The transparency penalty: why disclosing AI costs you more than using it

Hire a human being to write your LinkedIn posts and almost nobody objects. Say a language model helped you draft one and the same audience marks you down for it. That gap is not a moral position anyone reasoned their way into. It is a measured effect, it now has a name in the research literature, and it should change how you build, disclose and price this work.

The name comes from Nakano, Takezawa, Matulic, Yang and Yatani, whose paper on how reader perception shifts once AI authorship is disclosed (arXiv:2510.24011, 23 January 2026) calls it the transparency penalty. The finding underneath the label is not one study or one field. Separate experiments, run by different teams in different research areas, keep arriving at the same uncomfortable place: tell a reader that AI was involved and their trust in the writing falls, whether or not the writing is any worse.

Meanwhile the older version of the same practice, a person writing under somebody else's name, sits comfortably inside professional norms. A survey of 291 PR professionals by the Institute for Public Relations found a general industry consensus in favour of undisclosed organisational ghost blogging, provided the content originates from the executive and the executive approves it, though a substantial minority disagreed. Same output, same reader, two completely different standards. What follows is what the evidence actually shows, exactly where it stops, and what I think an honest practitioner does with the gap.

What is the transparency penalty?

The name is recent. The underlying work is older, and it sits across three research communities that do not usually read each other's journals. Start with the strangest of them. A within-subjects experiment published in the Journal of Science Communication in 2026, with 433 participants, tested what an AI disclosure does to how believable a piece of content seems.

Read the second half of that finding again, because it is the part people skip when they quote it. The label did not help readers sort true from false. It pushed their judgement in the wrong direction on both sides of the line, penalising accuracy and rewarding error. Whatever a disclosure is doing inside a reader's head, it is not working as information about quality. It is working as a signal about whether an author was present, and readers respond to that signal before they evaluate a single claim.

The second study is from news. Toff and Simon, published by Sage in 2025, looked at what an AI label does to audience trust, and crucially they separated trust from accuracy.

The third comes from advertising, where the commercial stakes sit closest to ours. Note that the damage there did not stop at the message. It reached the organisation behind it.

There is a fourth paper in the same area, 'The Transparency Dilemma: An Experiment on How AI Disclosures Affect Credibility Perceptions and Engagement Across Topics', presented at AAAI/AIES in 2025, which put the same question across a spread of subject matter instead of inside one field. Three findings, in three research communities, pointing the same way. That is not a fluke and it is not a feeling. It is the closest thing this whole argument has to a settled fact.

Across three published research settings, news (Toff and Simon, Sage, 2025), science communication (Journal of Science Communication, 2026) and advertising (Journal of Interactive Advertising, 2025), disclosing AI involvement lowered trust in the content even where the content itself was judged no less accurate.

	Field	What it found	Tested on B2B buyers?
Journal of Science Communication, 2026	Science communication	Credibility of correct information fell after disclosure	×
Toff and Simon, Sage, 2025	News	Trust fell while accuracy judgements did not	×
Journal of Interactive Advertising, 2025	Advertising	Trust in the ad and the organisation fell	×
AAAI/AIES, 2025	Multiple topics	Experiment on credibility perceptions and engagement	×
B2B thought leadership	Commercial content	No published test exists	×

Why does a human ghostwriter get a pass when an AI assistant does not?

The Institute for Public Relations put this question to practitioners directly, in a study whose title asks whether ghost blogging is like speechwriting. The answer that came back was, broadly, yes.

Now look at the conditions attached, because they are the entire thing. The consensus was not that ghostwriting is fine. It was that ghostwriting is fine when the content originates with the executive and the executive approves it. And a substantial minority of those 291 still disagreed, which is worth carrying around with you rather than filing away.

A Hindi film credits a lyricist, a composer and a playback singer, and the person you are watching on screen is lip-syncing to somebody else's voice. Nobody in the hall feels defrauded. The division of labour is understood, the performance is still the star's responsibility, and if the song lands badly the star wears it. That is the shape of the ghostwriting bargain: many hands, one accountable name.

Here's the thing. The reason AI breaks that bargain is not the reason people say out loud. It is not that a machine touched the words, because a spellchecker is also a machine touching the words and nobody has ever demanded a footnote for it. It is that a machine can now produce the substance as well as the sentences, and once that is possible the reader loses the one guarantee the old arrangement gave them. A ghostwriter's presence proved that somebody, somewhere, held an opinion. AI removes that proof. The penalty is aimed at absent authorship, and disclosure of tooling is only the nearest available proxy for it.

The reader is not policing your tools. They are checking whether anyone was home when the thing was written.

Does the transparency penalty apply to B2B thought leadership?

Nobody knows, and I want to be blunt about that, because the temptation to pretend otherwise is enormous and the whole argument of this piece would collapse under it. Every published test of this effect sits in news, science communication or advertising. Not one of them tested a founder's LinkedIn post, a company blog, or anything a buying committee would read on the way to a purchase order. Applying those results to B2B content is extrapolation. It may well turn out to be correct extrapolation, but it is not evidence, and a piece arguing for honest sourcing cannot fudge that in its own defence.

There is also a specific reason to think B2B might behave differently, and it is not wishful thinking.

A person who spent the morning asking a model to summarise vendor reviews is not the same subject as a news reader meeting an AI label on a report about a flood. One of them already has AI inside their own process and a purchase decision with consequences attached to it. The other is being asked, in effect, whether a stranger deserves belief. I would not assume the second group's reaction transfers cleanly onto the first.

1. 8 July 2025
   Edelman and LinkedIn publish the 2025 B2B Thought Leadership Impact ReportNearly 2,000 management-level professionals; 79% more likely to advocate for a vendor's proposal in an RFP if that vendor consistently produces high-quality thought leadership
2. 12 November 2025
   6sense publishes its 2025 Buyer Experience Report4,000+ buyers; 94% used LLMs to summarise reviews or analyse data during the buying process
3. 23 January 2026
   Nakano and colleagues post their paper on reader perception shifts upon disclosure of AI authorshiparXiv:2510.24011, the paper that names the transparency penalty
4. 30 July 2026
   LinkedIn ships a 'Seems like AI slop' report buttonAlso discontinues its AI 'enhance your post' feature, replacing it with a proofreader that does not alter the writer's voice
5. 2 August 2026
   EU AI Act Article 50 transparency obligations become applicableMarking and detection obligation on systems already on the market deferred to 2 December 2026

Does the EU AI Act force you to label AI-assisted posts from 2 August 2026?

No, on the European Commission's own reading, and the panic being sold around this date deserves to be named as a product. Consultants have spent months telling executives and ghostwriters that from August 2026 every AI-assisted LinkedIn post carries a mandatory disclosure. The Commission's own Article 50 guidance says something quite different.

Three things in that guidance narrow the obligation far more than the headlines suggest.

• The Article 50(4) text-labelling duty covers text published to inform the public on matters of public interest: politics and democratic processes, public administration and services, the administration of justice, and law enforcement. Commercial and professional content generally is not what the provision is aimed at.
• Text that has undergone 'human review or editorial control' does not need to be labelled. The Commission defines that review as deliberate examination of the substance by a person with relevant knowledge, which is not a glance and a click.
• Purely personal, non-professional use falls outside the definition of a deployer entirely.

So the regulation everybody is frightened of arrives at the same destination as the trust research, from a completely different direction. Both say the thing that resolves the problem is a knowledgeable human examining the substance. Not a badge, not a footer, not a hashtag on the fourth line. The law treats substantive review as what discharges the obligation, and readers treat it as what earns belief. When the lawyers and the psychologists agree, it is usually worth building your process around the overlap.

The European Commission's Article 50 guidance limits the AI text-labelling duty to content published to inform the public on matters of public interest, and states that text which has undergone human review or editorial control does not need to be labelled.

• EUR 15M or 3% of worldwide annual turnoverMaximum Article 50 penalty
• 2 August 2026Transparency obligations apply from
• 2 December 2026Grace period for systems already on the market ends

What LinkedIn actually did on 30 July 2026

Platforms do not usually delete features that increase posting volume. LinkedIn did exactly that, and it did it in the same announcement where it handed every member a button for reporting the output of features like it.

The company's policy post, 'Keeping conversations real on LinkedIn', written by executive editor Laura Lorenzetti, is careful to say that using AI to help you write is acceptable as long as posts represent your own voice and perspectives, and that overuse at scale and in an automated way dilutes what real conversation produces. LinkedIn also says it has systems trained to recognise signals of AI slop. Chief Product Officer Hari Srinivasan called the problem a top priority. What LinkedIn has not published is a single number: no ranking weights, no measured reach impact by content type, nothing that would let anyone calculate what any of this costs you.

That absence has not stopped the market from supplying numbers anyway. You will see it claimed that LinkedIn's authenticity update reduces AI content reach by up to 47%, or that AI posts get 45% less engagement. No LinkedIn source exists for either figure. Both appear only in vendor and agency blogs with no stated methodology, which means somebody is selling protection against a penalty they invented. If either number has appeared in a deck you sent this year, that is a good place to begin your own audit, and it will be more useful than anything you do about disclosure.

• LinkedIn's writing assistant'Enhance your post' rewrote your draftA proofreader that fixes grammar and spelling without altering your voice
• Reader reportingNo AI-specific report option'Seems like AI slop' reporting on posts and comments
• Author feedbackNonePrivate dashboard flag when people believe your content reads as inauthentic
• IdentityExisting verificationExpanded profile verification

If detectors decide who gets penalised, who actually pays?

Every enforcement mechanism in this argument, platform side, employer side, reader side, eventually reduces to somebody deciding that a piece of writing looks like AI. That decision is increasingly made by detection tools. The academic evidence on those tools is not comforting for anyone quoting them.

Sit with the second number for a moment. Those essays were written by human beings, and the detectors called them machine-made 61.3% of the time, while returning near-zero false positives on native English writing. The difference being measured there is not AI use. It is which English you learned first.

LinkedIn is a global platform, and English is not the first language of everyone writing on it. So a penalty administered through detection does not land evenly across the people it touches. It lands hardest on the writer whose sentences run slightly more formal than a native speaker's, whose vocabulary arrived from a textbook rather than a dinner table, who learned the language in a classroom and writes it carefully because of that. That person is not using AI more than anyone else. They are simply easier to accuse.

The penalty does not fall on AI use. It falls on the appearance of it, and the appearance is judged by tools with a documented bias.

This is also why the single most-shared statistic in the entire AI content argument deserves far less respect than it currently gets. Originality.ai published a study in July 2026 reporting that 81% of long-form LinkedIn posts are likely AI-generated, and it has been repeated everywhere since by people who did not open it. The company sells AI detection, so the finding doubles as its own advertisement. The sample was 5,000 posts drawn from 10 pages of results across 90 topic-and-date searches, which the authors themselves concede is narrower than a feed-based study, and which over-represents search-shaped content by construction. The scoring ran on the publisher's own detector at a 15% AI allowance threshold. And the study page discloses that significant AI editing was used on the post itself.

Originality.ai's July 2026 finding that 81% of long-form LinkedIn posts are likely AI-generated was produced by a company that sells AI detection, scored by its own detector on 5,000 search-discovered posts rather than on a sample of the LinkedIn feed.

Zoom out to where all of this is landing, and the environment matters more than the mechanics.

Trust is not disappearing so much as relocating. It is moving away from institutions and towards people the reader can identify and place. That is the ground your disclosure decision lands on, and it explains why a label is the wrong instrument. A label is an institutional gesture, the compliance department's idea of honesty. What is actually gaining trust is the opposite of that: a specific person, named and locatable, saying something you can check against how they behave next year.

The strongest argument against everything I have said

The best objection to everything above is not commercial and it is not legal. It is moral, and it runs like this: if a practice only works while it stays quiet, the practice is telling you something. Disclosure being unpopular is not an argument against disclosure, because stopping at a red light is unpopular too. And the profession is not unanimous here either, since a substantial minority of those 291 practitioners rejected undisclosed ghost blogging outright, which means treating the consensus as settled is its own small dishonesty.

I take that seriously, and here is where I think it actually lands. There is a real difference between concealment and non-announcement, and most of professional life runs on the second one. A keynote does not carry a footnote naming which paragraphs the speechwriter drafted. An annual report does not itemise the finance team. What makes that acceptable in the ghostwriting case is precisely the condition those practitioners named: the substance originates with the person whose name is on it, and that person approves it. The condition is doing all the work. The silence is doing none of it.

Apply the identical condition to AI and the ethics resolve without a badge anywhere. If the argument, the position, the specifics and the judgement come from the named human, and that human can defend every part of it in a room full of sceptics, then the tool that shaped the sentences is a production detail. If those things did not come from the named human, no disclosure rescues the piece, because the problem was never that AI helped. The problem is that nobody wrote it. One line stays non-negotiable in either case: if you are asked directly, you answer honestly. A practice that requires you to lie is a practice you should stop, whatever the research says about the price of telling the truth.

What to do instead of a disclosure badge

A disclosure badge is the worst available option, because it pays the full cost of the penalty and buys almost nothing back. It tells a reader the one thing the research says they punish, and it still fails to tell them the thing they actually want to know, which is whether a competent human examined the substance. The Commission's test is a better instruction than any label anyone has designed: deliberate examination of the substance by a person with relevant knowledge. Build the process so that sentence is true, then build it so you could prove it.

In practice that comes down to five things, and none of them are about the tool.

1. The position comes from the named person first, on record, before a word is drafted. A recorded conversation, a voice note, a marked-up outline, anything with a timestamp. If you cannot point to where the opinion originated, you do not have a post, you have text.
2. Keep the raw input. The question 'can you defend this' needs a physical answer sitting somewhere, and the day somebody asks it is a bad day to start looking.
3. No model supplies a fact, a figure, or a claim about your own work. Every number in a post traces to a named source you have personally read. This one rule would have killed most of the zombie statistics currently circulating in personal branding.
4. Publish the things only you could have written. First-hand specifics, the case you actually ran, the decision you got wrong two years ago and what it cost to unwind. A model can produce a plausible opinion. It cannot produce your Tuesday.
5. If somebody asks whether AI was involved, say what is true, specifically. Being caught inside a denial is a different order of damage from having admitted a tool.

The reason this is worth the effort is that it is exactly what decision-makers say they want, and the numbers on that are not vague.

Look at what those three findings have in common. Overlooked risks, challenged assumptions, and an argument somebody will carry into an RFP on your behalf when you are not in the room. Every one of them requires a position that could turn out to be wrong, held by a person who will still be around when it gets tested. That is not a formatting problem and it is not something a generative tool can supply, because the tool has never sat in the meeting where the risk became obvious.

The position you occupy in a buyer's head is also settled long before you know that buyer exists.

That is the real economics of this argument. Your published thinking does the ranking work while you are asleep, and the shortlist it produces converts most of the time. Which is why I keep returning to something I told Forbes in January 2026: manufactured personas work like credit, where you take the attention now and pay for it later in lost trust, while authentic branding is cash, 'slower to build, but it never defaults.'

So here is what I would actually do this week, and it has nothing to do with labels. Open your last five posts. In each one, find the single sentence that could only have come from you: the number you know because you lived through it, the situation you actually sat in, the thing you argued for in 2023 and quietly abandoned in 2024. If a post does not contain that sentence, no disclosure would have saved it and no amount of hiding the tool would have made it yours. Fix that first. The disclosure argument is genuinely interesting, and for almost everybody reading this it is the second problem, not the first.

Questions I get asked about this

Does disclosing that you used AI actually reduce trust?

In the published experiments, yes. A within-subjects study with 433 participants in the Journal of Science Communication (2026) found AI disclosure reduced the perceived credibility of correct information while increasing the perceived credibility of misinformation. Toff and Simon (Sage, 2025) found AI-labelled news was trusted less even when the articles were not judged any less accurate or unfair. All of this evidence comes from news, science communication and advertising, not from B2B content.

Is human ghostwriting considered ethical by practitioners?

A survey of 291 PR professionals by the Institute for Public Relations found a general industry consensus in favour of undisclosed organisational ghost blogging, provided the content originates from the executive and the executive approves it. A substantial minority of respondents disagreed, so the consensus is real but not unanimous.

Does the EU AI Act require you to label AI-assisted LinkedIn posts from August 2026?

Not on the European Commission's own guidance. The Article 50(4) text-labelling duty covers text published to inform the public on matters of public interest, such as politics and democratic processes, public administration and services, the administration of justice and law enforcement. Text that has undergone human review or editorial control does not need to be labelled. Article 50 becomes applicable on 2 August 2026, with a grace period to 2 December 2026 for the marking obligation on systems already on the market.

Did LinkedIn announce a reach penalty for AI-generated content?

No. On 30 July 2026 LinkedIn shipped a 'Seems like AI slop' report button on posts and comments, discontinued its AI 'enhance your post' feature in favour of a proofreader that does not alter the writer's voice, said it will privately flag content that reads as inauthentic in users' own dashboards, and expanded profile verification. It has published no ranking weights and no measured reach impact by content type. The widely quoted 47% reach reduction has no LinkedIn source behind it.

Can AI detectors reliably tell whether a post was written by a machine?

Not reliably, and not evenly. Leading detectors incorrectly flag human-written text as AI between 12% and 26% of the time, and a widely cited evaluation found seven detectors showed 61.3% false-positive rates on TOEFL essays by non-native English speakers versus near-zero on native English writing.

Sources

1. Understanding Reader Perception Shifts upon Disclosure of AI Authorship (arXiv:2510.24011), Nakano, Takezawa, Matulic, Yang and Yatani (arXiv), 23 January 2026.
2. Visible sources and invisible risks: exploring the impact of AI disclosure on perceived credibility of AI-generated content (JCOM_2501_2026_A09), Journal of Science Communication, 2026.
3. Or They Could Just Not Use It?: The Dilemma of AI Disclosure for Audience Trust in News, Toff, B. and Simon, F.M. (Sage), 2025.
4. Disclaimer! This Content Is AI-Generated: How AI-Disclosures Influence Trust in Advertisements and Organizations, Journal of Interactive Advertising, 2025.
5. The Transparency Dilemma: An Experiment on How AI Disclosures Affect Credibility Perceptions and Engagement Across Topics, AAAI/AIES, 2025.
6. Is Ghost Blogging Like Speechwriting? A Survey of Practitioners About the Ethics of Ghost Blogging, Institute for Public Relations, undated.
7. Transparency obligations under Article 50 of the AI Act (FAQ), European Commission, 2026.
8. LinkedIn adds a button to report AI-generated slop, TechCrunch, 30 July 2026.
9. LinkedIn adds a seems like AI slop button after blocking billions of automated comment attempts, Fortune, 31 July 2026.
10. LinkedIn AI Content Study: 81% of Long-Form Posts Are Likely AI, Originality.ai, July 2026.
11. Evaluating the accuracy and reliability of AI content detectors in academic contexts, International Journal for Educational Integrity (Springer), 2026.
12. AI Detection and assessment: an update for 2025, Jisc National Centre for AI, 24 June 2025.
13. 2026 Edelman Trust Barometer: Trust is In Peril As Society Slides from Grievance into Insularity, Edelman, 2026.
14. The Timeline for Influencing B2B Buyers Is Shrinking: Insights From 6sense's 2025 Buyer Experience Report, 6sense, 12 November 2025.
15. 2025 B2B Thought Leadership Impact Report: Invisible Influence, Edelman and LinkedIn, 8 July 2025.
16. Why Leaders Can't Fake Authenticity In The Age Of Social Media, Forbes (William Arruda, Senior Contributor), 13 January 2026.

Every figure above is linked to its original source. Where I am predicting rather than reporting, I say so in the panels marked My take.