Written from 36 named sources · Oct 16 · first result A research companion to VO Infusion Episode 4, "Why You Are Right to Be Skeptical." Researched with Pig Knuckle and reviewed by Deb Munro and Keith Norton, with sources for every claim. Deb admitted she has been scared of AI too, and I have defended the right to be skeptical. Clients are using ElevenLabs voices right now, and those lost jobs are real [36]. This companion lays out what controlled research actually found, where the limits are, and what you can do this month to protect your work. WPP Media with AMP, Choreograph and MediaScience tested 55 participants in labs in Chicago, New York and Austin, Texas, and found synthetic voices matched human voices on cognitive attention, physiological arousal and purchase intent in audio advertising [10][13]. Less than half of participants correctly identified generic AI voices, at 42 percent for individual sentences and 47 percent for complete ads, and voices believed to be human were rated more relatable and emotional even when that belief was wrong [11][13]. Peer-reviewed detection work shows listeners are not dependable detectors. A March 2025 Scientific Reports study by Barrington, Cooper and Farid found about 60 percent correct identification of AI voices and about 80 percent same-identity judgments for clones, while a 2025 PLOS One study led by Nadine Lavan at Queen Mary University of London found 58 percent of clones judged human against 62 percent for real voices and only 41 percent for generic AI voices [12][7]. A March 2026 study by Chen, Pell and Jiang in Computers in Human Behavior: Artificial Humans, volume 7, article 100261, used Mandarin-speaking university students and Likert humanlikeness ratings, and found cross-source same-speaker judgments around 54 percent when prosody matched and around 36 percent when prosody mismatched [15]. Consumer and business tiers run under different rules. As of September and early October 2026, trackers found consumer chat tiers often train by default with an opt-out, while business, Team, Enterprise and Workspace tiers generally state no training by default, with important exceptions and version differences [16][21][27]. NAVA offers a free Synthetic Voice/AI rider, Version 4.0 updated July 2025, for education only and not legal advice, and CAVA advises Canadian performers that Canadian law does not currently give people a general property right in their voiceprint, so written limits matter [22][2][24]. What the WPP Media and MediaScience research actually found I came into this wanting a clean answer, and the WPP work does not give us one. It gives us something more useful, a split between what people do and what people feel. In labs in Chicago, New York and Austin, 55 participants who used social media, podcasts and radio heard audio clips in a simulated living room setup while researchers measured emotional, neurological and behavioral responses [13]. Synthetic voiceovers matched human-voiced ads on brand engagement and purchase intent, with similar neurological engagement on cognitive attention and physiological arousal [10][11][13]. That result does not prove synthetic voices work everywhere. The public material describes audio advertising, not animation, games, audiobooks, character work or long-form narration that demands sustained emotion [13]. No full peer-reviewed paper or numerical purchase-intent effect size was supplied in the material reviewed here, so we should treat this as useful applied research for ads rather than proof that a generated voice can replace an actor across categories [13]. The perception gap is where the human edge shows up. Respondents consistently scored voices they believed were human as more relatable and emotional, even when the underlying voice was synthetic or their belief did not match its actual origin [11]. Nancy Hall, Chief Client Officer for WPP Media in the US, said it comes down to tone and inflection, and that those qualities help drive brand recall and attachment [11][13]. That fits the barbecue joint with grandpa's recipe that Deb and I talked about on the show. People respond to a sense of human care behind the message, and belief about who is speaking shapes that response [11]. Age mattered in a narrow way. Younger participants aged 18 to 30 were more adept at spotting AI voices than participants aged 45 and older, according to the study summary [11]. The supplied material does not establish whether that came from greater exposure, different listening habits or another factor, so I would not build a casting argument around it [11]. Can listeners reliably tell AI voices from human ones The short answer is sometimes, under some conditions, and not reliably enough to bet your business on. In the WPP work, less than half correctly identified generic AI voices, at 42 percent for sentences and 47 percent for full ads [11]. Those figures concern generic AI voices in advertising, not every model, language, recording condition or performance style, and a listener can hear a stiff read without knowing the cause was synthesis [11]. Other peer-reviewed work points the same direction with different numbers because designs differ. Barrington, Cooper and Farid published March 31, 2025 in Scientific Reports with 604 listeners using a commercially available cloning tool, finding about 60 percent correct for AI voices, 67.4 percent correct for real voices, and about 80 percent same-identity judgments when a clone was compared to its real counterpart [12]. Lavan and colleagues at Queen Mary University of London, published in PLOS One in 2025 and reported via secondary coverage, found 58 percent of voice clones judged human against 62 percent for genuine human voices, while generic AI voices not cloned from a speaker were judged human only 41 percent of the time [12][7]. An earlier University College London study in PLOS One in August 2023 with 529 participants found 73 percent detection of deepfake speech in English and Mandarin, with only slight improvement after training [12]. A 2026 industry review from Whispeak argues human listening is now unreliable for scale and real-time protection, and points to the ASVspoof 2024 challenge and a 2025 Speech Deepfake Arena benchmark [20]. The supplied page does not provide underlying benchmark tables to independently assess each result, and its claim to rank first on one leaderboard and top four at ASVspoof with 4.16 percent equal error rate is vendor-reported [20]. The practical takeaway stays narrow. Do not assume a client, listener or platform will catch an unauthorized clone by ear, and do not assume a detector settles every case either. Keep contracts, source files, delivery records and written permissions, because provenance beats asking everyone to become a forensic analyst [20][23]. Where humans still hold an edge, and why prosody matters The clearest human advantage in this evidence is perceived relatability and emotion tied to belief. When listeners thought a voice was human, they rated it higher on those dimensions even though attention and arousal looked similar [11]. That means a human voice may carry more weight where a client is selling trust, care or intimacy than in a short promotional spot optimized for speed and volume [11]. It does not mean humans always win, because synthetic voices still held attention and drove purchase intent in the ad test [11][13]. The March 2026 prosody study adds texture without overturning that result. Chen, Pell and Jiang recorded 24 native Chinese speakers with acting, speech or music training, trained Huawei Celia-system avatars, then had AI generate 123 novel Mandarin sentences and had the original speakers return about a month later to produce the same sentences with confident or doubtful prosody [15]. Forty-eight listeners rated 11,808 clips on 7-point humanlikeness and confidence scales, and 80 different listeners did same-or-different speaker judgments on a filtered set of 768 clips [15]. Human voices were rated substantially more humanlike than AI voices across prosodies with large effects, even though mean fundamental frequency differences were minimal to small [15]. In identity judgments, cross-source same-speaker accuracy was around 54 percent when prosody matched and dropped to around 36 percent when prosody was inconsistent, while human-human and AI-AI same-speaker pairs stayed near ceiling [15]. The authors frame this as conservative for simple AI, and they warn that newer expressive systems may close the gap and that lab results with students hearing Mandarin should not be stretched into a general claim about all English commercial detection [15]. That caveat matters for my own test on the show, where I compare my clone to my real voice and hear it lag on depth. The research gives a plausible reason a clone can sound close yet still miss, because timing, emphasis and choices across a line carry identity beyond vocal color, but it does not tell us how any one performer's clone sounds [15]. Turning off model training in the tools you actually use Own the file, the raw files and the API, and turn off model training in every tool you use. That line from the episode is right, but the switches differ by plan and by version, and training off is not the same as deleted. As a rule, check the setting before the first sensitive prompt, because opt-outs apply going forward and do not pull back data already used in training runs or trained models [21][27]. Rating a response with thumbs up or down can also create a separate permission, with vendors stating the rated conversation may be used for improvement even if you opted out [21]. The table below reflects vendor pages as read in September and early October 2026. Menus change, so confirm against the current help page before you paste client scripts [16][21][27]. Tool and tier Default as documented Where the control lives and what it covers ChatGPT Free, Go, Plus, Pro Yes, trains by default for services for individuals [21][27] Settings then Data Controls, switch called Improve the model for everyone. It covers the account across web and mobile but keeps chats in history. It does not cover the separate Codex full-environment control, and Temporary Chat is kept up to 30 days for safety rather than purged instantly [21][27]. ChatGPT Business, Enterprise, Edu No training by default [16][21] Workspace or admin level. OpenAI wording is By default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API [21]. Claude Free, Pro, Max User choice at signup, pre-selected position not published [21][27] Name then Settings then Privacy, at claude.ai/settings/data-privacy-controls. Engadget labels it Help improve our AI models while Anthropic wording varies to Help improve Claude and You can help improve Claude. Off returns accounts to a 30-day deletion window described in comparisons, while on extends retention up to 5 years in those comparisons [27][21]. Claude Team, Enterprise No training by default, per Digital Applied comparison [21] No training toggle described. Owners can disable thumbs feedback under Organization settings then Data and Privacy [21]. Gemini on personal Google account Yes, Keep Activity on by default for adults 18 and over [21][27] Gemini site then Settings and help then Activity, plus myactivity.google.com/product/gemini. When on, activity is used for training generative AI models, and a subset may get human review retained up to 3 years even after you delete history [27]. Off stops training use and stops saving to history, but Google still keeps each conversation 72 hours to run the service [27]. Gemini in Google Workspace No training without permission [21] Admin controlled. Wording is Workspace does not use customer data for training models without customer's prior permission or instruction [21]. Microsoft Copilot, older app, signed in Yes for signed-in adult consumers in eligible countries unless opted out [32] Older-app FAQ describes opt-out of conversation activity for model training, with 18-month default storage for conversation activity and separate personalization controls [32]. Excluded are Entra ID work accounts, Microsoft 365 Personal or Family in-app use, signed-out users, under-18s, and users in Brazil, China excluding Hong Kong, Israel, Nigeria, South Korea and Vietnam [32][21]. Microsoft Copilot, new app from August 18, 2026 No foundation-model training as stated on new-app pages [21] New privacy-controls page carries no training toggle in the Digital Applied read, which conflicts with older Engadget how-to that listed Training on conversation activity and Training on voice conversations. Treat this as version-dependent and verify which app you run before you rely on a menu path [21][27][32]. ElevenLabs No verified training switch in supplied material [36] The supplied September 29, 2026 third-party tutorial documents text-to-speech, cloning, dubbing, speech-to-text, sound effects, music and agents, but no primary help page confirming a current training toggle, location or scope [36]. Open account, privacy and data-use settings now, and get written confirmation before you upload a voice if the control is not explicit. Meta AI Yes by default, with regional difference [21][27] Engadget reports US users have no opt-out for training, while UK and EU users can object through a form in Meta's privacy center under GDPR [27]. Since December 2025 Meta AI chats also feed ad targeting across its apps without a decline path for US users in that comparison [27]. Treat Meta AI as unsuitable for confidential scripts or voice material unless current policy shows a clear control. Across tools, training off does not delete history, recordings, files, memories, logs or material already incorporated, and each of those has its own control [16][27]. If the material belongs to a client, ask permission before it goes into any consumer service, even with training off. What NAVA is asking for, and what contract language exists NAVA's position centers on permission before cloning, limits on reuse, and pay that matches the use. Its free Synthetic Voice/AI rider, Version 4.0 updated July 2025 on the page supplied here, is meant as an addendum when the client's agreement does not already address AI or cloning [22][19]. NAVA states plainly the rider is for education only and does not constitute legal advice from NAVA or its officers or members, and you should have your own attorney or agent review terms [22][2]. The core problem is redistribution after the session. A file recorded for one defined use can later be copied, modeled, placed inside another service or used to generate new performances the session never covered [2][22]. Tim Friedlander of soundBOX made that point on the show, that the right fee does not fix an AI job because you cannot stop later unpaid reuse once the audio leaves your hands. The rider approach responds by requiring separate written agreement for a digital double, synthetic voice or learning model and by restricting transfer, sale or sublicensing of audio assets to third-party AI vendors, though clause specifics beyond excerpts were not verifiable in the material supplied [2][22][23]. NAVA's fAIr Voices pledge covers six areas, training, consent, control, compensation, security and sourcing, and asks companies handling audio to bar unauthorized use of sound files for creating or training synthetic models and to give assurances about performer data [23]. Current members updated October 12, 2025 include AudioBee, Backstage, Bodalgo, CastVoices, Ethovox, LA Casting, Leonardo Audio, Voice123, VoiceArchive, Voices.com, VoCaster, VO Planet, Voquent, soundBOX and others listed on the pledge page [23]. Canadian performers face a different legal reality. CAVA says Canadian law does not currently give people a general property right in their voiceprint, and it urges written limits on machine learning and generative AI or synthetic voice use plus limits on transfer or sale to third parties [24]. CAVA tells performers to watch for simulation, synthesization, digital double and machine learning, to take the job but tell ACTRA or CAVA about the language, and for non-union jobs to add its AI rider language or ask an agent to do it [24]. That contract-first advice fits rates, usage, term, exclusivity and storage terms CAVA lists, including start and end dates instead of in perpetuity buyouts where possible [24]. One history point is worth getting straight. The Illinois Biometric Information Privacy Act, or BIPA, passed in 2008 and provides a private right of action, and it lists voiceprint as a biometric identifier that requires written release before collection [29][30]. The legislative history widely ties it to the collapse of Pay By Touch, a biometric payment startup, and concern that consumer fingerprint data could be sold as an asset in bankruptcy [29][30]. That is a US state law with real teeth in Illinois, not a Canadian rule and not a national voice-ownership law. What this means for your voice business Audit your tools this month and screenshot the switches. List every service that touches your voice, scripts, auditions or client data, open each privacy page, record the date, note whether the control covers text, audio, voice conversations, files or only future chats, and confirm whether you are on a consumer or business tier [16][27]. Set ChatGPT Data Controls, Claude Privacy, Gemini Keep Activity and Copilot privacy before the first real prompt, and adopt a no-rating rule for sensitive chats because feedback can override an opt-out [21][27]. Separate paid work from experiments. Use a business, Team, Enterprise or Workspace account where its terms state no training by default, and do not assume a paid consumer plan carries the same promise [21]. Keep client-owned scripts out of Meta AI and out of any tool where you cannot find a documented control, and get client permission in writing before any consumer-service upload [27][24]. Put AI limits in writing before you record. State whether the client may clone, train, generate new performances, transfer files, sublicense results or reuse after the campaign, and name storage, access and term dates [24]. Use NAVA Version 4.0 as a starting point for the conversation, then have an agent, union rep or lawyer review any change, since NAVA says the rider is educational and not legal advice [22][2]. Keep provenance you can show without arguing about ears. Save the signed contract, session date, delivered and raw files, invoices, approvals and correspondence about intended use [11][20]. Human detection missed a large share in the studies above, and automated detectors have their own error rates and telephony limits, so a clean paper trail beats a debate over whether a clip sounds synthetic [12][20]. Test the business case in the work you actually chase. Compare human and synthetic reads where appropriate and measure completion, response or client preference rather than slogans [11][13]. The ad research shows synthetic can match on engagement and purchase intent while perceived humanness still lifts relatability and emotion, which points to a hybrid worth pricing fairly, human core for connection with consented replicants for tags and versioning where speed matters [11][13][14]. Go further WPP Media research summary. Best single read for engagement, purchase-intent, detection and belief effects in audio ads, including method with 55 lab participants [13]. Phillip Lomax and AMP participant summary. Clearest supplied figures for identification at 42 percent for sentences and 47 percent for full ads, plus the Nancy Hall quote on tone and inflection [11]. Chen, Pell and Jiang March 2026 prosody study. Shows why identity involves timing and expressive choices, with matched versus mismatched prosody results and Likert humanlikeness framing [15]. Barrington, Cooper and Farid March 2025 and Lavan 2025 via Flaunt summary. Useful pair for 60 percent and 80 percent clone-identity figures alongside 58 percent versus 62 percent and the 41 percent generic-AI detail [12]. NAVA rider Version 4.0 July 2025. Download the current rider and instructions, then get professional review for your situation [22][2]. CAVA AI guidance. Canadian-focused checklist for transfer limits, watch terms like digital double, and when to contact ACTRA or CAVA [24]. Digital Applied October 5, 2026 settings comparison and Engadget how-to. Use as a checklist for defaults, paths, retention and feedback overrides, then recheck vendor pages because versions change [21][27]. About VO Infusion VO Infusion is a podcast about artificial intelligence and the voice acting industry, co-hosted by voice actor and coach Deb Munro and voice actor Keith Norton. Episodes cover voice cloning and licensing, legal protections for performers, prompt engineering for voice actors, AI tools for the business side of voiceover and why human performance still matters. Listen wherever you get podcasts or on YouTube at @VOInfusion, and find more at voinfusion.com. About Pig Knuckle Pig Knuckle (pigknuckle.ai) is a research orchestration platform from Alchemy Agentic, founded by Paul Langtry and Keith Norton. It breaks a research question into steps, sends each step to the AI model best suited to it, and builds its citation list from the documents it actually read, so every claim can be traced to a source. Sources [2] AI Rider - NAVA — https://navavoices.org/ai-rider [7] Speech Technology - comprehensive, independent coverage of information impacting speech technologies — https://www.speechtechmag.com/Search/Default.aspx?Query=text-to-speech&rowcount=0 [10] How AI voices impact audio ads: study reveals surprising results WPP Media posted on the topic LinkedIn — https://www.linkedin.com/posts/wpp-media_wpp-media-ai-voices-in-audio-ads-activity-7388530909278334977-HJf_ [11] AI Voices in Ads: Consumer Trust and Engagement Phillip Lomax posted on the topic LinkedIn — https://www.linkedin.com/posts/philliplomax_ai-voices-in-audio-ads-unpacking-consumer-activity-7388719639318507520-dC8K [12] Can People Tell AI Voices From Human Ones? (2026 Data) — https://flauntaudio.com/can-people-tell-ai-voices-from-human [13] AI Voices in Audio Ads: Unpacking Consumer Trust and ... — https://www.wppmedia.com/news/ai-voices-audio-ads [14] Synthetic Voices: When and Why to Use AI Voice Technology — https://www.siriusxmmedia.com/insights/ai-vs-human-voices-when-and-why-to-use-synthetic-voices [15] Prosodic cues strengthen human-AI voice boundaries: Listeners do not easily perceive human speakers and AI clones as the same person — https://www.sciencedirect.com/science/article/pii/S2949882126000125 [16] AI Privacy Policies Compared: ChatGPT, Claude, Gemini, Microsoft Copilot, Codex, Abridge, Fathom, Zoom, Dragon + 48 More (2026) Voibe Resources — https://www.getvoibe.com/resources/ai-privacy-tracker [19] Updated AI Rider with instructions now available. Notes on t LinkedIn — https://www.linkedin.com/posts/navavoices_ai-voice-contracts-activity-7117884355565244416-AO5V [20] What Solutions Exist to Detect AI-Generated Voices? — https://whispeak.io/detect-ai-generated-voices [21] Do AI Chat Apps Train on Your Data? Settings for 9 Apps — https://www.digitalapplied.com/blog/do-ai-chat-apps-train-on-your-data-settings [22] SYNTHETIC VOICE/ARTIFICIAL INTELLIGENCE RIDER - NAVA — https://navavoices.org/2023/01/23/artificial-intelligence-rider [23] fAIr Voices - NAVA — https://navavoices.org/fair-voices [24] AI Issues in Voice Over CAVA — https://www.cavavoices.org/ai-issues-in-voice-over [27] How To Stop AI Companies From Training On Your Data — https://www.engadget.com/2268598/how-to-stop-ai-companies-training-your-data [29] The Illinois Biometric Privacy Act: History, Developments, And ... — https://digitalcommons.mainelaw.maine.edu/cgi/viewcontent.cgi?article=1007&context=sjipl [30] Biometric Information Privacy Act — https