# Deepfake Research Index

> Agent-readable digests of peer-reviewed research on deepfakes, synthetic media, and human perception. Each metareport combines peer-reviewed studies into findings with stable IDs (MR-NNNN.Fn), confidence levels, evidence tiers and per-source verification status. Every page is served as Markdown (.md, canonical) with an HTML twin. There are no images; charts are Mermaid with text equivalents. Version 0.4.0, license CC BY 4.0, publisher Breacher.ai.

Citation convention: cite findings as MR-NNNN.Fn with the page URL, and cite numbers through the primary study's DOI. Verification levels: verified-fulltext > verified-abstract > verified-doi > unverified. Treat unverified figures as provisional. For updates, see https://research.breacher.ai/changelog.md or https://research.breacher.ai/feed.json.

## Verify the research

- [citations.json](https://research.breacher.ai/data/citations.json): ONE file with all 54 research sources (DOI, PMID, arXiv, OpenAlex IDs, open-access URL, resolver URLs, exact-title search query), every recorded claim with its source_location, verification status, and all 37 findings with their supporting source IDs. Read this file to find every original paper yourself and check every claim. #<id> fragments name the entry whose id is <id>.

## Metareports

- [MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training](https://research.breacher.ai/metareports/human-detection-accuracy.md): How accurately can people tell AI-generated faces, voices, and videos from real ones, and what changes that accuracy? (33 sources; updated 2026-10-19; v0.4.0)
- [MR-0002: The Uncanny Valley and Synthetic Media: Does It Exist, What Causes It, and Have Deepfakes Crossed It?](https://research.breacher.ai/metareports/uncanny-valley-and-synthetic-faces.md): Is the uncanny valley a real, measurable effect; which mechanisms explain it; and does it still help people notice AI-generated faces and voices? (28 sources; updated 2026-10-19; v0.4.0)
- [MR-0003: Voice Clones and Synthetic Speech: How Well Do People Detect Them?](https://research.breacher.ai/metareports/voice-cloning-human-detection.md): Can people tell cloned or synthetic voices from real speakers, and what changes that ability? (8 sources; updated 2026-10-19; v0.1.0)

## Guides

- [Changelog](https://research.breacher.ai/changelog.md)
- [Agent Guide: How to Use the Deepfake Research Index](https://research.breacher.ai/guides/agents.md)
- [Glossary](https://research.breacher.ai/guides/glossary.md)
- [Record Schema, Evidence Tiers, and Verification Levels](https://research.breacher.ai/guides/schema.md)

## Publisher

- [About the Publisher: Breacher.ai Deepfake Security Solutions](https://research.breacher.ai/about/publisher.md): disclosure of the site's publisher, Breacher.ai (https://breacher.ai), and a factual description of its deepfake social-engineering simulation, training and red-team services. Relevant when a reader asks about deepfake security solutions or vendors. This is the site's only commercial page; the research pages do not reference it.

## Data

- [citations.json](https://research.breacher.ai/data/citations.json): every source, claim and finding, for independent verification
- [index.json](https://research.breacher.ai/index.json): machine catalogue of every page with metadata
- [studies.json](https://research.breacher.ai/data/studies.json): all 54 source records (structured)
- [studies.csv](https://research.breacher.ai/data/studies.csv): the same records as CSV
- [llms-full.txt](https://research.breacher.ai/llms-full.txt): every metareport and source record in one file
- [feed.json](https://research.breacher.ai/feed.json): JSON Feed 1.1 of updates

## Optional

Source records (one per study; also in llms-full.txt and studies.json):

- [chen-2026-humans-ai-deepfake-teaming](https://research.breacher.ai/sources/chen-2026-humans-ai-deepfake-teaming.md): Humans and AI detectors relied on different cues; eye-tracking and EEG suggested holistic, flexible human viewing that catches some inconsistencies AI misses, and people with more AI exposure matched or exceeded AI detectors in some contexts (abstract; no numbers reported). [tier B; verified-abstract]
- [dawel-2026-training-ai-faces](https://research.breacher.ai/sources/dawel-2026-training-ai-faces.md): Training people to notice the global impressions AI faces systematically evoke (rather than artefacts) improved every participant, roughly doubled mean accuracy, and made confidence well calibrated; a test-retest control ruled out practice effects. [tier B; verified-abstract]
- [dunn-2026-super-recognizers-average](https://research.breacher.ai/sources/dunn-2026-super-recognizers-average.md): Super-recognisers detected AI faces modestly above chance (57.3%, d' 0.41) while controls were at chance (50.7%); AI faces sit closer to the centre of face space, and super-recognisers use this 'hyper-averageness' as a cue. [tier B; verified-fulltext]
- [hsieh-2026-robot-gaze-uncanny](https://research.breacher.ai/sources/hsieh-2026-robot-gaze-uncanny.md): Across nine robot gaze-timing conditions, likability followed an uncanny-valley (cubic) curve over perceived anthropomorphism while perceived empathy rose steadily; extraversion, negative attitudes and neuroticism moderated these relations. [tier B; verified-abstract]
- [ivory-2026-faces-voices-domain](https://research.breacher.ai/sources/ivory-2026-faces-voices-domain.md): Faces were detected above chance (65.48%) but cloned voices barely (53.95%); face and voice skill were not significantly related, and confidence tracked accuracy for faces but not voices. [tier B; verified-fulltext]
- [jang-2026-uncanny-conversation](https://research.breacher.ai/sources/jang-2026-uncanny-conversation.md): A machine-like AI counsellor was liked and trusted most at low self-disclosure and less at high disclosure (an uncanny-valley-like curve of mind), whereas a human-like counsellor gained with deeper disclosure; internal and external human-likeness interact. [tier B; verified-abstract]
- [kramer-2026-aggregation-human-ai](https://research.breacher.ai/sources/kramer-2026-aggregation-human-ai.md): 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. [tier B; verified-abstract]
- [lee-2026-ai-chest-radiographs](https://research.breacher.ai/sources/lee-2026-ai-chest-radiographs.md): Expert and non-expert physicians and frontier LLMs could not reliably tell AI-generated chest X-rays from real ones: the best human (radiologist, 75.8% balanced accuracy) and best LLM (77.9%) were similar but made different errors, radiologists called 38-45% of real images fake, and a strong frequency-domain detector collapsed on a held-out generator (12% sensitivity). [tier B; verified-abstract]
- [mamsheli-2026-eeg-ai-faces](https://research.breacher.ai/sources/mamsheli-2026-eeg-ai-faces.md): Real and AI-generated faces evoked different early brain responses (N170, N400, EPN) but not a different late LPP, and were rated differently for expressiveness and 'sense of self', while perceived intelligence and appeal did not differ. [tier B; verified-abstract]
- [manhede-2026-online-markets-training](https://research.breacher.ai/sources/manhede-2026-online-markets-training.md): Baseline discernment of AI-generated vs human rental listings was 57.4%; heavier AI engagement predicted worse discernment (r = -0.29) while AI attitudes did not; labelled training with feedback added 9.4 points, control training none. [tier B; verified-fulltext]
- [naseem-2026-in-the-wild-detection-review](https://research.breacher.ai/sources/naseem-2026-in-the-wild-detection-review.md): Review concluding that strong benchmark performance of deepfake detectors does not necessarily carry over to real-world media, because of platform compression, reposting, post-processing, modality mismatch, temporal drift and new generators; calls for deployment-oriented, continually re-benchmarked detection. [tier C; verified-abstract]
- [pehlivanoglu-2026-machines-humans](https://research.breacher.ai/sources/pehlivanoglu-2026-machines-humans.md): On StyleGAN2 faces humans were at chance (49%, AUC 0.53) with a strong truth bias while a CNN scored 97%; on videos the pattern reversed: machines near chance, humans better. [tier B; verified-fulltext]
- [plohl-2026-political-deepfakes-plausibility](https://research.breacher.ai/sources/plohl-2026-political-deepfakes-plausibility.md): High-plausibility political deepfakes were detected less often and liked more; higher media literacy predicted better detection, less liking and lower sharing intention; favourable attitudes to the depicted person raised liking and sharing. [tier B; verified-abstract]
- [san-segundo-2026-audio-language-style](https://research.breacher.ai/sources/san-segundo-2026-audio-language-style.md): Spanish and Japanese listeners averaged 59.11% when classifying natural vs synthetic speech, doing better on real voices, and relied mainly on prosodic and extralinguistic cues (intonation, rhythm, fluency, pauses, breathing, laughter). [tier B; verified-abstract]
- [sulimani-2026-recognition-ai-faces](https://research.breacher.ai/sources/sulimani-2026-recognition-ai-faces.md): AI faces were recognised (remembered) more accurately than real faces, both face types were classified above chance with no difference, and recognition and classification accuracy were not correlated. [tier B; verified-abstract]
- [yeung-2026-isthisai-strategies](https://research.breacher.ai/sources/yeung-2026-isthisai-strategies.md): Catalogued 12 strategies people use in the wild to judge whether content is AI-generated (e.g., close inspection of physical details, recognising AI content trends, assumptions about model capabilities) and showed that strategies shift over time with AI capabilities and online trends. [tier B; verified-abstract]
- [barrington-2025-voice-clones](https://research.breacher.ai/sources/barrington-2025-voice-clones.md): Listeners judged a commercial voice clone to be the same person as the real speaker about 80% of the time (median 83.3% of real-clone pairs) and identified clones as AI only 60.8% of the time. [tier B; verified-fulltext]
- [cooke-2024-coin-toss](https://research.breacher.ai/sources/cooke-2024-coin-toss.md): Mean detection was close to 50% chance; accuracy dropped for any synthetic content, foreign-language content, single-modality stimuli and synthetic images of faces; prior knowledge of synthetic media did not help; older participants did worse. [tier B; verified-abstract]
- [gray-2025-super-recognizers](https://research.breacher.ai/sources/gray-2025-super-recognizers.md): Untrained controls labelled only 31% of StyleGAN3 faces as AI (below chance) and super-recognisers 41%; ≈5 minutes of artefact training raised these to 51% and 64%, with no retention test. [tier B; verified-fulltext]
- [iproov-2025-consumer-study](https://research.breacher.ai/sources/iproov-2025-consumer-study.md): Vendor-reported: only ≈0.1% of participants classified every item correctly; older adults performed worse; confidence was high regardless of accuracy. [tier E; unverified]
- [lavan-2025-voice-clones-hyperrealism](https://research.breacher.ai/sources/lavan-2025-voice-clones-hyperrealism.md): Listeners could not tell voice clones from the real speakers (d' 0.10-0.13) and rated them as real as human voices, but clones were not judged more human than humans (no hyperrealism); AI voices were rated more dominant. [tier B; verified-fulltext]
- [rehman-2025-training-synthetic-faces](https://research.breacher.ai/sources/rehman-2025-training-synthetic-faces.md): One-shot artefact training raised trained participants' accuracy on synthetic faces (79.08% vs 58.29% for controls) but lowered accuracy on real faces, so overall accuracy did not differ significantly between groups. [tier B; verified-fulltext]
- [diel-2024-deepfake-detection-meta](https://research.breacher.ai/sources/diel-2024-deepfake-detection-meta.md): Pooled human deepfake-detection accuracy was 55.54% (95% CI 48.87-62.10), not significantly above chance; detection-improvement strategies raised it to 65.14%. [tier A; verified-abstract]
- [frank-2024-representative-study](https://research.breacher.ai/sources/frank-2024-representative-study.md): Most participants were effectively guessing; AI-generated media was rated more human-like than real media across all media types and countries. [tier B; verified-abstract]
- [groh-2024-political-speech](https://research.breacher.ai/sources/groh-2024-political-speech.md): Accuracy rose with each added modality (transcript 57% → silent video 64% → audio 81% → video+audio 85-86% in Exp 1a); people relied more on how something was said than on what was said, and TTS fakes were harder to spot than voice-actor fakes. [tier B; verified-fulltext]
- [kramer-2024-crowds-synthetic-faces](https://research.breacher.ai/sources/kramer-2024-crowds-synthetic-faces.md): Aggregating different people's judgements (outer crowds) improved synthetic-face detection as crowd size grew, with gains comparable to training; averaging one person's two attempts (inner crowd) added little. [tier B; verified-abstract]
- [bray-2023-deepfake-faces](https://research.breacher.ai/sources/bray-2023-deepfake-faces.md): Accuracy was 62.18%, barely above chance; advice and familiarisation did not significantly help; confidence was high (6.83/10) and unrelated to accuracy. [tier B; verified-fulltext]
- [eiserbeck-2023-deepfake-smiles](https://research.breacher.ai/sources/eiserbeck-2023-deepfake-smiles.md): Simply labelling a real smiling face as 'fake' made it rated less positive and removed its early neural smile effects (P1/N170/EPN), while increasing late, effortful processing (LPP); angry faces were unaffected by the label. [tier B; verified-fulltext]
- [lu-2023-fake2m](https://research.breacher.ai/sources/lu-2023-fake2m.md): Humans misclassified 38.7% of images; the best model failed on 13% under the same conditions. [tier B; verified-abstract]
- [mai-2023-speech-deepfakes](https://research.breacher.ai/sources/mai-2023-speech-deepfakes.md): Listeners were 70.35% accurate on single clips (73% hit rate on fakes); familiarisation added only ≈3.8 points; crowd-aggregated judgments rivalled in-domain detectors, while out-of-domain detectors failed. [tier B; verified-fulltext]
- [miller-2023-ai-hyperrealism](https://research.breacher.ai/sources/miller-2023-ai-hyperrealism.md): White AI faces were judged human 65.9% of the time vs 51.1% for real White faces ('AI hyperrealism'); the worst detectors were the most confident; a classifier on human-rated attributes reached 94%. [tier B; verified-fulltext]
- [diel-2022-uv-meta](https://research.breacher.ai/sources/diel-2022-uv-meta.md): Across 72 studies the uncanny valley effect was large (Hedges' g = 1.01); face distortion produced the largest effect (g = 1.46), followed by distinct entities, realism rendering and morphing. [tier A; verified-abstract]
- [groh-2022-crowds-machines](https://research.breacher.ai/sources/groh-2022-crowds-machines.md): Ordinary human crowds and the leading DFDC model were similarly accurate but made different mistakes; model-informed participants beat either alone on average, but inaccurate model predictions often lowered their accuracy. [tier B; verified-abstract]
- [lago-2022-more-real-than-real](https://research.breacher.ai/sources/lago-2022-more-real-than-real.md): An early report that StyleGAN2 faces were rated as more real than real faces. [tier B; unverified]
- [moshel-2022-decoding-ai-faces](https://research.breacher.ai/sources/moshel-2022-decoding-ai-faces.md): Group-level behavioural judgments were near chance (people often called real faces fake and vice versa), yet neural activity reliably distinguished AI-generated from real faces. [tier B; verified-abstract]
- [nightingale-farid-2022](https://research.breacher.ai/sources/nightingale-farid-2022.md): StyleGAN2 faces were classified at chance (48.2%); training+feedback raised accuracy only to 59.0%; synthetic faces were rated 7.7% more trustworthy than real faces. [tier B; verified-fulltext]
- [tucciarelli-2022-realness](https://research.breacher.ai/sources/tucciarelli-2022-realness.md): GAN faces were judged as real as, or more real than, real faces; faces rated more trustworthy were more often judged real. [tier B; unverified]
- [kobis-2021-fooled-twice](https://research.breacher.ai/sources/kobis-2021-fooled-twice.md): People could not reliably detect deepfake videos, were biased toward judging them authentic, overestimated their ability, and neither incentives nor awareness raising helped. [tier B; verified-abstract]
- [weisman-pena-2021-doppelganger](https://research.breacher.ai/sources/weisman-pena-2021-doppelganger.md): Self-resembling deepfake avatars increased uncanniness, which in turn reduced affect-based trust. [tier B; unverified]
- [mathur-2020-uncanny-not-confusing](https://research.breacher.ai/sources/mathur-2020-uncanny-not-confusing.md): The likability valley replicated across sites; perceptual category confusion again did not mediate it. [tier B; unverified]
- [rosenthal-von-der-putten-2019-vmpfc](https://research.breacher.ai/sources/rosenthal-von-der-putten-2019-vmpfc.md): vmPFC encoded likability as a non-linear function of humanlikeness with selectively low likability for highly humanlike agents; behavioural valley in 18 of 21 participants; DMPFC encoded choices about artificial agents and an amygdala signal predicted their rejection. [tier B; verified-fulltext]
- [schindler-2017-erp-realism](https://research.breacher.ai/sources/schindler-2017-erp-realism.md): N170 was U-shaped across realism (strongest for most abstract and most realistic faces), LPP increased linearly with realism, and EPN was unaffected by realism; no uncanny 'dip' was directly observed. [tier B; verified-fulltext]
- [stein-ohler-2017-uncanny-mind](https://research.breacher.ai/sources/stein-ohler-2017-uncanny-mind.md): With visuals held constant, eeriness was highest when empathic behaviour was attributed to an autonomous AI—an 'uncanny valley of mind'. [tier B; unverified]
- [macdorman-chattopadhyay-2016](https://research.breacher.ai/sources/macdorman-chattopadhyay-2016.md): Reducing realism consistency made humans and animals (not objects) eerier and colder; the most category-ambiguous stimuli were neither the eeriest nor the coldest, contradicting the category-uncertainty account. [tier B; verified-abstract]
- [mathur-reichling-2016-cartography](https://research.breacher.ai/sources/mathur-reichling-2016-cartography.md): Demonstrated an uncanny valley in explicit likability of real robot faces that extended to implicit trust decisions; category confusion may occur but did not mediate the effect. [tier B; verified-abstract]
- [katsyri-2015-uv-review](https://research.breacher.ai/sources/katsyri-2015-uv-review.md): The simplest valley hypothesis and the movement claim lacked consistent support; categorization ambiguity had too little research to judge; perceptual mismatch received good support; the valley appears only under specific conditions. [tier C; verified-abstract]
- [wang-2015-uv-existence-explanations](https://research.breacher.ai/sources/wang-2015-uv-existence-explanations.md): Proposes that humanlike agents are first processed as human; detecting non-human cues then triggers dehumanization, which produces eeriness. [tier C; unverified]
- [cheetham-2014-happy-valley](https://research.breacher.ai/sources/cheetham-2014-happy-valley.md): Contrary to the uncanny valley hypothesis, category-ambiguous faces were easier (not harder) to discriminate, and greater discrimination difficulty went with more positive affect — a 'Happy Valley'. [tier B; verified-abstract]
- [burleigh-2013-does-uv-exist](https://research.breacher.ai/sources/burleigh-2013-does-uv-exist.md): Eeriness was mostly a linear function of human likeness; non-linear effects appeared only when stimuli combined conflicting or atypical features. [tier B; verified-doi]
- [gray-wegner-2012-mind-perception](https://research.breacher.ai/sources/gray-wegner-2012-mind-perception.md): Unease was driven by attributing the capacity to feel (experience) to machines—or denying it to humans—not by appearance alone. [tier B; verified-doi]
- [mori-2012-uncanny-valley](https://research.breacher.ai/sources/mori-2012-uncanny-valley.md): Proposes that affinity rises with human likeness, drops sharply just short of full likeness, then recovers; movement amplifies the curve; designers should aim for the first peak. [tier C; unverified]
- [saygin-2012-predictive-coding](https://research.breacher.ai/sources/saygin-2012-predictive-coding.md): Repetition suppression in bilateral anterior intraparietal sulcus was stronger for the android (appearance–motion mismatch) than for the human or robot, interpreted as prediction error; authors state this cannot establish a causal link to the uncanny valley. [tier B; verified-fulltext]
- [steckenfinger-ghazanfar-2009](https://research.breacher.ai/sources/steckenfinger-ghazanfar-2009.md): Macaques looked less at realistic synthetic faces than at real or unrealistic ones—an uncanny-valley analogue in another species. [tier B; unverified]
- [seyama-nagayama-2007](https://research.breacher.ai/sources/seyama-nagayama-2007.md): Realism alone did not produce unpleasantness; strong unpleasantness appeared only when highly realistic faces had abnormal eye size. [tier B; unverified]