Changelog

Dated record of changes to the Deepfake Research Index, newest first. Finding IDs (MR-xxxx.Fn) and source IDs are stable; this page records when they are added, revised or re-verified.

2026-10-19 · v0.4.0 · Citations file, MR-0003 voice clones, expert readers and new uncanny-valley studies

• Site:
• New /data/citations.json: one file with all 54 source records (identifiers, resolver URLs, exact-title search query, every claim with source_location, verification) and all 37 findings with their supporting source IDs, for independent verification. data/studies.json is unchanged for backward compatibility.
• Inline [source-id] citations in every metareport and on the homepage now link to /data/citations.json#<id>; References entries show the citations.json link, the source record, an APA-style citation and the DOI link; source pages link to their citations.json entry.
• citations.json is listed in llms.txt ("Verify the research", first), the llms-full.txt header, index.json, the agent guide, schema.md (new "Citations file" section and locator fields pmid, pmcid, arxiv_id, openalex_id, oa_url, source_location), the homepage, the sitemap, _headers and the homepage JSON-LD.
• New build errors: citations.json must parse and match studies.yaml; every cited source and every finding must be present, with existing source IDs; every source needs a locator or search query; no entry may mention the publisher; public/index.html must exist.
• llms-full.txt now carries compact frontmatter for source records (full data remains in the record tables and citations.json), keeping it under 400k characters.
• Locators recorded from identifiers already read: PMCID for nightingale-farid-2022, PMID for sulimani-2026-recognition-ai-faces, arXiv IDs for frank-2024-representative-study, cooke-2024-coin-toss, lu-2023-fake2m and groh-2022-crowds-machines; source_location: abstract set on 71 results of abstract-verified records.
• New metareport MR-0003: Voice Clones and Synthetic Speech (F1–F9, 8 sources), split out of MR-0001 so MR-0001 stays under its size limit. MR-0001.F12 and F15 remain and cross-reference it.
• Added 6 source records:
• lee-2026-ai-chest-radiographs: tier B, verified-abstract.
• chen-2026-humans-ai-deepfake-teaming: tier B (HFES proceedings), verified-abstract; abstract reports no figures.
• naseem-2026-in-the-wild-detection-review: tier C (narrative review), verified-abstract.
• mamsheli-2026-eeg-ai-faces: tier B (ECCE proceedings, exploratory), verified-abstract.
• hsieh-2026-robot-gaze-uncanny: tier B, verified-abstract.
• jang-2026-uncanny-conversation: tier B, verified-abstract.
• MR-0001 v0.4.0: revised in place F9 (humans and AI use different cues) and F10 (expert readers vs LLMs on AI chest X-rays; detector collapse on a held-out generator; in-the-wild transfer). Voice detail moved to MR-0003; TL;DR, Key numbers, evidence table, Themes 1–7, moderators, gaps, implications and FAQ condensed. No finding text or values were dropped; the second Groh et al. chart was folded into Theme 6 text with its values.
• MR-0002 v0.4.0: revised in place F1 (valley-shaped likability across robot gaze timing), F5 (uncanny valley of mind in conversational self-disclosure) and F10 (early ERP differences between real and AI faces).
• Homepage: citations.json added first under "Data and reference" and in the agent entry points; key facts 9–10, open questions and the Metareports table updated for MR-0003.
• Searches run: general web search was unavailable (blocked for this account). Discovery used OpenAlex listings (search "deepfake detection human participants" from 2026-09-15; title searches for deepfake, voice clone / synthetic speech / AI voice, and AI-generated / synthetic faces / uncanny from 2026-08-15) and OpenAlex single-work lookups; OpenAlex then rate-limited further requests, and the Springer publisher page could not be fetched. No qualifying peer-reviewed studies were found for real-time video-call deepfakes, provenance/C2PA effectiveness or deepfake security-awareness training.
• Not added: Chen, Pell & Jiang 2026; Chow et al. 2025; McGuire et al. 2026 (still unread; lookup rate-limited); preprints on voice-clone familiarity and emotion prosody (arXiv 2609.38794), intonation perception in real vs synthetic speech (arXiv 2610.08839), a registered report on the valence-dominance model and deepfake detection (PsyArXiv), "Hyper-Expressions: AI-Generated Faces Emulate and Amplify Stereotypical Human Emotions" and "EEG and Eye-Tracking Evidence That AI Disclosure Shapes Face Evaluation" (preprints, not assessed); "Known-Voice Cloning and the Limits of Personality Transfer" (preprint; personality transfer, not detection); "When Agents Have Faces: Deepfake Avatars…" (ACM workshop; abstract not read).

2026-10-18 · v0.3.0 · Aggregation limits, plausibility, cross-linguistic speech and AI-text training

• Added 6 source records (all tier B, peer-reviewed):
• kramer-2026-aggregation-human-ai: verified-abstract.
• plohl-2026-political-deepfakes-plausibility: verified-abstract.
• san-segundo-2026-audio-language-style: verified-abstract.
• manhede-2026-online-markets-training: verified-fulltext.
• sulimani-2026-recognition-ai-faces: verified-abstract.
• yeung-2026-isthisai-strategies: verified-abstract (archival conference, AIES).
• MR-0001 v0.3.0:
• New: MR-0001.F14 (content plausibility lowers detection; sharing driven by liking and attitudes) and MR-0001.F15 (listeners rely on prosodic/extralinguistic cues; folk heuristics shift over time).
• Revised in place: F5 (no metacognitive sensitivity for video; confidence-weighted pairs worse), F7 (labelled training improves AI-text discernment), F9 (wrong AI advice offsets gains, replicated), F11 (media literacy; AI-platform engagement predicts worse discernment).
• Condensed TL;DR, Key numbers, Themes 2 and 4–7, moderators, caveats and FAQ to keep the page under 40k characters; removed the duplicate synthetic-image chart. No finding text or values were dropped.
• MR-0002 v0.3.0: F9 revised (AI faces remembered better than real faces, covertly).
• Homepage: interventions table (AI advice, labelled training) and metareport table updated.
• Searches run: general web search was unavailable (blocked for this account); several OpenAlex and Europe PMC requests were rate-limited. Discovery used OpenAlex keyword listings ("deepfake detection human participants", "voice clone listeners perception", "deepfake participants experiment detection accuracy", "AI label provenance trust experiment"; publication date from 2026-06 to 2026-10) and a PubMed listing (deepfake / AI-generated faces / voice clone, 2026-09-15 onward). No qualifying peer-reviewed studies were found this week for real-time video-call deepfakes or provenance/C2PA effectiveness.
• Not added: Chen, Pell & Jiang 2026; Chow et al. 2025; McGuire et al. 2026 (still unread); "Complementary Error Profiles of Human Readers and LLMs in Detecting AI-Generated Chest Radiographs" (J Imaging Inform Med, 2026; publisher page could not be read); arXiv preprints on known-voice cloning and deepfake detection across hearing abilities (not yet assessed).

2026-10-12 · v0.2.0 · Voice clones, training and individual differences

• Added 7 source records (all tier B, peer-reviewed):
• ivory-2026-faces-voices-domain: verified-fulltext.
• lavan-2025-voice-clones-hyperrealism: verified-fulltext.
• pehlivanoglu-2026-machines-humans: verified-fulltext (Study 1; Study 2 from abstract).
• dunn-2026-super-recognizers-average: verified-fulltext (repository version).
• rehman-2025-training-synthetic-faces: verified-fulltext.
• dawel-2026-training-ai-faces: verified-abstract.
• kramer-2024-crowds-synthetic-faces: verified-abstract.
• Verification upgrades:
• barrington-2025-voice-clones: unverified → verified-fulltext. Added DOI, N = 634, and exact figures (same-speaker median 83.3%; clones identified 60.8%).
• gray-2025-super-recognizers: unverified → verified-fulltext. Corrected authors, DOI, N = 664 and generator (StyleGAN3, not "StyleGAN-family"). The earlier approximate percentages (31%/41% → 51%/64%) were confirmed.
• MR-0001 v0.2.0:
• New: MR-0001.F12 (voice clones near-indistinguishable from the real speaker) and MR-0001.F13 (detection skill is not domain-general; AI faces are hyper-average).
• Revised in place: F5 (confidence uninformative for voice clones), F6 (truth bias extends to voice clones; confidence raised to High), F7 (artefact training criterion shift; impression-based training), F10 (detectors fail on video; crowds help for faces), F11 (super-recogniser evidence).
• MR-0002 v0.2.0: F9 revised (face-space centrality evidence). New MR-0002.F13 (voice clones reach realism parity, not hyperrealism).
• Homepage: key fact 9, accuracy table, myths, interventions and open questions updated for voice clones and training.
• Searches run: web search tools were unavailable this run (blocked); Europe PMC and Crossref were rate-limited. Discovery used a PubMed query ("deepfake human detection participants", 2025–2026), OpenAlex record lookups, and the reference lists of Ivory et al. 2026 and Dunn et al. 2026.
• Not added (could not be read): Chen, Pell & Jiang 2026 (voice clones and identity, Computers in Human Behavior: Artificial Humans); Chow et al. 2025 (JEP: General; abstract incomplete); McGuire et al. 2026 (AI faces becoming more trustworthy; venue unknown).

2026-10-07 · v0.1.0 · Initial seed release

• Added metareport MR-0001: Human Detection of Deepfakes. 11 findings (F1–F11) from 17 sources.
• Added metareport MR-0002: The Uncanny Valley and Synthetic Media. 12 findings (F1–F12) from 22 sources.
• Added 35 source records, an evidence-briefing homepage (key facts, myths vs evidence, interventions table), data exports (/data/studies.json, /data/studies.csv), llms.txt, llms-full.txt, index.json, and JSON/Atom feeds.
• Added publisher disclosure page, describing Breacher.ai's services. It is kept separate from the research pages.
• Verification at release:
• 9 records checked against full text.
• 12 checked against the official abstract.
• 2 with metadata/DOI only.
• 12 pending.
• Corrections made during verification, before release:
• Mai et al. 2023: 73% is the hit rate on fakes. Overall single-clip accuracy is 70.35%.
• Diel et al. 2022: overall effect g = 1.01 and face distortion g = 1.46. This replaced a mischaracterization of the overall trend.
• Cheetham et al. 2014: ambiguous faces were easier to discriminate.
• Eiserbeck et al. 2023: the LPP effect appears for presumed-fake smiles. Early components were lost.
• Schindler et al. 2017: EPN showed no realism effect.
• Rosenthal-von der Pütten et al. 2019: DMPFC encodes choices about artificial agents.
• Groh et al. 2022: N = 15,016.
• Cooke et al.: now published in Communications of the ACM (2025).