In 2023, game studio Riot Games published a detailed post-mortem on using AI tools during the development of League of Legends skin concepts. Their concept artists reported that AI image generators were most useful in the earliest divergent phase — generating twenty rough texture directions in minutes so that a human artist could immediately feel which ones resonated with their instinct. Artists were explicit: the ideas themselves still came from them; AI had accelerated the moment of reaction, not the moment of invention.
When the same artists tried using AI to produce near-final assets, the results required so much corrective redrawing that the time savings evaporated. The leverage existed in one narrow zone — and outside that zone it became friction.
The common framing — AI either helps or it doesn't — is too coarse. Where in your creative process you introduce AI determines almost everything about the outcome. Research on creative workflows distinguishes between divergent phases (generating options) and convergent phases (selecting, refining, committing). AI tools tend to provide most measurable lift in divergent phases, where volume and variety are valuable and where your judgment hasn't yet narrowed the field.
When you push AI into convergent phases — the moment you're committing voice, deciding what the work is actually about, doing the final edit that makes something uniquely yours — the calculus reverses. The tool's tendency to smooth and generalize actively works against the kind of commitment convergent work requires.
The leverage zone is real and repeatable. It's not a vague aspiration. Several documented cases from professional creative industries show consistent patterns about where AI multiplies effort versus where it costs more than it saves.
Pattern 1 — Volume at the start of ideation. In 2022, Spotify's internal design team documented that using AI tools to generate rough brief alternatives before any human designer touched a project reduced the time teams spent on initial alignment by roughly 30%. The AI wasn't designing — it was producing rough material fast enough that the team could react to real options instead of abstract descriptions.
Pattern 2 — Constraint testing. Several novelists who spoke publicly at the 2023 Nebula Conference described using AI to rapidly generate sentences under different constraints (different sentence lengths, different POV shifts, different tonal registers) to test which direction felt most alive. The AI generated the variations; the writer provided the recognition — the felt sense that one direction was right. This is a leverage-zone use because the writer's judgment is the operative force.
Pattern 3 — Research translation. Science journalists at outlets including Undark and Wired have described using AI to produce rough plain-language summaries of dense academic papers as a starting point, then rewriting completely. The AI provides a scaffold that saves the journalist from the blank-page problem on complex material, but the journalist's domain judgment, narrative instinct, and source relationships remain entirely irreplaceable.
AI is in the leverage zone when your judgment is the operative force and AI is producing raw material for that judgment to act on. AI has left the leverage zone when AI's output is the thing that ships — and your role has shifted to approving rather than creating.
The leverage zone is not identical for every creative. A novelist's leverage zone differs from a UI designer's, which differs from a filmmaker's. What remains consistent is the structural feature: leverage exists where volume and speed matter more than singularity, where reaction is more useful than blank-page generation, and where the work produced by AI is material for your process rather than output from it.
To locate yours, ask two questions about any proposed AI use: Am I generating options to react to, or am I accepting output to deliver? And: Is my judgment the thing that determines quality here, or have I delegated that judgment to the model?
Think about a creative project you're working on now, or one you've completed recently — writing, design, music, video, code, anything. Describe a specific phase of that project to the AI assistant and ask it to help you identify whether that phase is likely inside or outside the leverage zone. Push back, probe, and refine your understanding through conversation.
In April 2023, Sports Illustrated published articles under bylines of authors who did not exist — fabricated profiles, complete with AI-generated headshots from the service This Person Does Not Exist. The articles were competent in the mechanical sense: grammatically correct, structurally sound, formatted like sports journalism. What they lacked was specificity — the particular reporter's access, the source relationships, the observation that could only come from someone actually at the event.
When the deception was exposed by Futurism's Drew Ortiz in November 2023, readers described the articles less as bad writing and more as hollow — technically adequate but missing the texture that makes journalism worth reading. The content passed mechanical quality checks precisely because it had been trained on competent work. What it couldn't reproduce was idiosyncrasy, access, and the irreplaceable weight of a specific perspective.
Language models are trained on vast corpora of human writing and optimize for what statistically follows what. This produces prose that is, in the aggregate, smooth and competent. It also means the model tends toward the center of any distribution — the most common way something has been said, the most typical structure for this type of piece, the most expected emotional note for this kind of scene.
This is not a bug in AI systems. It is a direct consequence of how they work. The problem for creative practitioners is that distinctiveness often lives at the edges of distributions — in the choices that are characteristic of a specific writer or artist, the decisions that deviate from convention in ways that turn out to be more true, more resonant, more alive than the center would have been.
Flattening is not about quality dropping below a threshold. It's about quality converging on a mean that is technically acceptable but distinctively no one's.
A 2024 study published in Science Advances examined nearly one billion Amazon product reviews over twelve years and found a significant increase in language patterns consistent with AI generation after the release of ChatGPT in late 2022. More importantly, the study found that the reviews that increased in AI-like language were less useful to other consumers — rated lower for helpfulness — even when rated higher for grammatical correctness. Readers could sense the absence of specific experience even when they couldn't identify the mechanism.
The same dynamic appears in creative work. When a piece is generated or heavily shaped by AI, readers often describe a vague dissatisfaction they struggle to articulate — a sense that the work is hitting all the expected notes while somehow missing the thing that would make it worth re-reading. The literary term for this is voice, and voice is precisely what AI tends to average out of existence.
Structure, grammar, format conventions, genre expectations, common phrases, appropriate register, sentence variety, factual accuracy (when trained data is accurate), pacing templates.
Idiosyncratic word choices, earned awkwardness, specific observation, the unexpected metaphor that works, structural risk-taking, the silence before a sentence that gives it weight, point-of-view as lived experience.
A subtler version of flattening occurs gradually. Designer and writer Robin Rendle, writing publicly in 2023, described noticing that after months of using AI writing tools for first drafts, he found himself less certain about his own instincts on sentence-level choices. The tool hadn't replaced him — it had gradually eroded his trust in his own ear. He described having to spend weeks writing deliberately without AI assistance to recover the feel of his own voice.
This pattern — reduced confidence in one's own creative judgment following sustained AI use — has been described by multiple practitioners publicly. It suggests that the flattening effect is not only about the quality of individual outputs; it may also affect the creative practitioner's own capacity over time if AI is used in phases that should be developing and exercising that practitioner's judgment.
Ask yourself: When did I last make a creative decision that surprised me? If you're struggling to remember, it may be worth auditing whether AI has moved into phases of your work where your own judgment should be the thing getting exercised and strengthened.
Share a short piece of writing with the assistant — either your own AI-assisted draft, a sample you've come across, or describe a piece you're thinking about. Ask the assistant to analyze it for signs of the flattening effect: phrases that tend toward the center, structural choices that feel conventional rather than felt, moments where a more specific observation was available but a generic one was used instead.
Musician and producer Holly Herndon has been one of the most publicly articulate practitioners about AI and creative identity. In interviews with The Guardian and Wired in 2022 and 2023, Herndon described a specific process structure she developed: she uses AI tools only on stems and fragments that have already passed through her own compositional instinct — material she has already made decisions about. The AI works with her material, not before it.
Herndon contrasted this with what she called "prompt and accept" workflows where the AI produces and the artist selects from AI output — a structure she argued inverts the creative relationship. In her framing, the question was not whether to use AI but at what point in the causal chain AI enters the work. Her guardrail was structural: AI had no access to the blank canvas. It only ever encountered work-in-progress that already carried her fingerprint.
Research and practitioner testimony converge on three structural approaches that preserve creative distinctiveness while using AI effectively. These are not rules about when AI is morally acceptable — they're practical structures that protect the quality of the work itself.
Always produce your own rough version before consulting AI. Even a terrible, incomplete, unpolished version that captures your instinct provides a fingerprint the AI can then work with rather than replace. The first move is always yours.
Ask AI to generate versions that are deliberately different from yours — not better versions, but versions you can react against. Disagreement with AI output activates your own judgment more reliably than accepting or refining AI output.
After any AI-assisted session, spend time with the output and ask: which parts of this feel like mine, and which parts feel like something I would not have chosen? Actively discard the second category, even if it's technically good.
Periodically write without AI assistance and compare to your AI-assisted work. If you notice the AI-assisted work has begun to sound more consistent with itself than with your logged voice, that's a signal to pull back on AI in convergent phases.
In 2023, the Recording Academy's task force on AI published a set of practitioner principles developed with working musicians, including members of Radiohead's team and producers from the Black Music Collective. The principles included a structural recommendation that has since been widely cited: "AI should enter the production chain only after the musical identity of a track has been established by human creative decisions."
This is essentially Guardrail 1 formalized as professional practice. The logic is that once a track's core identity — its harmonic language, rhythmic character, emotional register — has been determined by human decision-making, AI can operate on that established identity without erasing it. It can extend, process, and expand without replacing the foundational act of creative choice.
The pattern is consistent across fields: visual artists who describe effective AI use almost always describe it as working with established compositions, color palettes, or visual languages they've already developed. Writers who describe effective AI use describe it as expanding from material they've already written, not generating material from scratch. The common structure is: human establishes identity, AI operates within that established identity.
Holly Herndon's framing is precise and useful: at what point in the causal chain does AI enter your work? If AI enters before you've made any decisions, you're building on AI's foundation. If AI enters after you've made foundational decisions, AI is building on yours. Only the second structure keeps you as the creative author.
It's also worth noting that guardrails applied too rigidly become their own form of creative limitation. If you spend so much energy protecting your voice from AI contamination that you never actually use AI for the genuine leverage it provides, you've overcorrected. The goal is not AI abstinence — it's intentional use. Guardrails exist to keep you in the leverage zone, not to wall you off from it.
Describe your creative medium and typical workflow to the assistant. Work with it to design a specific guardrail system tailored to your practice — not generic advice, but a set of concrete structural rules that fit how you actually work. Push back on suggestions that don't fit, and ask for alternatives until you have something actionable.
In early 2024, visual artist and researcher Lev Manovich — whose work on cultural analytics spans decades — published a widely read essay arguing that the most significant risk of generative AI for creative practitioners was not plagiarism or economic displacement but what he called "stylistic amnesia": the gradual loss of a practitioner's relationship to their own aesthetic history. Manovich documented his own process of experimenting with AI image tools and noticing that after extended use, he found it progressively harder to distinguish between what he would have chosen and what the AI had normalized him toward choosing.
His prescription was specific: maintain a body of work produced entirely without AI assistance as an ongoing reference — not as a rejection of AI, but as a calibration instrument. He called this the practitioner's "immune system."
The single-session version of the flattening question is: does this particular output sound like me? The career-length version is: after five years of AI-assisted work, does my body of work still have a recognizable aesthetic trajectory that is mine? These are related but different questions, and the second is harder to answer in real time because the drift is gradual.
Calibration drift occurs when a practitioner's sense of what their work should sound or look like has gradually shifted toward AI's defaults — not through a single bad session but through thousands of small acceptances of AI-adjacent choices. Each individual choice seemed reasonable. The aggregate effect is a voice that has been averaged toward the center.
The challenge is that calibration drift is largely invisible while it's happening. It becomes visible retrospectively — when you look at work from three years ago and can't produce anything that feels like it anymore. At that point, recovery is possible but takes significant deliberate effort.
Regular AI-free periods. Multiple practitioners who have spoken publicly about sustained AI use — including novelist Jonathan Lethem in a 2023 Paris Review interview and composer Max Richter in interviews with The Wire — have described maintaining deliberate periods of working without AI assistance. Not as a rejection, but as a way of keeping the unconditioned voice active and available. Richter described this as analogous to a musician practicing acoustically even when their primary performance context is electronic.
Retrospective audits. Looking back at work from before significant AI use and asking: what choices am I no longer making that I used to make? What tendencies have disappeared from my work? This audit is not always comfortable, but it provides the clearest signal about whether drift has occurred.
Specificity as a discipline. AI tends toward the general. Training yourself to push toward specificity — the specific smell, the specific gesture, the specific chord voicing — actively works against AI's averaging tendency even in AI-assisted work. Where you insert your specific, irreducible observation, the flattening effect cannot reach.
Maintain a body of AI-free work as an ongoing reference point — not to reject AI, but to keep a calibrated sense of your own aesthetic baseline. This is the immune system: it doesn't prevent you from encountering AI, it keeps you able to recognize when something foreign has been normalized.
The practitioners who have sustained distinctive creative identities through technological transitions — from analog to digital, from pen to word processor, from film to digital photography — tend to share a structural characteristic: they adopted the new technology for the leverage it provided while maintaining active engagement with the practices that built their aesthetic identity in the first place.
Photographer Andreas Gursky described digital editing as something he uses to realize visions he had before sitting at the computer — not to discover visions he wouldn't have had otherwise. This is a career-length version of the causal chain principle: AI and digital tools are downstream of a creative vision that has its roots in practice that predates and is independent of those tools.
The question for practitioners using AI now is: what are the practices that built your aesthetic identity — and are you maintaining them? If the answer is yes, AI can be a long-term amplifier. If those practices have been largely replaced by AI-assisted workflows, the risk of permanent calibration drift is real.
Use this lab to conduct a structured retrospective audit of your creative practice. Describe to the assistant how your work has evolved over the past year or two, what AI tools you've introduced, and what you've noticed about your own instincts and aesthetic choices. The assistant will help you identify calibration drift signals and build a maintenance plan using Manovich's immune system concept and the long-game practices from this lesson.