Brand safety — the silent risks · AI for Marketing & Content

The dramatic brand-safety failures get the headlines: the deepfake, the fake testimonial, the AI image that fools nobody and embarrasses everybody. They are real, and the previous section was about not getting caught hiding the work that produces them. But the quieter failures — the ones that do not make the news cycle, the ones that the marketing team itself often does not notice — are where most of the actual damage happens. A fabricated statistic that nobody fact-checks. An AI-generated face that looks just slightly generic-South-Asian. A festival greeting that picked the wrong honorific. Each one is small. Each one tells your audience the brand does not quite know them. Stacked over a year, they cost more trust than a single big incident does.

The fabricated-customer risk

The most common silent risk in 2026 Nepali marketing is the AI-generated person presented, with no caveat, as a real human connected to the brand. A “happy customer.” A “team member.” A “community elder” in an INGO appeal. The marketer who used the image often did so under deadline pressure; finding a real customer to photograph would have taken three days the campaign did not have. The model produced a smiling face in thirty seconds.

The problem is that audiences are getting fast at the tells. The hands — six fingers, missing thumbs, fingers that bend at the wrong knuckle. The eyes — slightly mismatched pupils, irises that do not quite catch light the same way. The jewellery — earrings that do not match, a necklace that ends mid-collar, a tika that floats above the skin rather than sitting on it. The faces — uncannily symmetric, skin that is too smooth, no asymmetric small details that a real face has. The background — text on a sign behind the person that says nothing in any language, a doorway that does not quite line up.

The Nepali audience is getting practised at this in 2026 the way every audience is. When they spot it, the post spreads — not because they are angry, but because it is satisfying to spot. “Hera, yo Khalti ko ‘customer’ AI ho — auli herchu.” By the time the campaign manager sees the screenshot in the group chat, the post has been shared a thousand times with the tell circled.

The fabricated-claim risk

AI is a fluent liar. It will, with no flicker of hesitation, invent a market share, a customer count, a “study shows” reference, a price comparison, a feature your product does not have. The prose around the claim will be confident. The grammar will be clean. The marketer who pastes it without checking will publish a brand statement that is — depending on how careful you want to be with the word — false.

The pattern is familiar from the long-form-content section, but the brand-safety stakes are higher here. A blog post with a fabricated statistic is bad. A LinkedIn post from the founder’s account with a fabricated statistic is worse, because the founder’s name is on it. An RFP response with a fabricated capability is worst of all, because somebody is going to write that capability into the contract and your team is going to have to deliver something the model invented.

The riskiest categories of AI fabrication for a brand are: specific numbers (market shares, customer counts, performance metrics), named comparisons (claims that your product is faster, cheaper, or better than a named competitor), regulatory or compliance claims (statements about what laws require, what licences you hold, what certifications you have), customer quotes (paraphrased testimonials that the customer did not actually say), and product capabilities (features the model thinks would be reasonable but that your product does not actually have).

Any of these in a customer-facing piece, unchecked, is a brand-safety failure waiting for the first customer or regulator who reads carefully.

The defamation and IP risk

AI is not just liable to make things up about your own brand. It will, just as confidently, make things up about competitors. A draft for a comparison post that says “unlike [competitor], we comply with NRB regulations” — when the competitor in fact does — is defamation in any jurisdiction worth worrying about. A draft that says “[competitor] charges 3% in hidden fees” when they do not is the kind of claim that, in 2026, gets a letter from a lawyer faster than it gets engagement.

The IP risk is the mirror image. AI image models have absorbed the work of named artists and named studios. Asking the model to generate something “in the style of [famous illustrator]” produces output that, in several jurisdictions, has already triggered legal complaints. Asking for “in the style of” a named Indian or Nepali artist — and several models will comply — produces work that the artist may, reasonably, want to do something about. The same applies to mascots, logos, and characters from existing brands; the model will cheerfully produce something that looks 80% like a known property, which is exactly the percentage that gets a cease-and-desist.

The defence is mechanical and boring: do not name competitors in AI prompts, do not ask for “in the style of [named artist]”, do not let the model generate logos or characters that resemble existing IP. These three rules, in the brand AI policy, prevent most of the legal trouble.

The cultural risk — the one only Nepal sees

This is the failure mode that is invisible to the global tooling and almost unique to brands serving Nepali audiences. The model does not know you. Asked to generate “a Nepali family for a Dashain ad,” it will produce a generic South Asian family — Indian-looking, dressed wrong, in a room that could be anywhere in the subcontinent. Asked to write a Tihar greeting, it will produce something that reads like a translated Diwali message. Asked to address a Nepali audience formally, it will pick the wrong honorific register — too casual for a banking ad, too formal for a youth-focused fintech, or simply wrong for the relationship the brand has.

These failures do not announce themselves. The audience reads the post, feels a small wrongness, and scrolls. They do not write a complaint. They simply update their unconscious sense of the brand: this company does not actually know us. Stacked over a year of campaigns, that signal is more damaging than one fabricated image, because it is constant and ambient.

The defences are specific. Brand voice docs in Nepali, not English translated. Cultural context briefs for each major festival — what the brand says at Dashain, what it says at Tihar, what it says at Maghe Sankranti, what it says nothing about because it is not the brand’s place. Real photography for hero shots in cultural campaigns, not generated faces. A Nepali editor’s veto on any customer-facing copy that touched a model — somebody whose judgement about register and tone the brand trusts more than the model’s.

The single highest-leverage defence

Most of the failures above — fabricated customers, fabricated claims, defamation drafts, cultural misses — fail the same way. A marketer publishes AI output that nobody read carefully first.

So the highest-leverage habit, the one that, if you do nothing else, defends against most of this, is read every AI output carefully before it goes live, and edit it. Not skim it for typos. Read it the way you would read copy from a junior whose work you have not yet learned to trust. Check the numbers. Check the claims about your product. Check the claims about competitors. Check that the image actually depicts what you wanted it to depict. Check the cultural register. Edit until the piece is something the brand could defend in any room.

Around that habit, the brand should have a written AI policy that lists, by content category, what the team is allowed to do unsupervised, what requires review, and what is forbidden. The policy is dull to write and decisive to have when a junior asks at 11pm whether they can generate a customer photo for tomorrow’s post. “No — see policy section 3.2: customer images must be real photography. Use a stock photo with disclosure or wait until tomorrow.” The policy makes the right answer the easy answer.

The third leg of the defence is an editor’s veto — one named person on the team whose call on “is this safe to publish” is final. Not a committee. One person. They reject anything that looks even slightly off, and the team accepts the rejection. That role is the one that catches the fabricated finger before it ships.

Check your understanding

Quick check

—

A Nepali brand is starting to use AI for content. They want one habit that, by itself, defends against the largest share of brand-safety failures — fabricated customers, fabricated claims, defamation drafts, cultural misses. What is the single highest-leverage habit?

Use the most expensive AI model available — premium models hallucinate less.
Read every AI output carefully and edit it before it goes live, the way you would read copy from a junior whose work you have not yet learned to trust.
Add a generic AI disclosure to every piece of content, regardless of how AI was used.
Never use AI for any customer-facing content — restrict it entirely to internal use.

What comes next

This is the second-to-last section of the course. The next — and final — section is a short essay on the shape of marketing work as AI scales: what changes, what stays and grows in value, what the new marketing craft looks like, and why the marketer who reads the room is needed more, not less.