Why Saying “AI-Generated” Is Not Enough

Credibility, risk and AI governance in business

by Andrea Viliotti

In business, transparency about the use of artificial intelligence is becoming an organisational requirement rather than a communication accessory.

But a simple label — “AI-generated content” — is not enough to create trust.

A study published in the Journal of Science Communication by Teng Lin and Yiqing Zhang shows, within a specific experiment on science communication texts in a Weibo-like setting, a more ambiguous effect than many managers would expect.

When an AI disclosure label was present, the perceived credibility of correct information decreased, while the perceived credibility of misinformation increased.

The result should not be generalised beyond its methodological boundary. But it does force a practical management question.

What does a reader actually see when they encounter an AI label: a source, a channel, a verification signal or a statement of accountability?

The operational answer is that these four things do not coincide.

AI governance in business therefore cannot stop at declaring the origin of a text.

It must make visible what was generated, what was checked, by whom, against which sources, with which accountability and within which decision limits.

This is the rationale behind the GDE Disclosure Stack: not a single label, but a readable layer of generation, verification, source, evidence and accountability.

Answer in brief

In business, an “AI-generated” label signals technology involvement, but it does not prove the source, verification status, accountability or quality of the claim. To build trust, organisations must make visible what was generated, what was checked, by whom, against which sources and within which decision limits.

AI Disclosure in Business

Opening: the paradox of AI transparency

The scene is now familiar. A company publishes a product page, a customer reply, an internal policy, a technical note or a sales summary and adds a reassuring sentence: “this content was generated with AI” or “AI-assisted content.”

At first glance, the move seems correct. It signals openness. It avoids the impression that the organisation is hiding the use of generative AI. It also responds to an emerging expectation: readers increasingly want to know whether a text was written, summarised, translated or refined by a machine.

Yet that sentence can create a misunderstanding. Some readers may see it as a reason to trust the content less. Others may see it as a sign that the content is more neutral, more technical or somehow more objective. In both cases, the label risks becoming a shortcut rather than a tool for judgement.

An AI label says something about origin or production process. It does not say whether a figure has been checked, whether a commercial promise is grounded, whether a compliance statement is up to date, whether a technical claim has been reviewed, or whether a human owner is accountable for the final output.

The thesis of this article is straightforward: genuine AI governance in business is not the generic declaration that AI has been used. It is the ability to make visible what has been generated, what has been verified, by whom, with which sources, under which responsibility and within which decision limits.

The paper: when a label does not produce the expected effect

The starting point is the study by Teng Lin and Yiqing Zhang, “Visible sources and invisible risks: exploring the impact of AI disclosure on perceived credibility of AI-generated content” published in the Journal of Science Communication. The paper examines whether AI disclosure changes the perceived credibility of science communication texts.

The experimental design is within-subjects: participants evaluate content under different conditions. The valid sample consists of 433 responses. The authors manipulate two elements: the presence of a label indicating AI generation, and the veracity of the information, distinguishing correct information from misinformation.

The most important result is what the authors call a truth-falsity crossover effect. Within the study’s boundary, AI disclosure reduces the perceived credibility of correct information and, unexpectedly, increases the perceived credibility of misinformation. The disclosure signal therefore does not simply make people more careful; it can redistribute credibility in a counterintuitive way.

The paper adds two useful elements. The first concerns negative attitudes towards AI: pre-existing distrust of AI can intensify the credibility penalty for correct content. The second concerns audience involvement: involvement plays a more limited and topic-dependent role than one might expect.

The limitations are as important as the findings. The study concerns text, not video, images or complex conversational interfaces. It focuses on two areas of science communication. It removes social cues such as avatars, usernames, reposts, likes and comments. It is situated in a Chinese social media environment, using Weibo-like posts. It does not authorise automatic generalisation to all countries, industries, media formats or corporate contexts.

Box — The paper’s paradox in one sentence

Within its experimental boundary, a simple AI label did not automatically help readers separate correct information from misinformation; it redistributed perceived credibility in a counterintuitive way.

The GDE reading: the issue is not AI, but the source-channel-verification-observer system

The GDE reading of the paper starts from a distinction that companies often neglect. A text never reaches the reader as pure content. It arrives inside a system: a source that authorises it, a channel that distributes it, a process that produces it, an observer who interprets it, and a context that defines the consequences of believing it.

When a company writes “AI-generated,” it makes only one part of that system visible. It says something about the production process, while leaving the rest in the dark. The reader may then make a category error: they may treat the channel as the source, the production process as verification, or the label as an implicit guarantee of quality.

This is the source-channel collapse. It happens when the fact that a piece of content has passed through AI is confused with the fact that it is grounded, checked, updated or accountable.

The operational distinction is clear. An AI label is not a source: it does not identify who stands behind the content. An AI label is not verification: it does not say whether the claim has been checked. An AI label is not accountability: it does not assign responsibility. An AI label is not claim quality: it does not prove that the statement is correct, proportionate or usable for decisions.

This creates the Disclosure-Verification Gap: the distance between what a label makes the reader perceive and what the content has actually demonstrated. If the gap is wide, the reader receives a signal but not the tools needed to interpret it.

Table 1 — AI label, verification, source and accountability: four different things

Object	What it signals	What it does not guarantee	Managerial question
AI label	AI involvement in producing or refining the text.	It does not identify the authoritative source behind the content.	Which part of the content is AI-assisted and which is not?
Verification	Checking the claim against sources, data, versions or policies.	It does not arise automatically from text generation.	Who checked the claim and with what evidence?
Source	Responsible origin of the knowledge used: document, expert, owner, database.	It does not coincide with the channel that wrote or distributed the text.	From which primary source does the statement derive?
Accountability	Editorial, technical or organisational owner who answers for the content.	It is not assigned by process transparency alone.	Who owns the decision and who updates the content?

Implications for business

For executives and business owners, the issue is not whether to use generative AI. The issue is how to insert AI into processes where trust, responsibility and decision-making remain legible. A company that uses AI to accelerate text production must also accelerate its ability to qualify, verify and own what the text says.

In marketing, the risk is overstating a promise with elegant but unchecked copy. In knowledge management, the risk is that an AI summary is treated as the source rather than as a derivative. In HR, the risk is that a job description, internal policy or performance note becomes detached from review, fairness and accountability. In customer care, the risk is that a fast answer becomes a wrong promise.

In technical documentation, the stakes are even more concrete. A manual, procedure or release note generated with AI does not become reliable because it is well written. It must remain connected to product version, safety constraints, engineering review and quality responsibility.

The organisational point is this: AI increases the speed of text faster than many companies increase their capacity to verify it. When production grows and verification does not, the transparency label becomes a thin cover over a deeper governance deficit.

The GDE Disclosure Stack for business

The answer is not to remove disclosure. It is to make disclosure more informative. The operational proposal is a GDE Disclosure Stack: a set of clear, readable and risk-proportionate signals. The purpose is not to burden every text with bureaucracy; it is to give the reader enough context to interpret the content correctly.

The first layer is AI involvement: which part of the content was generated, summarised, translated, rewritten or merely assisted. The second is verification status: unverified draft, internally checked, expert-validated, source-linked or decision-ready. The third is the responsible source: document, database, expert, function or owner. The fourth is evidence at claim level: what supports the important statements. The fifth is decision risk: informational, operational, sensitive or not suitable for decision-making without further validation. The sixth is accountability: who approved the content and who updates it.

In practice, a company does not need to write a treatise on AI in every document. It must decide which minimum signals should be visible for each risk level. A social post may require a light disclosure. A technical note requires source and version. A compliance memo requires jurisdiction, date and accountable review. A customer-facing promise requires owner, verification and escalation path.

This is the difference between transparency and verifiable trust. The first says: “we used AI.” The second says: “this output was produced in this way, checked in this way, supported by these sources, owned by this person or function, and usable within these limits.”

Table 2 — Simple disclosure vs GDE Disclosure Stack

Dimension	Simple disclosure	GDE Disclosure Stack
Origin	States that AI was used.	Distinguishes generation, rewriting, summarisation, translation and assistance.
Verification	Often unspecified.	Indicates whether the claim was checked and by whom.
Source	Implicit or confused with the channel.	Connects the content to a source, document, owner or expert.
Risk	Same label for all content.	Varies by marketing, HR, technical, compliance, customer care and sales use.
Accountability	May remain anonymous.	Assigns editorial, technical or organisational accountability.
Decision use	Leaves interpretation to the reader.	States the use limit: informational, operational, sensitive or to be validated.

How disclosure should change by business function

Minimum governance cannot be identical in every department. The point is not to create silos, but to recognise that each function has a different relationship between speed, risk and responsibility.

In marketing, disclosure should clarify whether AI was used for drafting, rewriting, translation or creative variation, while verification should focus on product claims, prices, availability and regulated statements. In knowledge management, the central question is not whether a summary reads well, but whether it points back to the primary document, version and subject-matter owner.

In HR, generative AI can help with job descriptions, onboarding content and internal communications, but human review must remain visible when fairness, evaluation or sensitive employee information is involved. In customer care, the critical point is escalation: the reader must know when an AI-assisted answer has been verified and when it must be treated as preliminary.

In technical documentation and scientific communication, the label is weakest when the claim is strongest. The more a text can influence a technical, legal, medical, safety or strategic decision, the more disclosure must be linked to source, verification and accountability rather than to the mere use of AI.

Operational map — Where disclosure must change form

Area	Typical AI use	What must be verified	Candidate owner
Marketing	Copy, campaigns, product pages, newsletters.	Product claims, prices, availability, promises, regulated statements.	CMO / communications / legal for sensitive claims.
Knowledge management	Summaries, knowledge bases, internal memos, wikis.	Primary source, version, date, knowledge owner.	Knowledge manager / subject-matter expert.
HR	Job descriptions, policies, performance summaries, internal messages.	Fairness, policy alignment, human review.	HR director / people operations.
Customer care	Chatbots, FAQ, ticket triage, assisted emails.	Updated answer base, guarantees, complaints, escalation rules.	Customer operations / legal for contractual language.
Technical documentation	Manuals, SOPs, datasheets, release notes.	Product version, safety constraints, compatibility, quality review.	Product owner / engineering / QA.
Compliance	Regulatory summaries, internal policies, training.	Official source, date, jurisdiction, specialist validation.	Compliance officer / legal / DPO where relevant.
Sales	Offers, RFPs, proposals, account planning.	Prices, SLAs, promises, case studies, customer data.	Sales director / sales operations / legal.
Technical-scientific communication	White papers, R&D notes, technical explainers.	Claim-level evidence, methodological limits, expert review.	Technical-scientific owner plus editorial owner.

What to do on Monday morning

The checklist does not require a large governance structure. It requires a simple discipline: do not allow AI to produce more content than the organisation can responsibly govern.

Start by mapping the visible AI outputs: website copy, sales material, customer care replies, internal knowledge bases, HR communication, technical documentation, training content and executive summaries. Then classify the risk: informative, operational, sensitive or decision-impacting.

Assign an owner to each output class. Define the minimum level of verification for each class. Separate texts that may be published after editorial review from texts that require subject-matter validation, legal review or technical approval. Finally, define when the label must say not only “AI-assisted,” but also “verified by,” “based on,” “approved by,” or “not suitable for autonomous decision-making.”

The first improvement is often cultural rather than technical: stop treating the AI label as the end of transparency. Treat it as the beginning of a verification conversation.

Box — Checklist for executives

1. Map visible AI outputs.

2. Classify content by decision risk.

3. Assign accountable owners.

4. Define verification levels.

5. Separate draft, verified, expert-validated and decision-ready content.

6. Link critical claims to sources.

7. Make escalation paths visible.

8. Review labels when products, policies or contexts change.

Risks to avoid

The first risk is cosmetic transparency: adding a label and considering the problem solved. This is the simplest form of apparent governance, but also the weakest. A poor label can make the organisation look transparent while leaving the reader unable to assess the claim.

The second risk is generalised distrust. If every AI-generated text is framed as suspect, or if every AI-generated text is framed as automatically neutral, the company teaches its audience to read in binary terms. But credibility is not binary. It depends on the claim, the source, the verification process and the decision context.

The third risk is internal overconfidence. Management can trust an output too much because it is coherent, fast, fluent and apparently well structured. It can also trust an output too little because the AI label becomes a stigma. Both reactions are governance failures.

The fourth risk is evaporated accountability. If the content is “from AI,” who updates it? Who answers the customer? Who corrects a technical sheet? Who decides that an HR summary should not be used for evaluation? Without an owner, the label becomes a way to distribute responsibility until no one truly holds it.

Box — ClaimGuard: what this article is not saying

It does not say that every AI label has negative effects.

It does not recommend hiding AI use.

It does not turn AI into a general misinformation accelerator.

It does not present the paper as a universal law of AI communication.

It does not claim that GDE is empirically proven by the paper.

It does not treat tiered labels as a solution tested by the paper.

From cosmetic transparency to verifiable trust

A company using generative AI must move beyond a false alternative: either hiding the machine or declaring it in such a generic way that nobody is helped. The stronger path is a third one: designing disclosure as part of a trust architecture.

The Lin and Zhang paper is valuable because it challenges a naive assumption: that showing a label is enough to orient the reader’s judgement. In their experiment, the label can generate effects that are not aligned with the intended purpose of transparency.

The managerial task is therefore to design better signals. Not only “this text was generated with AI,” but “this text was generated with AI, checked against these sources, approved by this owner, and usable only within these limits.”

Trust is not obtained by naming the technology. It is obtained by making the path from content to source, verification, responsibility and decision controllable. For a company that wants to use AI seriously, this is where governance begins.

Final box — Ten rules for using AI in business without confusing disclosure and trust

1. Declare the use of AI when the reader has a legitimate interest in knowing it.

2. Do not treat disclosure as verification.

3. Separate AI involvement, source, evidence, owner and decision limit.

4. Classify AI outputs by risk, not by department alone.

5. Assign a human or organisational owner to every decision-relevant output.

6. Use stronger verification for technical, legal, HR, compliance and customer-facing claims.

7. Keep a link between summaries and primary sources.

8. Make escalation paths visible for customer-facing AI content.

9. Review AI-generated content when context, law, product version or policy changes.

10. Build verifiable trust, not cosmetic transparency.

Frequently asked questions for executives and managers

Why is “AI-generated content” not enough in business?

Because the label indicates the production process, but it does not say whether the content has been verified, who is accountable, which sources support it, or which decision use is appropriate.

What is the difference between AI disclosure and verification?

Disclosure informs the reader that AI was involved. Verification concerns the checking of the claim against sources, data, policies, versions and accountable human or organisational review.

What is the GDE Disclosure Stack?

It is an operating grid that separates AI involvement, verification status, responsible source, available evidence, decision risk and editorial or organisational accountability.

When can transparency become ambiguous?

When the reader interprets the AI label as a guarantee of quality or, conversely, as an automatic reason for distrust. Transparency works only when it also clarifies verification and responsibility.

What should a manager do immediately?

Map visible AI outputs, classify risk levels, assign owners, define verification levels and make the limits of use explicit for AI-generated or AI-assisted content.