Press. https://global.oup.com/academic/product/how-brands-grow-9780195573565

Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., and Anderson, R. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y

Smith, N. J., and Levy, R. (2013). The effect of word predictability on reading time is logarithmic. Cognition, 128(3), 302–319. https://doi.org/10.1016/j.cognition.2013.02.013

Summerfield, C., and de Lange, F. P. (2014). Expectation in perceptual decision making: neural and computational mechanisms. Nature Reviews Neuroscience, 15(11), 745–756. https://doi.org/10.1038/nrn3838

Tom, S. M., Fox, C. R., Trepel, C., and Poldrack, R. A. (2007). The neural basis of loss aversion in decision-making under risk. Science, 315(5811), 515–518. https://doi.org/10.1126/science.1134239

Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need. arXiv:1706.03762. https://arxiv.org/abs/1706.03762

Zhang, J., et al. (2025). Verbalized sampling: how to mitigate mode collapse and unlock LLM diversity. arXiv:2510.01171. https://arxiv.org/abs/2510.01171

Zhu, Y., Lu, S., Zheng, L., Guo, J., Zhang, W., Wang, J., and Yu, Y. (2018). Texygen: a benchmarking platform for text generation models. Proceedings of SIGIR 2018, 1097–1100. https://arxiv.org/abs/1802.01886