Can people reliably spot content made by AI?

This PRISMA-guided systematic review finds that human ability to tell generative AI content from human-produced content, across text, image and voice, "varied widely but generally clustered around chance performance." The search ran in Scopus, restricted to 2025 and 2026, and returned 22,541 records. Titles and abstracts were screened in relevance order until 1,200 had been reviewed, and 30 studies entered the synthesis. The review's summary is blunt: "humans are generally unreliable detectors of gen AI content." The Discussion makes the same point about the average. Humans "do not distinguish AI-generated content from human-generated content reliably better than chance," and human accuracy "has not meaningfully improved over time or at least has not improved at a pace that keeps up with the increasing realism of gen AI content."

The review does not compute a pooled estimate. It says "substantial methodological heterogeneity across studies limited the ability to compute pooled effect sizes," so its case rests on consistency across studies. It also reports that screening reached saturation, since "no eligible studies had been identified in the last 100 records" before it stopped. The modality pattern is the most specific claim in the excerpt. Voice detection succeeds more often than image detection, and image detection more often than text. For voice, the review cites work suggesting listeners may pick up "subtle artifacts in timing, pitch, or prosody." For images, the visual heuristics that might help (inconsistent reflections, unnatural textures, impossible lighting) still produce accuracy that "mostly did not vary convincingly from chance." These mechanisms are proposals from the cited literature. The review does not test them.

Read against the library, the review turns single-study findings into a cross-modality pattern. Can humans detect AI text if machines can measure it? and Can human judges detect measurable differences in AI text? both describe text that differs from human writing on measurable dimensions while judges fail to notice. This review extends that pattern across 30 studies and into image and voice. It does not check whether the measurable differences exist, because it measures human accuracy and nothing about the content itself. The text-level result, which is the most consequential for evaluation, is developed in Does polished writing actually signal better quality work?.

The excerpt does not establish several things a reader might want. It gives no per-study accuracy values, no per-modality figures, no confidence intervals and no chance baseline for each task. Its Table 1 is referenced but not included. So "around chance" is a summary of how the reviewed literature is distributed, and it carries the strength of a non-pooled review. The scope is bounded by design too: Scopus only, English only, 2025 to 2026, and the first 1,200 of 22,541 records by relevance, which the authors say may miss preprints and fast-moving computer science venues. The implication is that unaided human judgment is a weak basis for flagging AI content. The excerpt says nothing about automated detectors, and nothing about whether training or tools change the picture, so it cannot support a claim either way on those.