Think You Can Spot AI? A Confidence Subgroup Result

70% of respondents who described themselves as extremely confident about identifying AI-generated ads classified at least one AI-generated study image as real in Magic Hour's four-image task.

This is an exploratory subgroup finding from our July 2026 survey of 1,019 U.S. adults. The number of respondents in the extremely confident subgroup is not published, so the 70% figure has an unknown subgroup base and no published subgroup confidence interval.

The result challenges a familiar assumption, “I would know it when I saw it,” within one specific task. It does not show that confident people perform worse than less-confident people, that every image fooled them, or that seven in ten Americans cannot recognize AI.

This article is a focused reading of the existing AI advertising survey, not a new experiment. The cover is a conceptual illustration, not one of the survey's test images.

Exploratory result: 70% of extremely confident respondents classified at least one AI study image as real in a four-image task. Subgroup count unpublished; full study 1,019 U.S. adults; July 10-11, 2026.

What Participants Were Asked to Do

The original questionnaire presented four ad images and asked respondents to select the ones they believed were AI-generated. The task combined AI-generated images and a real image.

The 70% result concerns one direction of error: an AI-generated image being classified as real. It counts people in the extremely confident group who made at least one such error. It is not average accuracy across the four images.

Separately, 1.4% of the full sample classified all four images correctly. A perfect score required getting every classification right. One incorrect answer was enough to miss that standard, so the perfect-score rate must not be reported as per-image detection accuracy.

Confidence Was Not a Guarantee in This Task

Confidence is a person's assessment of their own ability. The identification task records their answers about particular images. The published subgroup finding shows that an expression of extreme confidence did not guarantee avoiding an AI-as-real mistake in this task.

It does not establish a relationship between confidence level and error rate across all participants. To make that comparison, we would need the counts and error rates for each confidence group, along with an appropriate uncertainty analysis.

The source article also reports that 85% were at least slightly confident about identifying AI ads. That is a broader confidence category than “extremely confident.” It must not be substituted for the subgroup denominator or paired with the 70% result as if the two describe the same people.

Why Four Images Cannot Describe the Whole Internet

An identification result depends on the images shown, how they are displayed, the question wording and the task design. A different selection of ads could produce a different result.

The source page does not provide a downloadable set of test images, answer key, randomized presentation details, response-level files or subgroup cross-tabs. That limits independent replication. This article therefore treats the 70% figure as a reported exploratory result, not an independently reproduced estimate of the public's detection ability.

The result should not be extended to video, audio, every current image model or a reader's ability to identify all synthetic media. It also says nothing about whether a particular AI detector works.

What a Reader Can Take Away

The modest practical lesson is to avoid treating personal certainty as proof of an image's origin. Visual inspection and a confident guess are not the same as checking an asset's source.

When an image matters to a decision, the useful questions concern provenance: who published it, what its creator says about its production, and what evidence supports the event or product it depicts. That is an editorial recommendation, not a verification method tested by this survey.

For advertisers, an asset looking plausible should not be confused with its claims being substantiated. The survey tested classification of images, not product truthfulness, effectiveness or compliance.

A Citation-Ready Finding

Magic Hour's published survey reports that 70% of its extremely confident respondents classified at least one AI-generated ad image as real in a four-image task; the subgroup count is not published.

Include the task and missing-base qualification when citing the figure. Do not rewrite it as “70% of people cannot spot AI” or “confident viewers are the easiest to fool.”

Methodology and Limits

SurveyMonkey conducted the survey for Magic Hour July 10-11, 2026. The full sample comprised 1,019 U.S. adults. Responses were unweighted. The 70% statistic is a confidence subgroup result, not a percentage verified against the full sample of 1,019.

The published summary does not supply the subgroup count, subgroup uncertainty or underlying response-level files. We have not reconstructed those values from rounded percentages. The result is task-specific and exploratory, and the full-sample margin of error must not be applied to this subgroup.

This article publishes no identities, private responses or user-generated media. See the full survey results and limitations for the related confidence, trust and classification measures.