Which Parts of Your Marketing May an AI Invent, and Which Must Be Real?

On a Thursday at 4:40 pm, a message lands in a dental clinic's Instagram inbox: "Who is the woman smiling after her treatment in your post? I came in with her photo and the dentist said she is not a patient." The image had been generated with the clinic's AI assistant, and the caption read "Real results, real smiles." The clinic had no good answer.

This is a composite of situations I see, not one client. The clinic has 22 staff and is in Beirut, and I have kept the numbers consistent so you can follow the arithmetic.

Why "was AI used?" is the wrong question

Most owners I speak with hold a reasonable-sounding belief. If the image looks good, the quote reads well and nobody is harmed, it is simply efficient marketing, and ethics is something large companies worry about. A common refinement is that the bar is naming the tool: say that AI was involved and the matter is closed.

In my work I find both ideas miss the point. A viewer does not care which software produced a pixel. A viewer cares whether the thing in front of them is something they can rely on. A generated texture behind a headline asks nothing of the viewer. A smiling woman captioned "real results" asks for belief, and a prospective patient may act on it by booking a procedure.

So the question I use is not "was AI used?" It is this: is the viewer using this item to decide whether to trust or buy? If not, it is decoration, and AI may invent it. If yes, it is evidence, and it must be real, or plainly labelled as something else. A label such as "illustration" repairs a caption. It does not repair a claim.

The proof matrix

Take six kinds of item a small business publishes. For each, ask what the viewer does with it. The right-hand column gives the verdict.

Four verdicts: generate freely, generate and label, real only, real only plus consent.

Two things stand out when owners first use the grid. Most of what a business posts lands in the first row and is harmless. And the dangerous cells are few and specific: people who look like customers, quotes that look like reviews, results that look like proof.

What the clinic's thirty days looked like

I asked the clinic to pull its last 30 days of posts and place each one in a row. There were 38 posts in all:

• 21 decor posts: backgrounds, abstract graphics, seasonal artwork. Fine as they were.
• 6 product or facility images: 5 real, 1 generated waiting room that did not match the actual one.
• 7 images of people: 4 generated, 3 real.
• 4 testimonials: 2 real, with consent; 2 written by the assistant "from patient feedback themes."

That is 21 + 6 + 7 + 4 = 38. Seven items sat in the wrong cell: 1 facility image, 4 generated people presented as patients, and 2 invented testimonials. Seven of 38 is 18 percent of the month's posts making an evidence claim the clinic could not back.

The fix was ordinary. Where a real patient had given consent, a real photo replaced the generated one. The two invented testimonials were taken down and the clinic asked two patients for real ones; one agreed within three days. The remaining generated images were recaptioned as illustrations, and the clinic put a one-line correction under the original post. The audit took 25 minutes.

The lesson I draw is that the damage was not the generated pixel. It was the caption that said "real."

Rules the law is already moving toward

I am not a lawyer, and this is not legal advice. But two public texts show the direction, and both matter to a Lebanese firm that sells to customers in the EU or the US, or to the diaspora, which many Instagram- and WhatsApp-led businesses do.

In the EU, Article 50 of the AI Act sets transparency obligations. In my reading, paragraph 4 says that those who deploy an AI system to generate or manipulate realistic image, audio or video content, the kind that resembles real people, places or events and could pass as authentic, must disclose that it is artificially generated or manipulated. There are exceptions. For evidently artistic, creative or satirical work the duty is lighter. The page I link gives the application date as 2 August 2026. Read the text for your own situation.

In the US, the Federal Trade Commission announced a final rule on 14 August 2024 banning fake reviews and testimonials. Its announcement says this includes reviews that are AI-generated or that misrepresent who wrote them or whether the writer had any real experience. It is a US rule and applies to commerce reaching US consumers.

Neither text asks you to stop using AI. Both ask that evidence be what it appears to be. My earlier piece on when an AI decision owes the affected person an explanation covers a different duty; this one concerns what a business publishes.

A 20-minute audit you can run today

1. Collect. Pull the last 30 days of posts, images, captions and short videos into one list.
2. Sort. Put each item in a row of the matrix, and decide whether the viewer treats it as decoration or evidence.
3. Count. Mark every item in a Real-only cell that was generated, or edited beyond lighting and cropping.
4. Record. Keep one line per item: the item, its cell, and who confirmed it is real.
5. Repair. Replace, relabel or remove each flagged item, and add a one-line correction wherever the original claimed to be real.

Repeat it each quarter, or whenever a new tool or a new agency joins your process. The record in step 4 matters more than it looks: when a patient asks, you can answer in a minute.

Sort your last 30 days of posts into the matrix this week and fix anything that lands in Real only.

Composite note: the clinic, the patient's message and every figure above are an illustrative composite built from situations I see in my work. They are not one client and not a measured result.

Frequently asked questions

Do I have to label every AI-generated image?

Not under the matrix. A generated texture, pattern or abstract illustration that nobody relies on for a decision is decoration, and a label adds nothing. Label an image when it is realistic and a viewer could take it as showing real people, a real place or a real event. If you sell into the EU, Article 50 may require disclosure for such content; read the text for your case. This is general guidance, not legal advice.

Can I use AI to polish a real customer's testimonial?

You can correct spelling and trim length if the customer approves the edited wording in writing. You cannot add praise, change what the customer said, or merge two customers into one quote. The safest routine is to send the final text back to the customer and keep the reply with the item.

Is a generated background behind a real product photo acceptable?

Usually yes, provided the product itself is untouched in shape, colour, size and finish, and the background does not make a claim, such as showing a place or setting you do not actually offer. The product is the evidence; the background is decoration.

What if an AI tool made a post I already published that falls in a Real-only cell?

Replace it with a real item if one exists, or relabel it as an illustration and rewrite the caption so it no longer claims to be real. If the original said real, add a one-line correction under it rather than deleting it quietly. Then record what you changed and who confirmed it.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.