Before sharing an AI-generated visual in a business presentation or report, state what it represents and check the claim it appears to support. “AI-generated” describes a production method. It does not tell the reader whether they are seeing a proposed product, a fictional training scene, a reconstruction or evidence of something that happened.
The practical rule we propose is to keep illustrations and evidence visibly distinct through the entire handoff. Put the necessary context beside the visual, verify the surrounding headline and inspect the copy that recipients will actually see. A concept image can help people discuss an idea without being allowed to stand in for proof that the idea has been built or tested.
This guide concerns communication practice, not a legal compliance determination. Its examples and review method are editorial proposals. No images, audience studies, content-credential tools or assistant products were independently tested for this article.
Start with the claim a reader could take away
An attractive image acquires meaning from its setting. A fictional control room beneath “Our operations today” suggests something different from the same image beneath “A possible future workspace.” A generated interface beside a performance claim can appear to document a working system even when it was created only to explore a design.
Before asking an assistant to generate or place an image, write down the job it must do. Is it setting a mood, explaining an imagined scenario, proposing a layout or documenting an observation? Then identify the strongest plausible misunderstanding.
For a product concept, that misunderstanding might be that the pictured feature is available. For a training scene, it might be that the people shown are actual employees or customers. For a fictional dashboard, it might be that the displayed figures came from operating data.
Our recommendation is to resolve that ambiguity in the artifact, not leave it to the presenter to remember. If an observation needs evidence, provide the real source and explain what it establishes. A synthetic illustration can accompany that source, but it cannot become a substitute for it.
Labeling research does not make disclosure a complete solution
In Labeling AI-generated media online, published in PNAS Nexus in May 2025, Chloe Wittenberg and colleagues studied process-based and harm-based labels in two preregistered survey experiments. Participants viewed misleading, AI-generated images in social-media posts. The experiments were conducted in 2023 with American respondents.
Labels reduced belief in the presented claims, but labels simply identifying AI generation had little effect on respondents’ stated engagement intentions. Belief and willingness to share were separate outcomes. The study measured reported intentions, not observed redistribution inside a company.
This is evidence about particular labels and misleading social-media stimuli. It does not validate the captions proposed here, measure current business audiences or establish that a label prevents every misunderstanding. Some authors disclose previous platform research funding or subsequent platform employment; those disclosures are available with the paper.
The implication we draw is to define what the disclosure must explain. If the concern is that readers will mistake a proposed installation for a completed one, naming AI alone leaves the status question unanswered. Say that it is a concept and identify the observation it does not document.
Separate origin information from factual support
The C2PA version 2.4 explainer describes Content Credentials as a way to bind provenance information to an asset and make that information verifiable. Provenance concerns origin and history. The explainer explicitly distinguishes this from judging whether digital content is true, accurate or factual.
For a business reviewer, these are different checks. Provenance may help establish that an image was generated or edited using a particular workflow. It does not establish that a photographed-looking facility exists, that the pictured product meets its specification or that a dashboard value is an observed result.
Preserve available provenance information when your tools support it, while retaining the source and claim review. Do not describe a generated illustration as “verified evidence” merely because a credential validates. Equally, do not infer that an image is genuine because it has no visible AI label.
Partnership on AI’s March 2025 practitioner account discusses disclosure context beyond a simple AI-or-not signal. It draws lessons from cases contributed by Google and civil-society organizations. That is implementation guidance and case reporting, not a controlled test of this business-review method.
A fictional facilities proposal
Imagine a team presenting a proposed reception area. It uses a generated image to discuss seating, signage and the visitor journey. No reception area has been installed, and the image is not a photograph of the organization’s premises.
A caption such as “AI-generated concept illustration of a proposed reception area; not a photograph of an installed site” explains both origin and status. The slide headline should also describe a proposal. Calling it “The visitor experience we delivered” would contradict the caption and imply a result the team has not supplied.
If the team wants to claim that the design reduces waiting time, it needs separate evidence. The image cannot measure that outcome. A nearby source could support a finding from an actual evaluation, or the team could mark the waiting-time effect as a hypothesis for a future pilot.
This example is invented. No premises were assessed and no visual was generated for it. Its purpose is to make the review decision concrete: does the artifact invite a reader to treat an imagined scene as a completed business achievement?
The same distinction can guide other uses:
| Intended visual | Context to make visible | Claim requiring separate support |
|---|---|---|
| Product concept | Proposed appearance or feature; availability not established by the image | The product is shipping or has passed a test |
| Fictional training scene | Invented scenario and people | A real employee or customer experienced the event |
| Interface mockup | Design proposal with illustrative content | The system is implemented or the displayed metrics were observed |
| Explanatory illustration beside a real finding | Illustration distinguished from the finding’s source | The picture itself documents the research or operating result |
These are proposed review categories, not an exhaustive disclosure standard. Choose wording that addresses the actual misunderstanding in the communication.
Carry context into the exported copy
Our proposed handoff check follows the visual through the formats the team plans to distribute. Inspect the slide shown during a meeting, the PDF sent afterward and any standalone image supplied for reuse. Each can carry a different amount of surrounding context.
Do not put the only explanation in speaker notes if recipients will receive a document without those notes. If the image will be reused alone, provide an appropriate labeled version or an explicit caption with the handoff. Check the label at the size and resolution at which people will read it.
Inspect cropping and layout as well. A disclosure placed near the edge can disappear when someone extracts part of a slide. Decide which exports are suitable for reuse and keep their purpose clear. Do not call a caption persistent unless you have checked the actual output path.
Alternative text and narration should preserve the same distinction. Describe a fictional concept as a fictional concept; do not let an accessibility description introduce an unsupported claim about real people or results.
These checks are operational recommendations. We have not tested a particular presentation tool’s metadata preservation, image extraction or credential behavior. Verify those features in the tools and formats your team uses.
Review the complete communication before release
Keep a small record of the visual’s purpose, origin, permitted use, proposed caption and supporting evidence for any surrounding factual claim. Identify who owns revisions when the concept changes or when a real implementation becomes available. A concept label should not quietly be removed just because the image has been used repeatedly.
Review the image, headline and supporting text together. Ask what a reasonable recipient could believe happened, exists or was measured. If that answer exceeds the evidence, change the framing or replace the visual. A disclosure is not a repair for an incompatible headline.
For numerical graphics, our guide to AI-generated charts covers the separate task of checking the underlying table, comparison and measure. Here the additional boundary is whether a visual is illustrative at all, rather than a representation of observed data.
Start with one recurring deck or report. Identify its generated visuals, give each a clear representational role and inspect one real distribution copy. The useful result is a communication whose readers can distinguish an idea being explored from evidence supporting a decision.