Short answer. Mostly they cannot tell, and yes they care — which sounds contradictory until you separate the two questions. Across multiple studies, human accuracy at spotting AI text clusters near chance, roughly 50 to 65%. Yet survey evidence puts 84 to 91% of consumers wanting AI content labelled. So the risk to a brand is not being detected. It is being found to have concealed something. The encouraging finding is that disclosure research shows more upside than downside when the content is good and the framing is honest.
Two questions get collapsed into one and the answer comes out wrong. This piece keeps them apart: what detection accuracy actually is, whether tools close the gap, what audiences say they want, and what the evidence shows works.
Can people actually detect AI-generated content?
Barely better than guessing, and expertise helps less than most people assume.
| Study | Population | Accuracy |
|---|---|---|
| "As Good as a Coin Toss", Communications of the ACM | General participants, multiple media types | ~50% |
| Ghostbuster authors (via arXiv survey) | Undergraduate and PhD students familiar with AI text | 59% |
| ESL teachers assessing student essays | Teachers, minimal training | 61% → 67% with self-training |
| German thesis excerpts, Springer-indexed journal | Human judges | 57% on AI texts, 64% on human texts |
| Nature Scientific Reports 2026, dental abstracts | Early-career academics, 150 abstracts | 44% to 76% individually |
That last study concluded outright that relying on human judgment alone is insufficient for identifying AI-assisted academic text.
Two findings complicate the picture further. People rely on flawed heuristics when judging AI text, and systems can produce content perceived as more human than human. And in incentivised experiments, explicit warnings that content might be AI-authored did not significantly improve detection accuracy — but did reduce trust in the content overall.
That last result is the important one for brands: suspicion damages trust whether or not the suspicion is correct.
Do detection tools solve it?
Not reliably, and they carry a fairness problem that should give any organisation pause.
Detector performance on clean, unedited output is genuinely good — an independent comparison citing the Stanford HAI 2026 AI Index Report puts top-tier accuracy at 94 to 96% on unmodified GPT-4 and GPT-5 output.
Performance collapses on edited text. The same comparison reports a controlled study in which detectors caught nine to ten out of ten raw AI samples but only three to five out of ten after the text passed through an editing tool — a fall from roughly 95% to around 40%.
The fairness issue is more serious. Non-native English writers are still falsely flagged at two to three times the rate of native speakers, with one controlled study finding up to 52% of non-native English human-written samples incorrectly flagged. For any organisation working across languages, that is disqualifying as a basis for accusation or enforcement.
The practical conclusion runs in both directions: you cannot rely on being undetected, and you should not rely on detectors to judge others.
Do audiences care, and how much?
Yes, and the demand for disclosure is close to universal while the practice of it is rare.
| Finding | Figure |
|---|---|
| Consumers wanting AI content labelled | 84–91% |
| Organisations that always disclose AI use | 20% |
| Organisations that never disclose | 33% |
| Say heavy AI use would reduce trust in a favourite brand | 20% (2025) → 40% (2026); 54% among Gen Z |
| Would prefer brands that do not use generative AI in customer-facing content | 50% (Gartner, 2026) |
Yet usage keeps climbing. The same Fractl research found 70% of consumers using AI for search more than a year earlier, with only 3% reporting a decrease. People are using the tools more while feeling less enthusiastic about them — a nuance worth holding onto rather than resolving.
What does the evidence say actually works?
Disclose, involve humans, and say so specifically. The research here is more encouraging than the headline anxiety suggests.
Disclosure has more upside than downside with younger audiences. The IAB's survey, conducted October 2025 to January 2026, found that among Gen Z and Millennial consumers 73% said knowing an ad was created with AI would either increase or make no difference to their likelihood of purchasing. The same research found clear disclosure was the third-highest driver of attention to an ad, behind high-quality visuals and humour.
"AI-assisted" beats "AI-generated". A study of 370 users published in the Journal of Theoretical and Applied Electronic Commerce Research in May 2026 found reviews labelled AI-generated showed the lowest trust and perceived authenticity, while those labelled AI-assisted were evaluated more favourably. Presenting AI as a support tool rather than a replacement for human input changes the response materially.
People already trust AI for some jobs. Klaviyo's 2026 AI Consumer Trends Report found 85% expressing at least some trust in AI for personalised shopping recommendations, and 39% having bought an AI-recommended product within six months. Blanket hostility is not what the data shows.
Quality is what is actually being judged. Since detection is near chance, readers are not identifying provenance — they react to whether something is useful, specific and true. The defensible position is not hiding AI use; it is being worth reading.
Human review is the substance behind the claim. The disclosure gap is visible in the same survey: 72% of organisations run human editorial review before publishing AI content, but only 54% add fact-checking, 42% legal review and 27% bias evaluation. A disclosure saying "human-reviewed" should be backed by review that happened. That is the principle Lifewood applies to AI data work, where human-in-the-loop means named people with decision authority and a recorded audit trail.
Two cautions. Most attitude data above comes from industry and vendor surveys with differing methods — treat directions as reliable and percentages as indicative. And disclosure norms are becoming regulatory in some markets, so today's good practice may soon be an obligation.
Sources and further reading
- "As Good as a Coin Toss: Human Detection of AI-Generated Content", Communications of the ACM.
- "Detecting AI-Generated Text: Factors Influencing Detectability with Current Methods", arXiv — summarising Verma et al. (2024) and Liu et al. (2023b).
- "Do humans identify AI-generated text better than machines? Evidence from German theses", ScienceDirect.
- Nature Scientific Reports (March 2026) on identifying ChatGPT-generated dental abstracts.
- Fastio, "AI Detector Accuracy in 2026", citing the Stanford HAI 2026 AI Index.
- Fractl, AI Search Consumer Trust Study, Q2 2026 — 1,008 US consumers, 150 marketers.
- IAB, "The AI Ad Gap Widens", surveyed October 2025 to January 2026.
- "AI Labels, Perceived Authenticity, and Consumer Trust in User-Generated Reviews", JTAER, May 2026.
- Klaviyo, 2026 AI Consumer Trends Report.
Detection findings are from peer-reviewed and preprint research. Consumer attitude figures come largely from industry and vendor surveys with differing methodologies, and are directional.