Improving human performance in deepfake detection: limits of aggregation and human–AI collaboration
Source record
kramer-2026-aggregation-human-ai· 2026 · Cognitive Research: Principles and Implications
Evidence tier: B (peer-reviewed-journal) · Verification: verified-abstract (compiler confidence: medium)
Bottom line: Individuals detected deepfake videos above chance but modestly; whole-sample aggregation helped, but confidence did not discriminate correct from incorrect trials, confidence-weighted pairing reduced accuracy, and gains from accurate AI guidance were offset by following incorrect advice.
Citation
Robin S. S. Kramer, Victoria P. M. Lister (2026). Improving human performance in deepfake detection: limits of aggregation and human–AI collaboration. Cognitive Research: Principles and Implications. https://doi.org/10.1186/s41235-026-00758-2
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#kramer-2026-aggregation-human-ai
Study design
| Field | Value |
|---|---|
| Design | Four experiments, real-vs-deepfake video judgements. Exp1: response aggregation across the sample. Exp2: confidence ratings, metacognitive sensitivity, confidence-weighted fusion of participant pairs. Exp3-4: decisions before and after simulated algorithmic guidance, varying guidance accuracy and deepfake prevalence |
| Modality | video |
| Stimuli | YouTube clips and their corresponding deepfakes (generator not in abstract) |
| Participants (N) | not recorded |
| N detail | not in abstract |
| Topics | human-detection, crowds, aggregation, metacognition, human-ai-teaming |
Results
| Metric | Value | Note |
|---|---|---|
| Individual accuracy | above chance but modest and error-prone | no values in abstract |
| Aggregation across full sample (Exp1) | reliable accuracy gains | wisdom of the crowd |
| Metacognitive sensitivity (Exp2) | no evidence | confidence modestly higher for correct decisions but did not reliably discriminate at trial level |
| Confidence-weighted fusion of pairs (Exp2) | reduced accuracy | |
| Simulated AI guidance (Exp3-4) | helped when highly accurate; offset by following incorrect advice | reliance unaffected by deepfake prevalence |
Moderators
- Guidance accuracy
- Aggregation method
Limitations
- Values not checked against full text
- Simulated, not real, algorithmic guidance
Cited in
- MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
Verification
| Field | Value |
|---|---|
| Status | verified-abstract |
| Compiler confidence | medium |
| Checked on | 2026-10-18 |
| Via | OpenAlex abstract record for doi:10.1186/s41235-026-00758-2 |
Values on this page have not all been checked against the full text. Confirm them against the primary source before high-stakes use.