Which Parts of a Company's Marketing Should Never Be Generated by AI?
Under Dr. Jonah Tebaa's framework distinguishing decorative marketing from decision-making evidence, a company must never use AI to generate items that customers rely upon as proof. Across six operational rows, customer testimonials, before-and-after results, depictions of actual products that ship, and figures presented as real clients must never be fabricated. In a composite clinic example, an audit flagged seven misplaced evidence items, including artificial patient quotes and an unmatched waiting room, requiring genuine replacements.
A patient writes to a dental clinic in Beirut to ask who the smiling woman in its Instagram post is. She came in carrying the photo, and the dentist told her no such patient exists. The image was generated, and its caption said "Real results, real smiles." Dr. Jonah Tebaa builds his argument on this scene, which he is careful to describe as a composite of situations he sees rather than a single client.
His reading is that the clinic did not get into trouble for using AI. It got into trouble because it asked a viewer to believe something that was not true.
Decoration and evidence
Dr. Tebaa argues that the usual debate, whether a business should say it used AI, points at the wrong target. Viewers do not weigh the tool. They weigh whether an item is something they can lean on. His distinction is between decoration, which a viewer ignores when deciding, and evidence, which a viewer relies on when deciding whether to trust or buy.
A generated background or abstract illustration is decoration, and he sees no problem in inventing it. A smiling patient, a customer quote, a before-and-after result and the owner's own voice are evidence. For these, the item must be real, or be plainly labelled as something else. A label such as "illustration" can correct a caption, he notes, but it cannot rescue a claim that was false.
Six kinds of item, four verdicts
Dr. Tebaa sorts what a small business publishes into six rows and gives each a verdict. Backgrounds and decor may be generated freely. Product images should be photographs of the real item; AI may help with lighting or the backdrop, but a generated product that differs from what ships is never acceptable. People appear in two ways: a generated crowd in a mood image can be used if labelled, while a generated figure presented as a customer must be replaced by a real, consenting one.
Testimonials and results are the strictest rows. A testimonial must be a real customer's words, used with written permission, and an AI-written quote offered as a customer's is out. A before-and-after must be genuine, and AI enhancement of a result is out. The owner's or a staff member's face and voice require that person's own written approval for each use. If a voice is cloned to read an offer, he says it should be labelled every time and should never say anything the owner has not said.
What the clinic found
In the example, the clinic sorted 30 days of posts into the grid. Of 38 posts, 21 were decoration and needed nothing. Six were product or facility images, of which five were real and one was a generated waiting room that did not match the real one. Seven showed people, four of them generated, and four were testimonials, two real and two written by the assistant from themes in patient feedback.
Seven items, or 18 percent of the month, sat in the wrong cell. The repairs were modest: real photos where consent existed, two testimonials requested again from patients, the remaining generated images recaptioned as illustrations, and a one-line correction under the original post. The whole audit took 25 minutes. Dr. Tebaa treats that figure as part of the argument, since the exercise is cheap compared with answering a patient who has noticed.
The legal direction, read cautiously
Dr. Tebaa says plainly that he is not giving legal advice. He points to two public texts as signs of direction. Article 50 of the EU AI Act, in his reading, requires deployers of realistic generated or manipulated image, audio or video content to disclose it as such, with lighter duties for evidently artistic or creative work. The page he cites gives 2 August 2026 as the date these provisions apply. In the United States, the Federal Trade Commission announced a final rule on 14 August 2024 against fake reviews and testimonials, which its announcement says reaches AI-generated ones.
For a Lebanese business that markets through Instagram and WhatsApp to customers in Europe, the United States or the diaspora, he suggests these texts are worth reading even if no one has yet asked for compliance. Both ask that evidence be what it appears to be.
The twenty-minute audit
His practical suggestion is a short routine. Collect a month of posts, sort each into a row, count the generated items that landed in a real-only cell, and keep one line per item: what it is, which cell it falls in, and who confirmed it is real. Then replace, relabel or remove what is flagged. He recommends doing this every quarter and whenever a new tool or outside supplier joins the process.
The full article, with the matrix table and the audit steps, is on jonahtebaa.com under the title "Which Parts of Your Marketing May an AI Invent, and Which Must Be Real?"
