Written from 15 named sources · Oct 16 · first result A research companion to VO Infusion Episode 1, "What Is This AI All About?" Researched with Pig Knuckle and reviewed by Deb Munro and Keith Norton, with sources for every claim. Summary Stanford founded the Institute for Human-Centered Artificial Intelligence, called Stanford HAI, in 2019 to guide and build the future of AI [3]. Its stated mission is to advance AI research, education, policy and practice to improve the human condition [1][2][5]. Human-centered AI treats AI as an institutional choice as well as a technical build. Stanford HAI connects computer science with business, economics, law, philosophy and more, with the goal of becoming an interdisciplinary hub for researchers, builders and policymakers [6]. The 2026 London School of Economics report on AI and labor markets, led by Cosmina Dorobantu, organizes the evidence around three deployment mechanisms: automation, augmentation and commoditization [14]. Commoditization is the most recently formalized and least tested empirically, in part because changing skill requirements are hard to observe [14]. A buyout is typically a one-time fee for a defined or sometimes very broad set of uses, so the scope, term and assignment language controls what a client actually bought. Whether an older buyout reaches synthetic or generative use depends on wording such as all media now known or hereafter devised or for any purpose, so voice actors need to read the clause and talk to a lawyer when AI is involved. A license grants permission to use a voice or recording on stated terms. Limits on use, duration, exclusivity and termination are whatever the contract says, and perpetual or irrevocable language is the version to negotiate against [10]. As of October 2026, SAG-AFTRA theatrical and television terms from 2023 ran November 9, 2023 through June 30, 2026 [9]. A Hollywood Reporter summary reports a new four-year deal ratified June 4, 2026, effective July 1, 2026 through June 30, 2030 [15]. ACTRA bargained for protections built on consent, control and compensation, and members ratified the Independent Production Agreement on January 21, 2025 [13]. As of October 2026, the practical move is to separate the session fee from any synthetic training permission. Name what may be recorded, what may be generated, who may use it, how long permission lasts and how new uses are paid. Stanford HAI and what human-centered means for performers I came to this topic from an odd angle. I spent years as a university fundraiser, and my interest in AI started around Stanford, where technical talk kept drifting away from working people. Deb has worked as a non-union voice actor from her home studio in Canada for more than 35 years, and I finished my demo in about six months, so we both know what it feels like to learn fast while the bills keep coming. Stanford HAI is the Stanford Institute for Human-Centered Artificial Intelligence [3]. Stanford founded the Institute for Human-Centered AI in 2019 to guide and build the future of AI [3]. The annual report says HAI officially launches in March 2019 and describes work since our founding in 2019 [5]. At launch HAI was to be led by John Etchemendy, professor of philosophy and former Stanford provost, and Fei-Fei Li, professor of computer science [1]. James Landay was named Denning Co-Director during the 2023-24 academic year, so any current leadership list should name all three [5]. The mission has stayed consistent. HAI exists to advance AI research, education, policy and practice to improve the human condition [1][2][5]. The institute describes research focused on developing AI inspired by human intelligence, studying and guiding human and societal impact, and designing applications that augment human capabilities [2]. The LinkedIn description adds the same point in plainer terms: AI should be collaborative, augmentative and enhancing to productivity and quality of life, drawing on business, economics, genomics, law, literature, medicine, neuroscience, philosophy and more [6]. Stanford HAI organizes research around three focus areas, sometimes called principles. They are Intelligence, Augment Human Capabilities and Human Impact [5][7]. Intelligence aims to develop technologies inspired by human cognition [5][7]. Augment Human Capabilities designs systems that assist rather than replace people [5][7]. Human Impact studies societal, economic, legal and philosophical consequences [5][7]. The 2024 annual report covers September 2023 to August 2024 and shows work across computer science, economics through the Digital Economy Lab, law through RegLab and law school grants, and policy [5]. For voice actors, that third pillar does most of the work. It asks who decides how a voice is used, who gets paid when a model speaks, and what remedy exists when consent was never clear. Legal scholars Jyh-An Lee and Xuan Sun, at the Chinese University of Hong Kong Faculty of Law, examine those questions in a paper forthcoming in the Singapore Journal of Legal Studies in 2026, a copy of which is hosted in a Berkeley Law archive [11]. They compare the right of publicity, personality rights and data protection, and they describe modern cloning as needing only a few minutes of audio to build a usable template [11]. I like that frame because it gets us out of what Deb and I called the prison of two ideas on the episode. One side says AI will save us and the other says it will end us, and both sides want you to pick a team. Human-centered work asks a harder pair of questions: does this system leave human judgment intact, and do the payment terms match the use [2][7]? A tool that helps you deliver alternate language versions or keep continuity through illness is augmenting. A hidden training clause that turns one commercial read into an endless regional campaign is extractive, and the math model is not the problem. The control and payment terms are the problem [11][13]. What automation transitions actually look like On the episode we asked the self-driving car question: what happens to people who drive for a living when driving gets automated. Some mid-2010s forecasts predicted large-scale driver displacement within years, but that shorthand needs care. The LSE review finds the effects of AI in autonomous vehicles largely unmeasured and what happens to displaced workers nearly unanswered [14]. A Goldman summary in the research library estimates call-center employment 39 percent below trend in the United States, which shows why customer support keeps coming up while driver data stays thin [15]. The LSE report is a 2026 edition contribution to the Global Forum for AI and the Social Sciences, led by Cosmina Dorobantu [14]. It is best read as a map of what has been studied, not as a verdict that one number explains every job. The report distinguishes potential evidence, which estimates exposure or structural capacity, from actual evidence, which measures realized employment, hours or wages with an identification strategy [14]. It warns that early diffusion data are weighted toward displacement, which materializes first, while scale and reinstatement effects work with lags through investment, firm entry, new tasks and reallocation [14]. Economic history suggests new technologies can take years or even decades to diffuse and generate absorbing tasks, citing Autor 2015 [14]. The report organizes the evidence around three mechanisms [14]. Automation substitutes for labor on a specific task, reducing labor needed for a given output [14]. Augmentation complements labor, raising productivity on tasks the worker performs [14]. Commoditization simplifies tasks without automating them, reducing the skill needed for acceptable performance, which expands who can do the job and can weaken bargaining position and compress wages [14]. The same technology can automate one worker's tasks, augment another's and commoditize a third's, depending on deployment and institutional setting [14]. The report says commoditization is the least studied by a wide margin, with no publication coded to it as the primary mechanism, in part because evolving skill requirements are hard to observe [14]. Translators, call-center staff and drivers help show why task language matters, even when the supplied sources do not provide a full case study for each occupation. A translator may use machine output as a draft while retaining responsibility for meaning and tone. A customer support agent may spend less time on routine answers and more time on unusual cases, and the LSE review includes a customer support productivity randomized trial by Brynjolfsson, Li and Raymond [14]. A driver may spend less time steering while still handling safety, service and logistics work. Those are possible task changes to test against local data, not findings the LSE report establishes for those jobs. For voiceover, the general mechanism is more useful than any borrowed statistic. The LSE literature does not measure voiceover session lengths, pickups or text-corpus buyouts, so I will not cite it for those details [14]. What it does offer is a way to name pressure: automation of a specific read, augmentation of editing and cleanup, and commoditization through lower skill thresholds and weaker bargaining power [14]. Rate compression and loss of repeat work can arrive before any formal replacement, and that pattern fits commoditization even when headcount looks stable. Workers benefit when they retain control over use, gain access to training and participate in workflow design, but the supplied sources do not provide a measured success rate for those interventions across occupations. Buyouts, usage terms, licensing and residuals in plain English I have told colleagues since 2018 that a buyout probably is not enough and that voice actors need to license their voices. That wording matters, because probably is not the same as always. The point is narrower and stronger: do not let a single session fee silently carry synthetic rights. A buyout is a commercial arrangement, not a legal category. It is usually a flat fee for recording plus a stated scope of use covering media, term and territory, with no residuals. Whether claims are extinguished depends on contract wording, so read scope, term and assignment language instead of relying on the word buyout alone. Many older buyouts did not mention AI, and whether they cover synthetic use depends on language such as all media now known or hereafter devised. When that phrase appears, talk to a lawyer before you assume it excludes training or generation. Usage terms commonly set medium, territory, duration, exclusivity and category. A 13-week regional broadcast run is different from a national paid social campaign, which is different again from internal training or a video game character. The AI risk is a gap: the contract names the broadcast window but never says whether raw stems may be ingested to train a model or generate alternate copy. A license is permission to use something on agreed conditions. A voice license can cover the original recording, a digital replica or both, and it may be exclusive or nonexclusive. Scope, duration and revocability are whatever the contract says. Perpetual, irrevocable licenses exist, and practitioner guidance advises negotiating against them and defining when rights terminate [10]. Voices as such are not copyrightable because they are not fixed in a tangible form, though recordings are, while personality and publicity rights can protect identity [8][11]. Residuals under SAG-AFTRA agreements are formula-based payments for reuse, such as time-and-salary formulas, percentages of distributor gross for advertising-supported streaming, and foreign subscriber factors [9]. The streaming success bonus is a separate payment triggered when a program reaches 20 percent of domestic subscribers in 90 days, not a residual [9]. Residuals arise under collective bargaining agreements, not under typical non-union buyouts. For non-union work, say plainly that SAG-AFTRA residuals do not apply unless the contract provides them. Contract model What it usually means Voice example AI risk to close Buyout One-time fee for a stated scope, often with no residuals Corporate narration or e-learning for a defined portal use Broad transfer language can be argued to reach training and generation, so define AI uses separately Usage terms Limits on media, territory, duration, exclusivity and category Regional radio for 13 weeks, or paid social for one campaign Add an explicit ban on model ingestion if the job is human-read only [11][13] Voice license Permission on stated terms for a recording, replica or both Branded voice or game engine voice for one year in defined media Limit sublicensing, assignment and term, and state deletion or reversion at end [10][11] Residuals and reuse payments Contract formulas for repeat distribution under union agreements Television, theatrical and covered streaming reuse under SAG-AFTRA formulas [9] Tie synthetic output to measurable units such as API calls, minutes or revenue share, because per-spot logic breaks when lines are infinite [9][13] A static file plays back what you said. A trained model says what you never said, in contexts you never approved [11]. That is why I keep saying price the asset, not the minutes. Charge for the session, then separately scope any modeling, term and output payment. The chart compresses the business choice the table describes in words. A flat buyout ends with a file, hidden rights end with an unpaid model, and an explicit license ends with a renewal decision. Use the chart in a client email when prose takes too long. What performer organizations recommend as of October 2026 NAVA, SAG-AFTRA and ACTRA do not share one contract system. Check the organization's own current guidance, contract department or representative before relying on a summary, and Canadian performers should get Canadian advice for Canadian engagements. NAVA, the National Association of Voice Actors, has published a synthetic voice and AI contract rider template to my knowledge. No supplied source confirms its current text or date here, so cite NAVA's own site as of October 2026 and treat it as a model clause set, not law or a union contract. Model language helps you start the conversation, but it does not bind a client until both sides sign it. SAG-AFTRA's 2023 tentative agreement followed a 118-day strike, with agreement reached November 8 [9]. The successor agreements covered theatrical motion pictures and scripted dramatic television and new media, effective November 9, 2023 through June 30, 2026 [9]. The summary defines Employment-Based Digital Replica and Independently Created Digital Replica [9]. Consent must be clear and conspicuous with a reasonably specific description of intended use [9]. Consent is not needed for use in the same picture if the performance remains substantially as scripted, and there are broad exceptions for post-production alterations, dubbing and doubles [9]. Employment-based replica use is generally paid at the pro rata daily rate or minimum, whichever is higher, for days the producer estimates the performer would have worked, with exceptions including Schedule F and work already covered by compensation [9]. Use outside the picture requires separate bargaining at no less than day-performer minimum [9]. Those 2023 terms are now superseded for current planning. A Hollywood Reporter summary reports SAG-AFTRA members ratified a four-year deal with the AMPTP on June 4, 2026, with 91.42 percent approval, effective July 1, 2026 through June 30, 2030 [15]. The summary says synthetic AI-generated performers may only be used when they provide significant additional value, studios need an articulable business reason to scan for a replica, minimum rates and residuals apply to independently created replicas, and replicas cannot circumvent strike participation [15]. As of October 2026, do not present the 2023 summary as current. Read the ratified text and ask the union when a voiceover, commercial or non-union job falls outside theatrical and television coverage. In Canada, ACTRA sought renewal of the Independent Production Agreement with producer associations as the prior agreement approached expiry at the end of December 2024 [13]. The FIA report describes ACTRA negotiating protections based on informed consent, control and compensation, with generative AI used to enhance rather than replace human performance [13]. Those were bargaining positions in December 2024, and the update says members voted to ratify the new agreement on January 21, 2025 [13]. Residuals modernization was listed as a separate priority driven by streaming changes, including shorter series and exclusivity effects, not as a subset of AI rights [13]. Jurisdiction changes the backstop when contract language is vague. California has Civil Code Section 3344, which names voice, plus a common-law right [10]. Tennessee protection comes from its Personal Rights Protection Act, expanded by the 2024 ELVIS Act covering voice including simulations [8][11]. Both apply to commercial use and are limited by the First Amendment [10]. Do not call Tennessee's protection common law. Canada has no single federal AI-use statute for voice cloning, so recourse remains uncertain without clear contract terms [8]. The essay often cited on this point is a student essay by Ambreena Ladhani, winner of the CBA Privacy and Access Law Section's 2025 essay contest, published November 3, 2025 [8]. It does not discuss passing off. It describes doctrinal hurdles: intrusion upon seclusion exists only in Ontario and needs a private affair, public disclosure needs truthful information, false light is novel and recognized only in Ontario, and misappropriation of personality needs commercial gain or endorsement-type use plus proof of damage at common law [8]. For commercial voice clones such as advertisements, the misappropriation tort is the most relevant [8]. Do not read that as no protection in Canada. Manitoba, Saskatchewan and Newfoundland and Labrador statutes expressly protect voice [8]. British Columbia covers name or portrait for advertising or sale and does not mention voice [8]. Ontario recognizes the common-law tort of appropriation of personality in Krouse [8]. Quebec's Civil Code protects voice, and Quebec's Charter also protects privacy [8][10]. Because the map is fragmented, non-union and union actors in Canada need explicit reservation of synthetic rights in the contract and should talk to ACTRA, an agent or a Canadian entertainment lawyer before recording [8][13]. What this means for your voice business Audit five recent agreements this month. Pull contracts, release forms and platform terms, then highlight perpetuity, assignment, derivative works, training, digital replica and sublicensing. Make a one-page list of clients holding your audio and any clause you cannot explain without help. Add a short AI rider to every quote and invoice. State that recordings may not be used for training, modeling, cloning or synthetic generation without a separate signed agreement. Check NAVA's own current rider as of October 2026, because organizational language changes with bargaining and litigation. Write a one-page synthetic policy before you need it. Name your minimum modeling fee, your term limit such as one year, your renewal or per-use rate, and categories you will not voice. Include script approval and a deletion process at term end, then send it with auditions that mention AI. Separate money for files and models. Price the session, then price creation, license term and output separately through renewals, per-minute fees or API-based royalties. If a client wants infinite lines, the payment structure should reflect continuing commercial value [11][13]. Verify jurisdiction in writing. If you work in Canada, do not paste American publicity language into a Canadian deal without review [8][10]. Ask ACTRA if you are a member, or consult a Canadian entertainment lawyer about personality rights and provincial limits [8][13]. If a job is union-covered, contact SAG-AFTRA or ACTRA about scale and reuse minimums before you sign [9][13]. Four questions to ask before you sign anything touching AI What specific technical actions are authorized with my audio. Ask the client to distinguish editing a file from training a model, building a replica or generating new speech [11][13]. If the job is human-read only, require the sentence that files may not be used for AI training, modeling or synthetic replication. Who owns any resulting model and can it be transferred. List the client, affiliates, vendors and platforms holding your data. Forbid sale, sublicensing, assignment or contribution to shared or open models without separate written consent [11][13]. Ask how affiliates are bound, not just the signing entity. What are the boundaries for media, territory, term and approval. Reject open-ended perpetuity for synthetic rights. Get channel, geography, expiration, exclusivity and a process to inspect scripts and demand deletion or return of data at term end [9][13]. Ask what happens if the model produces an unauthorized or misleading performance. What is the exact pay structure beyond the session fee. Name the creation fee, license retainer, per-minute or per-call rate, and any residuals or success payments [9][13]. If union coverage may apply, confirm minimums with the union. If the job is non-union, have an agent or entertainment lawyer review long-term vocal asset rights before you sign. Go further Stanford HAI overview: concise mission and 2019 founding history [3]. Stanford HAI 2024 Annual Report: shows cross-disciplinary work across research, education, policy and industry from September 2023 to August 2024 [5]. LSE State of AI and Society on labor markets: separates potential from actual evidence and organizes effects into automation, augmentation and commoditization [14]. FIA and ACTRA bargaining update: explains consent, control, compensation and separate residuals modernization, with January 2025 ratification [13]. CBA essay by Ambreena Ladhani: starting point on why Canadian privacy torts struggle with cloned artist voices [8]. Lee and Sun on vocal identity: comparative treatment of publicity, personality and data protection frameworks [11]. 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 [1] Stanford University launches Institute for Human-Centered Artificial Intelligence — https://idahobusinessreview.com/2019/03/20/stanford-university-launches-the-institute-for-human-centered-artificial-intelligence [2] Stanford Institute for Human-Centered Artificial Intelligence - Summary from LegiStorm — https://www.legistorm.com/organization/summary/200535/Stanford_Institute_for_Human_Centered_Artificial_Intelligence.html [3] Stanford Institute for Human-Centered Artificial Intelligence (HAI) — https://online.stanford.edu/schools-centers/stanford-institute-human-centered-artificial-intelligence-hai [5] Stanford Institute for Human-Centered Artificial Intelligence — https://hai-production.s3.amazonaws.com/files/2025-02/2024-hai-ann