The disclosure question — telling audiences AI was used · AI for Marketing & Content

A Kathmandu fintech launches a Dashain campaign with five smiling customer photos and a line of testimonial under each. The photos are AI-generated. The testimonials are paraphrased from real survey responses, but the faces are not the faces of the people who said the words. A Twitter user notices an extra finger on one of the hands. By Sunday evening, the campaign is not about the product. It is about whether the brand thinks its customers are stupid. The marketing manager spends the next two weeks defending a decision that, had it been disclosed with a single line — “illustrative customer image, generated with AI” — would not have been a story at all. This section is about that line, and the broader question it sits inside: when does AI use need to be told to the audience, and when does it not?

Where the norms and the law are heading

Nepal has not formalised AI disclosure rules. That is not a reason to behave as though they will never arrive — it is a reason to set a policy now, before the bad campaign forces you to set one in public.

The EU AI Act already requires disclosure of synthetic media in several contexts: deepfakes, AI-generated content in news and political communication, and chatbot interactions that a user might mistake for human. Several US states have passed laws on AI-generated political ads, on AI use in employment decisions, and — in California’s case — on broader transparency requirements for generative systems. India is publicly debating its own framework; the IT Rules updates being discussed in 2026 lean toward labelling of synthetic content, and the larger Indian platforms are already enforcing internal versions of it.

Nepal sells into these markets. A garment exporter shipping to the EU, an IT services firm with EU and US clients, a remittance product used by the diaspora — all of them touch jurisdictions where AI disclosure is becoming a baseline. Even when Nepali law is silent, the platforms (Meta, Google, TikTok, LinkedIn) are not — each has rolled out, or is rolling out, AI-content labelling requirements for advertisers.

The position that ages well

Here is the rule that holds up. Disclose AI use when it materially shaped customer-facing output — particularly anything an audience would assume is real.

That means: an AI-generated image of a person being presented as a customer, an employee, or a community member. A voice clip in a radio or social-video ad where the voice is synthetic but could be mistaken for a real spokesperson. A “user testimonial” where the words are AI-paraphrased beyond the original speaker’s intent, or where the speaker is composite. An AI-generated video of a place, an event, or a product demonstration that an audience would reasonably take as documentary footage. In all of these, a one-line disclosure protects the brand and the audience at the same time.

It also means: an AI-drafted long article that the brand is publishing under a named author. The byline says a human wrote it; if the model wrote 70% of the words, the byline is a soft lie. A footer line — “drafted with AI assistance, reviewed and edited by [name]” — keeps the relationship honest.

What you do not have to disclose

Equally important is the other side. Over-disclosure — labelling every keystroke of AI involvement — trains your audience to distrust everything and makes the brand sound paranoid. There are several uses of AI where disclosure is not expected, and announcing it would be performative rather than honest.

AI in internal first drafts that a human heavily edited. The model wrote a starting paragraph; the marketer rewrote it; the final version is the marketer’s voice and the marketer’s facts. The AI was a stage in the process, not the creative voice. You would not label the word processor; you do not need to label this either.

AI as a research or brainstorming tool. The model suggested ten campaign angles; you picked one and built it from scratch. The campaign is yours.

AI in templated communications where the model is not the creative voice — a routine support email, a standard product-update notification, a transactional message. These were templated before AI; they are templated now. Nobody mistakes them for hand-written prose.

The principle: disclose when the audience would feel misled to learn AI was involved. Do not disclose when the audience would simply shrug.

Concrete templates the team can use

These are the lines that work. Copy them, adapt them, and put them in the brand style guide so the team does not have to invent disclosure on a deadline.

On an AI-generated image of a person: Illustrative image — generated with AI. Not a photograph of a real customer.

On a hero image or product mock-up that does not depict a person: Image generated with AI assistance. (For a clearly illustrative graphic, this is often optional — but adding it costs nothing.)

On a long-form article or blog post: This article was drafted with AI assistance and reviewed by [editor name]. Put it in the footer, not the headline; the editor’s name is what does the work.

On a newsletter: A single line in the footer — Some sections of this newsletter were drafted with AI assistance. It does not need to be in the subject line. It needs to be findable.

On a video with AI voice-over: A short on-screen line at the start or end — Voice generated with AI. If it is the founder’s actual voice, no disclosure needed; if it is a synthetic clone of the founder, disclosure is mandatory.

On a social caption where AI helped write the copy but the visual is real: Nothing. The caption is part of the standard production process.

Why concealment is the expensive choice

The temptation to hide AI use is real. The team worries the audience will discount the work; the founder worries about competitor mockery; the agency worries the client will ask for a price cut. These are real worries. They are also outweighed, every time, by what happens when concealment surfaces.

And it surfaces. The AI-image detectors are getting better. The audience is getting more practised at spotting tells. The disgruntled ex-employee, the competitor’s intern, the journalist looking for a story — any one of them can break a campaign open. When it breaks, the story is no longer the campaign. The story is the deception. The brand spends the next month defending the practice instead of marketing the product. The trust the campaign was supposed to build is the trust the brand now has to rebuild.

The cynical math is also clear. A disclosed AI image gets you, at worst, a few comments from purists. A concealed AI image, once found, gets you a news cycle. There is no version of the trade where concealment is the cheaper option.

Check your understanding

Quick check

—

A Nepali fintech runs a Dashain campaign featuring AI-generated images presented as 'happy customers,' with no disclosure. A user spots an extra finger on one of the hands and the post spreads. What is the most likely brand-safety failure mode here?

The campaign cost more than expected because of AI tool subscriptions.
The story shifts from the product to the deception, and the brand spends weeks defending the practice instead of marketing.
The campaign performs poorly on engagement metrics from the start.
The AI image was technically illegal under current Nepali law.

Quick check

—

True or false — because Nepali audiences are less attuned to AI-content debates than EU or US audiences, brands operating mainly in Nepal can safely skip disclosure of AI use.

True — disclosure is a Western concern and does not apply to Nepal-only brands.
False — Nepali audiences do notice, the platforms enforce labelling regardless of market, and the brands selling into export markets touch jurisdictions where disclosure is becoming mandatory.
True — only brands selling internationally need to disclose.
False — Nepali law already requires full AI disclosure for all marketing content.

What comes next

Disclosure is one half of brand safety in the AI era — the half that protects you from being caught hiding something. The next section is about the other half: the quieter, easier-to-miss ways AI use damages a brand even when nobody is looking for deception. Fabricated customers, fabricated claims, cultural defaults that look generic-South-Asian, AI confidently citing things that aren’t true. The section walks through the silent risks and the single habit that defends against most of them.