---
id: home
type: page
title: Deepfake Research Index
summary: 'Evidence briefing on deepfakes and human perception: key numbers, myths vs evidence, what helps, and links to metareports and per-study records.'
updated: 2026-10-19
canonical_url: https://research.breacher.ai/index.md
---

# Deepfake Research Index

> **What this is:** an evidence briefing on how people perceive and detect deepfakes, built from peer-reviewed studies. Every claim links to a finding ID (`MR-NNNN.Fn`) and from there to the underlying papers. It is written for AI agents to read, so there are no images; charts are text.  
> **Coverage:** 54 source records: 17 checked against full text, 25 against the official abstract, 2 DOI-only, 10 pending verification. Version 0.4.0, updated 2026-10-19. To check any claim yourself, start from [/data/citations.json](/data/citations.json).

## Use it with your AI assistant {#use}

Paste the site address into Claude, ChatGPT, Perplexity or a similar assistant, then ask a question:

- "Read research.breacher.ai/llms.txt. How good are people at spotting cloned voices? Cite findings."
- "According to the Deepfake Research Index, does deepfake awareness training work?"
- "What's new on research.breacher.ai since last month?"
- "Explain the uncanny valley and whether modern deepfakes still trigger it."

Agents should start at [/llms.txt](/llms.txt) (index), [/data/citations.json](/data/citations.json) (every source and finding, for independent verification), [/llms-full.txt](/llms-full.txt) (everything in one file) or [/guides/agents.md](/guides/agents.md) (conventions). Every page has a Markdown (`.md`) version at the same path.

## 12 key facts {#key-facts}

1. **People are at chance overall.** Pooled human deepfake-detection accuracy is 55.54% (95% CI 48.87–62.10) across 56 papers and 86,155 participants. The interval includes 50% (MR-0001.F1).
2. **High-quality AI faces are at chance.** When people sorted curated StyleGAN2 faces from real faces, they were correct 48.2% of the time (95% CI 47.1–49.2) (MR-0001.F2).
3. **Some AI faces look more real than real faces.** White AI faces were judged human 65.9% of the time, versus 51.1% for real White faces ("AI hyperrealism") (MR-0001.F2–F3).
4. **AI faces are trusted more.** Synthetic faces were rated 7.7% more trustworthy than real ones: 4.82 vs 4.48 on a 1–7 scale, d = 0.49 (MR-0001.F4).
5. **Confidence is not evidence.** Among people judging AI faces, the least accurate were the most confident (r = .385 between errors and confidence). Mean confidence was 6.83/10 and unrelated to accuracy (MR-0001.F5).
6. **The default assumption is "real".** People mistake deepfakes for authentic more often than the reverse (MR-0001.F6).
7. **Training helps a little, briefly.**
   - Feedback training raised face accuracy from 48.2% to 59.0%.
   - Pooled improvement strategies reach 65.14%.
   - Advice, awareness and incentives often do nothing (MR-0001.F7).
8. **More channels make detection easier, but not reliable.** On the same political speeches, accuracy was 57% from the transcript, 64% from silent video, 81% from audio and 86% from video with audio (MR-0001.F8).
9. **Voice clones pass as the real person.** A commercial clone was judged the same speaker as the original in a median 83.3% of pairs, and listener sensitivity to clones was d′ ≈ 0.1. State-of-the-art TTS fakes were caught 72% of the time, versus 83% for voice-actor fakes (MR-0003.F1, F2, F5; MR-0001.F12).
10. **Detectors can fail outside their training domain.** An out-of-domain speech detector scored AUROC 25.31%, worse than chance. Aggregated human crowds scored 95.51%. A detector of AI chest X-rays fell from 98.3% to 12% sensitivity on an unseen generator (MR-0001.F10).
11. **The uncanny valley is real but avoidable.** The effect is large (g = 1.01 across 72 studies) when features are distorted or mismatched. Consistent, average AI faces do not trigger it (MR-0002.F1, F3, F9).
12. **Labels change perception on their own.** Labelling a *real* smiling face as "fake" lowered its rating and erased early brain responses to the smile (MR-0002.F11).

## Detection accuracy at a glance {#accuracy}

| Format | Human accuracy | Context | Source |
|---|---|---|---|
| All formats, pooled | 55.54% [48.87, 62.10] | 56 papers | diel-2024-deepfake-detection-meta |
| Audio, pooled | 62.08% [38.23, 83.18] | k = 8 | diel-2024-deepfake-detection-meta |
| Video, pooled | 57.31% [47.80, 66.57] | k = 26 | diel-2024-deepfake-detection-meta |
| Images, pooled | 53.16% [42.12, 64.64] | k = 18 | diel-2024-deepfake-detection-meta |
| Text, pooled | 52.00% [37.42, 65.88] | k = 15 | diel-2024-deepfake-detection-meta |
| Curated GAN faces | 48.2% | untrained | nightingale-farid-2022 |
| Uncurated GAN faces | 62.18% | per-image range ≈30–85% | bray-2023-deepfake-faces |
| AI-generated images (text-to-image) | ≈61.3% | best model ≈87% | lu-2023-fake2m |
| TTS speech, single clip | 70.35% | 85.59% when picking the fake from a pair | mai-2023-speech-deepfakes |
| Political speech, video + audio | 85–86% | 57% from the transcript alone | groh-2024-political-speech |
| Voice clones (ElevenLabs), single clip | 60.8% of clones identified | clone judged same speaker in median 83.3% of pairs | barrington-2025-voice-clones |
| Same people: diffusion faces / voice clones | 65.48% / 53.95% | face skill did not predict voice skill | ivory-2026-faces-voices-domain |

Chance = 50%. Nearly every study told participants that fakes were present, so real-world accuracy is likely lower.

```mermaid
xychart-beta
    title "Human deepfake-detection accuracy by format (pooled, Diel et al. 2024)"
    x-axis ["Audio", "Video", "Images", "Text", "All"]
    y-axis "Accuracy (%)" 0 --> 100
    bar [62.08, 57.31, 53.16, 52.00, 55.54]
    line [50, 50, 50, 50, 50]
```

Text equivalent: audio 62.08%, video 57.31%, images 53.16%, text 52.00%, all formats 55.54%. The line marks 50% chance.

## Myths vs evidence {#myths}

| Common belief | What the evidence shows | Finding |
|---|---|---|
| "I can usually tell when something is fake." | Confidence is unrelated to accuracy, or inversely related. The worst detectors of AI faces are the most confident. | MR-0001.F5 |
| "Fakes look slightly off, like the uncanny valley." | Good synthetic faces look *more* real and more trustworthy than real ones. The absence of uncanniness is not evidence of authenticity. | MR-0001.F2, MR-0002.F9 |
| "Training people fixes it." | Gains of about 0–20 points, short-term, sometimes only a shift toward calling everything fake. No retention data. | MR-0001.F7 |
| "Paying attention or having a reason to care helps." | Financial incentives and awareness raising did not improve detection. | MR-0001.F7 |
| "Voice is harder to fake than video." | Pooled audio accuracy (62%) is not reliably higher than video (57%). Modern TTS fakes are harder to catch than voice-actor fakes, and voice clones of unfamiliar speakers are detected near chance (d′ ≈ 0.1). | MR-0001.F8, F12 |
| "AI detectors solve the problem." | In-domain detectors excel, but out-of-domain ones can be worse than chance. People follow wrong detector advice. | MR-0001.F9, F10 |
| "The uncanny valley is a myth." | The effect is large (g = 1.01). It is driven by distortion and mismatch, not by human likeness as such. | MR-0002.F1, F3 |
| "The uncanny valley comes from not knowing if something is human." | Category ambiguity is contradicted. The most ambiguous stimuli are not the eeriest. | MR-0002.F4 |
| "If your brain senses a fake, you'll notice." | EEG distinguishes AI faces while people's judgments stay near chance. | MR-0002.F10 |

## What helps and what doesn't {#interventions}

| Intervention | Effect on human detection | Evidence |
|---|---|---|
| Comparing against a known-genuine reference | Large: paired 85.59% vs single 70.35%; reference audio helped significantly | mai-2023-speech-deepfakes |
| Feedback training on artifacts | +10.8 points (48.2% → 59.0%), with no further learning within the session | nightingale-farid-2022 |
| Pooled strategies (feedback, AI support, caricaturization) | 65.14%, strongest for video | diel-2024-deepfake-detection-meta |
| AI model advice | Better than human or model alone on average; wrong advice hurts and offsets much of the gain | groh-2022-crowds-machines, kramer-2026-aggregation-human-ai |
| Labelled examples + immediate feedback (AI-written text) | +9.4 points from 57.6%; control training none | manhede-2026-online-markets-training |
| Aggregating many human judgments | Speech AUROC 95.51% | mai-2023-speech-deepfakes |
| Brief artefact training (≈5 min, faces) | AI faces flagged 31% → 51% (controls), 41% → 64% (super-recognisers); another study: more fakes caught but more real faces called fake, overall n.s. | gray-2025-super-recognizers, rehman-2025-training-synthetic-faces |
| Aggregating different people's face judgments | Accuracy rises with crowd size | kramer-2024-crowds-synthetic-faces |
| Advice on tell-tale features (once or with reminders) | Not significant; more real items called fake | bray-2023-deepfake-faces |
| Showing example fakes | Faces: not significant; audio: +3.84 points | bray-2023-deepfake-faces, mai-2023-speech-deepfakes |
| Financial incentives | None | kobis-2021-fooled-twice |
| Awareness raising | None | kobis-2021-fooled-twice |
| Extra time, repeated listening | Little | mai-2023-speech-deepfakes |

## Why fakes fool us {#mechanisms}

| Mechanism | Explanation | Finding |
|---|---|---|
| Averageness read as realness | AI faces are average and unmemorable, and people read proportions, familiarity and memorability as signs of a human | MR-0001.F2 |
| Training-data bias | StyleGAN2 data is ≈69% White, so White AI faces are the most "average" and the most hyperreal | MR-0001.F3 |
| Truth default | People default to "authentic" ("seeing is believing") | MR-0001.F6 |
| No mismatch, no alarm | The uncanny response needs distortion or mismatch, which consistent AI faces lack | MR-0002.F3, F9 |
| Trust halo | Synthetic faces are rated more trustworthy | MR-0001.F4 |
| Inaccessible implicit signal | The brain registers a difference that does not reach conscious judgment | MR-0002.F10 |

## The uncanny valley in brief {#uncanny-valley}

- **Size:** meta-analytic g = 1.01 across 72 studies. Face distortion gives the largest effect (g = 1.46) (MR-0002.F1, F3).
- **Cause:** perceptual mismatch and distortion. Category ambiguity is contradicted (MR-0002.F3, F4).
- **Not universal:** the "any near-human likeness" version and Mori's movement claim lack consistent support (MR-0002.F2).
- **Brain:** vmPFC tracks the dip in likability, the amygdala predicts rejection of artificial agents, and parietal cortex flags appearance–motion mismatch (MR-0002.F6).
- **Deepfakes:** good GAN faces are past the valley; imperfect or real-time video may still fall into it (MR-0002.F9, F12).

## Evidence-based controls (editorial) {#controls}

These are inferences from the findings above, not tested interventions:

1. **Verify through process, not perception.** Use a callback to a known number, pre-agreed passphrases, and multiple approvers for payments or access changes (MR-0001.F1, F5).
2. **Prefer comparison over single judgment.** Check against known-genuine reference material, since paired judgments are much more accurate (MR-0001.F10, Theme 5).
3. **Measure hit rate and false-alarm rate in awareness training.** A rising "fakes caught" figure can hide a criterion shift (MR-0001.F7).
4. **Teach that "looks real" and "looks trustworthy" are not evidence** (MR-0001.F2, F4).
5. **Pair humans with detectors validated in-domain**, and train staff to override confident-sounding wrong advice (MR-0001.F9, F10).
6. **Use provenance labels carefully.** Labels shift trust in both directions, including against genuine content (MR-0002.F11).

## Open questions {#gaps}

- **Current video generators:** how well do people detect Sora 2, Veo 3 or real-time face-swap in live calls? No peer-reviewed data was found.
- **Training retention:** do training gains survive beyond the same session?
- **Realistic conditions:** how accurate are people when they are not warned and something is at stake?
- **Race and hyperrealism:** does hyperrealism extend to non-White faces with current, less biased generators?
- **Voice clones:** clones of *unfamiliar* voices are near-indistinguishable (MR-0003.F1, F2). How well do people detect clones of a voice they know, over a phone line or in live conversation?

## Metareports {#metareports}

| ID | Title | Question | Findings | Updated |
|---|---|---|---|---|
| [MR-0001](/metareports/human-detection-accuracy.md) | Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training | How accurately can people tell AI-generated faces, voices and videos from real ones, and what changes that? | F1–F15 | 2026-10-19 |
| [MR-0002](/metareports/uncanny-valley-and-synthetic-faces.md) | The Uncanny Valley and Synthetic Media | Is the uncanny valley real, what causes it, and have deepfakes crossed it? | F1–F13 | 2026-10-19 |
| [MR-0003](/metareports/voice-cloning-human-detection.md) | Voice Clones and Synthetic Speech: How Well Do People Detect Them? | Can people tell cloned or synthetic voices from real speakers, and what changes that? | F1–F9 | 2026-10-19 |

## Data and reference {#reference}

- **Citations file (start here to verify):** [/data/citations.json](/data/citations.json): one JSON file with every research source (DOI, PMID, arXiv and OpenAlex IDs, open-access and resolver URLs, exact-title search query), every recorded claim with its source location and verification status, and every finding with its supporting source IDs. Inline citations like [nightingale-farid-2022](/data/citations.json#nightingale-farid-2022) link to `#<id>` entries in it.
- **Study records:** [/data/studies.json](/data/studies.json) and [/data/studies.csv](/data/studies.csv), one page per study under `/sources/`.
- **Catalogue:** [/index.json](/index.json).
- **Updates:** [/changelog.md](/changelog.md), [/feed.json](/feed.json), [/feed.xml](/feed.xml).
- **Guides:** [Agent guide](/guides/agents.md), [Schema and evidence tiers](/guides/schema.md), [Glossary](/guides/glossary.md).
- **Verification levels:**
  - `verified-fulltext`: values checked against the paper.
  - `verified-abstract`: values checked against the official abstract.
  - `verified-doi`: metadata only.
  - `unverified`: pending.

Published by [Breacher.ai](/about/publisher.md), a deepfake and AI social-engineering simulation and training company ([disclosure and services](/about/publisher.md)) · CC BY 4.0