AI ad survey: 98.6% missed at least one image in a 4-image test

The same survey found 59.5% said discovering a brand used AI advertising would leave their trust unchanged or higher, while 40.5% said trust would decrease. These are stated reactions to survey questions, not observed purchase behavior.

Key findings

Methodology and limitations

SurveyMonkey conducted the survey for Magic Hour from July 10 through July 11, 2026. The sample contained 1,019 U.S. adults. Data was unweighted. The stated approximate margin of error for the full sample is ±3 percentage points at a 95% confidence level.

• Four-image task: a perfect score required classifying all four images correctly, so one error counted as not perfect.
• Stimulus dependence: results depend on the exact ads, image quality, display, question wording and order used in this study.
• Self-reporting: trust, emotion, influence and regulation responses are attitudes rather than observed purchases or behavior.
• Subgroups: age, gender, income and regional percentages have smaller bases and wider uncertainty; subgroup counts are not published on this page.
• Missing research materials: the questionnaire, randomized order, source images, answer key, response-level data and crosstabs are not downloadable here, limiting independent replication.
• No longitudinal inference: a July 2026 snapshot cannot establish how perceptions have changed since then.

59.5% said AI use would not reduce brand trust

When asked how learning that a brand used AI-generated advertising would affect trust, 59.5% said they would trust the brand the same or more; 40.5% said trust would decrease. The survey did not expose respondents to different disclosure treatments or measure purchase behavior, so it cannot show that disclosure raises trust.

Exploratory subgroup results included 42.6% of women reporting lower trust. The original article also reported a 37.8% value for respondents ages 45–60+, but that age label combines categories ambiguously; without the questionnaire and base sizes, it should not drive targeting decisions.

Curiosity, interest and deception were all common

Because emotions could be multi-select, positive and negative totals can overlap. A respondent could be both interested and uncomfortable. The result supports mixed reactions, not a conclusion that the public accepts or rejects AI ads.

Among respondents ages 18–29, 15% selected excited and 26% selected impressed; reported values for ages 60+ were 4% and 8%. Feeling deceived was similar in the reported gender split: 27.6% of women and 28.9% of men. Treat these as exploratory because subgroup bases are unavailable.

98.6% missed at least one image in the identification task

Eighty-five percent said they were at least slightly confident that they could identify AI-generated ads, but 1.4% classified all four study images correctly. The gap is between stated confidence and a strict perfect-score outcome, not between confidence and average per-image accuracy.

• Under 30: 93% reported confidence; 1.5% achieved a perfect four-image score.
• Extremely confident respondents: 70% classified at least one AI-generated study image as real.
• Age 60+: 68.4% classified the real study image as AI-generated.
• Gender and confidence: 34.3% of men and 13.0% of women selected extreme confidence.
• Real image misclassification: reported rates were 63.7% for women and 56.8% for men.

These values are diagnostic of the particular task. They do not validate an AI detector, generalize to video or show how current 2026 models would perform on the same respondents.

79.9% said AI ads could influence them at least slightly

The question measured perceived influence, not conversion. Reported values were 83.4% for men and 76.6% for women. Among extremely confident respondents, 91.2% said AI ads could affect them at least slightly.

Self-assessed influence can differ from observed behavior. A useful follow-up experiment would randomize otherwise identical AI-assisted and non-AI creative, use the same audience and offer, disclose consistently and measure qualified conversion rather than asking respondents to predict themselves.

90.1% supported some government oversight

Support covered different policy preferences, from disclosure to stricter regulation. Among respondents ages 60+, 58.2% said AI-generated ads should always be disclosed; the reported value was 48% among ages 18–44.

Exploratory findings included 62.4% in Pacific and Mid-Atlantic regions supporting full disclosure, 50.2% in the Pacific and 50.0% in New England supporting strict regulation, and 58.5% among respondents reporting $125,000–$149,999 income supporting strict regulation. Do not compare these subgroups without their base sizes and uncertainty.

Trust in AI ads from unfamiliar brands varied by subgroup

Reportedly, 43.9% of men and 26.0% of women said they were very or extremely likely to trust an AI ad from an unfamiliar brand. Thirty-five percent of respondents ages 60+ selected not likely at all.

Several income-bracket estimates in the original analysis reached 82%–95% trust. Those figures are especially vulnerable to small cell sizes; without subgroup counts and confidence intervals, they are not reliable grounds for media targeting or product strategy.

What advertisers can responsibly take from the study

• Do not treat visual confidence as verification. Review source files, generated claims, product details, people and disclosures.
• Do not infer behavior from stated attitudes. Measure qualified clicks, activation, purchases, refunds and complaints in controlled tests.
• Preserve substantiation. The FTC’s advertising guide says objective ad claims need a reasonable basis before dissemination. AI generation does not lower that requirement.
• Treat subgroup findings as hypotheses. Require base sizes and a preregistered follow-up before changing targeting.
• Disclose under current rules. Apply the destination platform, format, jurisdiction and advertiser-policy requirements in force when publishing.