TL;DR: The AI-disclosure-tanks-favorability research is real. Schilke and Reimann ran thirteen experiments and the trust penalty held across every one. The 99th Academy Awards just made AI-generated actors and screenplays ineligible. Both responses come from the same category error: treating visible AI assistance as morally distinct from the invisible human assistance (speechwriters, ghostwriters, comms teams, PR departments, polishers) that bylines, press releases, quotes, and speeches have been built on for centuries. Voice-as-a-skill is a public meta-skill I shipped today and you can run it on your own samples.
The Same Ad, Twice
A 2025 study by Oliver Schilke and Martin Reimann at Arizona ran thirteen experiments on AI disclosure. Show subjects a piece of communication (a memo, an analysis, a creative) and ask them to rate it. Show another set of subjects the same artifact, but tell them AI helped produce it. The second group rates it lower every time. Not “sometimes.” Every time. Across supervisors, subordinates, professors, analysts, and creatives. Whether disclosure is voluntary or mandatory. Whether the rater has favorable views of technology or skeptical ones. The penalty attenuates with high-AI-accuracy beliefs. It does not disappear.
That study did the cleanest job, but it isn’t alone. A separate thirteen paper meta-analysisthree-experiment study on AI in advertising shows the same pattern in commercial contexts. Identical ad, blind comparison, A and B. People prefer A when no one mentions AI. Tell them about A, they switch to B.
The honest read on this is not that consumers are wrong. It’s that we have a public moral stance on AI assistance that isn’t reciprocated by our private one. We dislike the AI ad until we forget it’s an AI ad, and then we like it fine. The discomfort lives in the disclosure, not in the artifact.
What We Already Tolerate
Bylines were ghostwritten in the third century first century BC. Cicero had Tiro. Modern presidents have speechwriters with first names you’d recognize from White House memoirs. Most CEO op-eds in the Wall Street Journal are drafted by a comms team. Most “thought leadership” content under an executive’s name is a writer’s draft with the executive’s edits. Most quoted spokespeople in press releases are reading lines another human wrote and approved on a call that ran twelve minutes over. None of this is a secret, and none of it is a scandal. The mechanism is so common we don’t have a clean public word for it. We have ghostwriting, which sounds shameful, and we have editorial, which sounds anodyne. Same act. Different framing.
The reason it isn’t a scandal is that the byline carries something specific: a commitment to the voice. The CEO doesn’t write the speech. The CEO owns the speech. If a line in it is wrong — strategically, ethically, factually — the CEO eats it. The writer worked under direction, the executive read every word, the executive can defend it in a hostile room. The voice is the ownership signal. The byline says this came through me, and I’m staking my reputation on the contents.
That’s the convention we’ve quietly run for centuries. AI is the same mechanism made visible, and the visibility is what’s triggering the moral panic.
To be precise, it’s the move I want to defend. We are not coming up with a new ethics. We are reacting badly to seeing the old one for the first time. The Schilke study is measuring the size of that reaction. The Academy is institutionalizing it. The AI ads research is pricing it. None of that means assistance has changed character. It means assistance got loud.
What the Speechwriters Knew
A short, partial list of lines you can probably finish in your head, and the writer who actually drafted them:
“The only thing we have to fear is fear itself.” Franklin D. Roosevelt, First Inaugural, 1933. Raymond Moley wrote the body of the speech; Louis Howe inserted the famous line as a late edit, possibly cribbed from a department-store advertisement that had run that week. FDR kept it and rewrote the ending of the sentence.
“Ask not what your country can do for you.” John F. Kennedy, Inaugural, 1961. Ted Sorensen drafted it. Sorensen spent the rest of his life politely declining to confirm or deny which lines were his and which were Kennedy’s, on the grounds that the byline was the President’s, not the speechwriter’s.
“We choose to go to the Moon, not because it is easy, but because it is hard.” Kennedy, Rice University, 1962. Sorensen again.
“Mr. Gorbachev, tear down this wall!” Ronald Reagan, Brandenburg Gate, 1987. Peter Robinson drafted the line. The State Department, Chief of Staff Howard Baker, and National Security Advisor Colin Powell tried three separate times to cut it from the speech text. Robinson and the other speechwriters elected not to pass the speech through the chain of command, knowing it would get edited out of existence. Reagan overruled the rest of his administration and kept it.
“A thousand points of light.” George H. W. Bush, Republican National Convention nomination, 1988. Peggy Noonan.
“Yes we can.” Barack Obama, 2008 New Hampshire primary concession. Jon Favreau on the page; the four-to-five-round drafting back-and-forth between Favreau and Obama was the actual creative process. Favreau is on record saying Obama is the better writer.
The byline of every one of those lines is the President. Nobody calls any of them dishonest. We accept the convention because the leader committed to the words in public, on the record, and bore the consequences when the words landed wrong. Sorensen’s declining-to-confirm posture is the cleanest articulation of the contract: the writer’s craft, the principal’s voice, the principal’s accountability.
The 99th Oscars
In May 2026, the Academy of Motion Picture Arts and Sciences adopted new rules for the 99th ceremony covering films released in 2026. Two of them speak to this directly. Performances must be “credited in the film’s legal billing and demonstrably performed by humans with their consent.” Screenplays must be “humanly created.” Producers must disclose generative AI used in scriptwriting, visual effects, and sound design.
The Academy paired the rule with language saying AI tools “neither help nor harm” the chances of a nomination, and that voters will judge “the degree to which a human was at the heart of the creative authorship.” That last clause is the one I want to defend before I take the rest apart. There is something specific about creative artistic labor that the Academy is trying to protect: the idea that a film is the residue of a human being’s choices, that the choices reveal a sensibility, and that the sensibility is what’s being honored. I take that seriously. It’s why I write a Substack instead of running a content farm.
But the line the Academy drew (visible AI versus invisible human assistance) doesn’t hold up under any pressure at all. Most major-studio screenplays are punched up by uncredited writers, sometimes a dozen of them, working in serial passes with a producer’s notes between each. Actors deliver lines that other writers wrote, with directorial coaching, with rehearsal scaffolding, with re-recorded ADR after the fact. The credited screenwriter on a tentpole release is, in many cases, the last screenwriter on a tentpole release. The Academy already lives on assisted authorship. It’s the entire industrial premise.
The new rule punishes visibility, not assistance. A film with twenty uncredited human polishers is a normal Oscar contender. A film with one disclosed generative-AI pass is now ineligible in the affected categories. The asymmetry is the tell.
I don’t think the Academy is acting in bad faith. I think they are doing what large institutions always do when a new technology forces an old practice into the open: they pick a defensible-sounding line in the wrong place, mistake it for a craft defense, and discover in eight years that they drew it under conditions they no longer recognize. Everyone who’s read the history of the Hays Code or the early MPAA ratings will know the shape of this.
The right defense of craft is not to ban a tool. It’s to demand a voice.
What Idea Flow and the Jagged Frontier Actually Show
Before the prescription, two pieces of evidence about why this question is louder this year than last.
The first is Jeremy Utley and Perry Klebahn’s framework from Stanford’s d.school. Their book, Ideaflow: The Only Business Metric That Matters, defines idea flow as the number of novel ideas a person or team can generate around a problem in a given amount of time. Their argument, distilled across years of teaching at the d.school: most organizations are not idea-poor. They are production-poor. The constraint isn’t getting to a good idea. The constraint is the human capacity to build, test, and ship around the ideas they already have.
That framework predicts what AI does to a competent writer’s workflow. It doesn’t write better thoughts than the writer. It removes the production bottleneck around the writer’s existing thoughts. The writer who had four good arguments and shipped one of them per quarter now ships three.
The second piece of evidence is the Dell’Acqua / Ethan Mollick / Lakhani / BCG study from Harvard, Navigating the Jagged Technological Frontier, now formally published in Organization Science (2026) after running as a working paper since 2023. 758 BCG consultants. Three groups: no AI, GPT-4 access, GPT-4 with a prompt-engineering primer. On tasks within AI’s capability frontier, the AI-assisted consultants were 12.2% more productive, finished 25.1% faster, and produced output rated more than 40% higher quality than the control. On tasks outside the frontier, the assisted group did worse. The boundary is jagged. Workers can’t always tell which side of it a given task sits on.
A follow-on field study from the same team, Cyborgs, Centaurs and Self-Automators (HBS Working Paper 26-036, December 2025), gave the question empirical shape. Same setup, this time tracking how 244 BCG consultants actually integrated AI across a seven-stage problem-solving workflow. Three modes emerged. Cyborgs (~60%) fused with the AI across every stage of the work, probing, redirecting, taking advice on some moves and pushing back on others — the writer is in the room throughout. Centaurs (~14%) ran structured, directed queries against the AI as a targeted tool, knowing what they wanted and asking specifically — the writer is in the room intermittently. Self-Automators (~27%) handed the task off and accepted what came back — the writer left the room. The disclosure backlash, the Academy’s rule, and the AI ads research are all really measuring fear of the third mode. Voice-as-a-skill is the operational form of the first two.
Read those findings together and the picture is specific. AI raises the production capacity around competent human judgment when judgment is in the room, and lowers quality when it isn’t. The disclosure backlash is measuring the fear that judgment isn’t in the room. The question isn’t whether AI helped. The question is whether the writer was there.
Voice as a Skill
Today I shipped a meta-skill called voice-as-a-skill. It’s a public repo. It’s an interviewer that walks a first-time agentic-system user (Cowork, Claude Code, Codex, anything Skills-aware) through producing a v1 voice scaffold of their own. Eight phases. Identity, modes, sample harvest, pattern extraction, tropes (the avoid-list, half the value), tensions (the patterns that overlap with bans because they’re also signature moves), worked examples, generation. The scaffold gets installed; the user runs harvest cycles for eight weeks; the voice compounds.
The skill refuses to proceed without writing samples. That’s deliberate. Without samples, the agent pattern-matches on its own training distribution and calls the result “you.” That isn’t voice. That’s slop with your name on it.
The discipline matters because it answers the disclosure question directly. If AI helped write this, in what sense is it mine? It’s mine because the patterns are mine: extracted from my own writing, tested against my own samples, bounded by my own avoid-list, harvested over months of use. The agent ran the codified pattern. I committed the pattern. The voice is the commitment. The output is mine the same way a CEO’s keynote is the CEO’s, the same way a Reagan speech was Peggy Noonan’s craft and Reagan’s commitment, the same way a press release is the spokesperson’s accountability and the comms team’s draft.
Ruben Hassid, who runs How to AI on Substack and built EasyGen, has been writing about the ghostwriting-versus-AI tension for the better part of two years from his own seat. He ghostwrites for a network of LinkedIn principals and now runs an AI-assisted version of the same craft. His how to finally stop writing like AI anti-AI style guide is a working artifact in this space, and the argument I’m making here owes him a tip of the hat. The version I’m landing on is downstream of his.
A small acknowledgment of the structural-tells problem before I close. There is still bad AI-assisted writing: flat lists, generic examples that prove nothing, restated conclusions that follow from nothing, no shape. I wrote about those four tells in a separate outline last week, leaning on Srinivas Rao’s work. They diagnose absence-of-voice, not presence-of-AI. Self-Automators leave traces; Cyborgs and Centaurs don’t. A piece can pass every word-level cliché test (”delve,” “tapestry,” “navigate the landscape”) and still read like nothing, because nobody chose anything. The fix is the same: have a voice; commit to it; encode it; run it.
What This Doesn’t Solve
A few things this argument doesn’t dissolve. I want to name them because the position will be stronger if I do.
The disclosure backlash is real even if it’s irrational, and “irrational” is not a marketing strategy. If you ship a campaign and disclose the AI work, your brand will take a hit. The Schilke study is the price tag, not the verdict. Pricing the bias is not the same as defending it.
The Academy’s craft instinct is also real, even if I think they drew the line wrong. There is a difference between an actor whose performance is a human being’s labored choice and a synthetic actor generated from a prompt. I’m comfortable saying both should be eligible if the credited human director, writer, and producer take responsibility for the result. The Academy is not yet comfortable with that. That’s a defensible read, even if it’s the wrong one.
And there is a category of AI use that is dishonest: a writer who claims to have hand-drafted a piece they prompted in fifteen minutes, no harvest, no voice file, no editing pass. That’s not assisted authorship. That’s misrepresentation. The line isn’t did AI help. The line is did the writer commit to the result. Voice-as-a-skill makes the commitment legible (receipts, samples, harvest cycle, a public repo) without requiring a paragraph of throat-clearing in every post.
Three Things I Cut From This Piece
To prove the discipline I’m preaching.
A long detour on Cicero, Tiro, and the Roman tradition of the amanuensis as the original ghostwriter. Interesting, slowed the argument, belongs in a different piece. A direct comparison between watermarking standards (C2PA, Adobe’s Content Credentials) and disclosure norms. The right question, but a structural-tells essay isn’t the place to litigate provenance infrastructure. A defense of disclosure in narrow cases (deepfakes, political ads, financial advice) where the asymmetry is real and the assistance is doing the deceptive work, not just the production work. The defense holds. It doesn’t change this argument; it bounds it.
Close
The disclosure penalty is human and mostly wrong. The Academy’s rule is well-intentioned and mostly wrong. The AI-assistance moral panic is loud and mostly displaced. What people are reacting to is the visibility of an old practice, not the arrival of a new one.
The right response is the response we’ve already been running for centuries. Have a voice worth committing to. Commit to it in public. Stake your reputation on the result. The pen, whether yours, your speechwriter’s, your comms team’s, or your Cowork session’s, is the medium. The byline is the contract.
Voice-as-a-skill is one operational form of the contract. There will be others. They will all rest on the same instinct: when the work lands, lands. When it doesn’t, eat it.
Good is good. The question that should be loud is who has the voice and the agency to commit to it. Not who held the pen.
Receipts
The disclosure research
Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes. — thirteen experiments, the canonical citation for this argument
Service ads in the era of generative AI: Disclosures, trust, and intangibility (Journal of Retailing and Consumer Services, 2025) — the parallel finding in commercial advertising
The Academy ruling
Variety — Oscar Rule Changes: AI Crackdown, Acting Nominations, International Film (May 2, 2026)
Academy Press Office — Awards Rules and Campaign Promotional Regulations Approved for 98th Oscars
TechCrunch — AI-generated actors and scripts are now ineligible for Oscars (May 2, 2026)
Production capacity and the jagged frontier
Utley, J., & Klebahn, P. (2023). Ideaflow: The Only Business Metric That Matters. Penguin Random House.
Dell’Acqua, F., McFowland III, E., Mollick, E., et al. (2026). Navigating the Jagged Technological Frontier. Organization Science. — formal publication of the BCG / Harvard study; the SSRN working paper is the prior-art version
Randazzo, S., Lifshitz-Assaf, H., Kellogg, K. C., Dell’Acqua, F., Mollick, E. R., Candelon, F., & Lakhani, K. R. (2025). Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise. Harvard Business School Working Paper 26-036. — the empirical follow-on; 244 BCG consultants, three collaboration modes
Prior art on the AI authorship question
Ruben Hassid — How to finally stop writing like AI (LinkedIn anti-AI style guide)
Srini Rao — Unmistakable Media (Substack) — the four structural tells
This piece’s operational form
voice-as-a-skill on GitHub — the public repo, the meta-skill, the eight-phase interview, the harvest cycle template
Editor's note, May 20, 2026. The morning after this published, an anonymous reader emailed a critique of it. The sharpest line: that the piece "cites studies the way a confident-but-not-careful person does."
Partly fair. The reader caught a real error: I had placed Cicero and his secretary Tiro "in the third century." They belong to the first century BC, off by two centuries, exactly as the email said. That one is fixed.
But the same email went after a citation that was fine. It flagged the repeated "thirteen experiments" as a tell of careless sourcing. That number is right; Schilke and Reimann did run thirteen experiments. What the critique missed is that a different number was wrong. I had called a three-experiment advertising study a "thirteen-paper meta-analysis," the "thirteen" carried over from the Schilke study just above it. That one is fixed too.
A critique about careless reading that read carelessly: it caught the small error, called a correct citation sloppy, and missed the one that was actually wrong. I'll take the trade. This piece argues that voice is the commitment, and commitment includes the part where the work doesn't land and you eat it. Two corrections, eaten.
And the critique arrived with no byline. This essay argues that the byline is the contract: you put your name on your words, in public, and carry what they cost. The reader diagnosed my voice at length and signed nothing. By the standard the essay defends, that isn't an indictment. It's an abstention. That's the next piece.