When it earns its place

The NYU and Emory field experiment ran banner ads across the Google Display Network and found fully AI-created ads achieved about nineteen percent higher click-through than ads made by human experts. That is a real result on real traffic, and it is the strongest number in this session. The same experiment found that AI-modified human designs failed to show improvements and sometimes performed worse, which is a strange and useful detail: the premium did not come from adding a model to a human process, it came from ads that were made entirely that way.

The Taboola study is bigger and more cautious. Across more than five hundred million impressions and three million clicks, AI ads took 0.76 percent click-through against 0.65 percent for human ads in the raw data, but the release states plainly that they "performed comparably when researchers applied the tightest statistical controls." And its most interesting finding is not about AI at all: the ads that did the best of any group were AI-generated ads that did not look AI-generated, significantly outperforming both human-made ads and AI ads that read as artificial.

Read on its own, this ledger says generated visuals can match or beat human ones on the metric advertisers care about. It does not say anything about what happens when people can tell.

Both are advertising studies, which means the outcome is a click, not a purchase, a brand impression or a reader's opinion of a report. An illustration in an annual review or on a product surface is not a banner ad, and nothing in either study speaks to it. Both are also recent enough that nobody has replicated them; the field-experiment numbers reached this page through the authors' own research summary and a write-up rather than the working paper itself, which sat behind a block, so the nineteen percent is a well-corroborated figure rather than one read from the source PDF.

The honest handling is the one session 1 used on vendor documentation: take the number, name where it came from, and do not put it in a headline without the qualification that shares its paragraph. Anyone quoting the nineteen percent without the thirty-one and a half from the very same experiment is quoting half a result.

The plus nineteen and the minus thirty-one and a half come from the same experiment, which is what makes this a finding rather than a disagreement between studies. The measured performance premium is conditional on non-detection. The measured trust penalty is conditional on detection or disclosure. They are two readings of one mechanism, and which one you get is decided by what the audience knows.

The TBWA and Ideally survey adds the part that removes the easy escape. Hybrid execution, a human finishing a generated image, reduces the trust penalty but does not remove it. And adding an AI watermark, the thing a responsible team reaches for, "actively erodes trust further, turning transparency into a measurable liability rather than a neutral safeguard." The Nuremberg survey found the same direction: identical ads labelled as AI-generated were rated more negatively, particularly on emotional appeal.

So a strategy that banks on the premium is a strategy with an expiry date, because disclosure is moving toward default and because one in two people in that survey already suspected AI in human-made work. What survives is the Taboola finding, arriving from the pro-AI side of the ledger: the work that wins is work that does not read as AI. Not because it is hiding, but because it was genuinely art-directed and finished by a person. That is an argument for craft and against volume, and it is the argument the five previous sessions have been building.

• The Getty survey sells the answer it found. Nearly ninety percent wanting disclosure and ninety-eight percent calling authenticity pivotal come from a company whose business is licensing human-made imagery and which is litigating against a model provider. The fieldwork also ran from July 2022 to September 2023, which makes it the oldest data here by a wide margin, in a subject that changes yearly.
• The TBWA and Ideally work is industry research, not a peer-reviewed study, with about 2,500 respondents across about 3,500 evaluations. Its watermark finding is the one that most deserves replication and has not had it.
• The Nuremberg survey is representative and small in scope: a thousand respondents in each of three countries, plus one labelling experiment in Germany. Its most transferable finding is a moderator rather than a headline: AI-generated ads were more accepted for innovative, high-technology products and less accepted for traditional ones.
• Several widely circulated figures are not on this page at all. Numbers from marketing-platform blogs claiming particular percentages of consumers trusting brands less, or particular click-through gains, reached this build as single snippets from vendor-published sources with no methodology attached. They are excluded rather than repeated, which is also why you will not see a figure here for how accurately anyone detects AI images.

One: what is the picture claiming? Decorative claims nothing; explanatory claims that the world works this way; evidential claims that this happened or this exists. Evidential is a no everywhere, because a generated image of a real thing is a false statement regardless of how it was labelled, and the trust penalty is the least of what follows.

Two: how much does this category live on trust? The Nuremberg survey found generated ads more accepted for innovative, high-technology products and less for traditional ones; the TBWA work found high-trust sectors such as airlines and banking hit hardest. A coffee subscription sits in the middle. A hospital does not.

Three: will it read as AI? This is the craft question and it is the one you control. The Taboola result says ads that did not look AI-made beat everything, including human work. Everything in sessions 1 to 5, a locked kit, a composed frame, refused lettering and hand-set type, is how a picture stops reading as generated. A yes on the grid is conditional on that pipeline, not on the tool.

The rule is worth four lines on one page, because in practice it is applied by people who were not in this session. Name the artefact classes you make and put each in one of the three verdicts. Name the two or three categories your organisation works in and mark their sensitivity. State the pipeline conditions that a yes depends on, in the words of the control sheet: dated version, named kit, reserved zone, hand-set type, named comparer. And state who decides when the grid says hybrid, by role, because "hybrid only" without a named person is a yes with extra words.

Then the disclosure question, which this page deliberately does not answer for you: what the law requires, where, and by when is owned by learn-ai-media-with-phoebe session 6, and it moves. The evidence here tells you what disclosure costs; the other course tells you when you have no choice. A rule that names both, and says which one wins when they conflict, is a rule that survives a legal review.