Written from 15 named sources · Oct 16 · first result A research companion to VO Infusion Episode 9, "The AI Dramaturg." Researched with Pig Knuckle and reviewed by Deb Munro and Keith Norton, with sources for every claim. Summary of findings A dramaturg centers on literary and research work on a script, including context, language, structure and subtext, and works before and alongside rehearsal. That research-first role is the useful model for using AI in script prep, while stage management and final performance choices stay human. I have said on the show that building a voice career can run $30,000 to $50,000, and that is my own multi-year planning estimate rather than a sourced total. The sourced evidence here covers only marketplace terms, including Fiverr at 20 percent, Upwork starting at 10 percent, and ACX per-finished-hour rates around $100 to $400 or more [15]. As of October 2026, the NO FAKES Act, identified as S. 4591, is proposed federal legislation on the Senate Legislative Calendar, not enacted law with a new civil claim available today [3], [10]. Existing state and federal protections described in current analysis continue to apply while that bill remains pending [10]. Illinois voiceprint litigation under the Biometric Information Privacy Act turns on written notice and signed consent before collection of voiceprints and other covered identifiers [14]. A commercial tracker preview generated October 9, 2026 lists 27 filings dated 2023 to 2026 with status unverified, and describes 156 additional companies as predicted targets that have not yet been sued [7]. ChatGPT, Claude and Gemini each offer a consumer training control for future chats, but opting out does not pull data back out of models already trained [6], [2]. Retention clocks keep running after you flip the switch, including about 30 days for ChatGPT Temporary Chats and Claude opt-out accounts, and longer windows where training is allowed [6], [2], [12]. As of October 2026, the supplied material describes no general United States consumer opt-out for Meta AI, so material typed into Meta AI should be treated as potentially usable for training until Meta says otherwise in its own policy. Microsoft Copilot evidence is mixed across the supplied guides, so the in-app Privacy setting should be checked directly [5], [12]. The practical rule that holds across tools is to keep confidential contracts, unreleased scripts, client data and identifying voice material out of consumer chatbots unless you have permission and have checked current terms and retention statements [2], [6], [12]. What a dramaturg does, and why the comparison helps voice actors In professional theater practice, a dramaturg focuses on the script itself through research, context, language and structure, and that work continues before and alongside rehearsal rather than ending before staging begins. The supplied research set for this piece includes no primary theater source defining the term, so this description reflects standard professional usage rather than a cited theater manual. A stage manager handles a different job centered on schedules, calls and show operation, while a dramaturg may also contribute program notes, new-play feedback and audience-facing context tied to the text. That distinction is why the AI dramaturg idea is useful in the booth. You can ask a language model to lay out the circumstances of a scene, summarize the likely audience, flag unfamiliar names or references, compare two readings of a line, or draft questions you want to answer from the client brief. I spent years as a former university fundraiser learning to do homework before a meeting, and script prep rewards the same habit of arriving with context and questions. The tool can gather and organize that homework, while you remain responsible for choices about meaning, tone and delivery. The limit is straightforward and worth keeping in front of you. Language models can sound confident while getting pronunciation, history, dialect or technical vocabulary wrong, and the material supplied for this piece provides no accuracy rate that would support trusting those answers blind. Treat every AI explanation as a lead to verify against an authoritative dictionary, a specialist source, a coach or the client brief before you record. Which booth and business jobs AI handles well, and which stay human The strongest fit is a first pass on non-confidential copy. Give the system a short practice script and ask for likely emphasis shifts, definitions of unfamiliar terms, questions about the speaker objective, and spots where the copy could be read two ways. Then check that output against the brief and your own judgment, and save the verified version with the project files. Routine business writing is the next practical use. AI can turn rough bullets into a polite invoice reminder, a short scheduling note, a deliverable list or a file-naming convention for auditions and finals. Those drafts still need a human review for rates, rights, deadlines, usage and payment terms, because a fluent draft is not an agreement and small wording changes can carry real consequences. Software help saves time in the same way. When Deb and I talked about getting the one answer you need about Adobe Audition, that captures the use case of asking for a plain-language explanation, a checklist for a repetitive workflow, or a starting point for a small utility. That kind of help does not replace the ear of George the Tech on a finished file, and AI-generated code should be tested before it touches client audio. Learning editing basics still matters, and de-breathing for long-form sits in that basics category as its own skill to practice. File organization needs a firmer boundary. AI can suggest folder structures, metadata fields and naming rules, and it can help you think through backup habits. Decisions about what to keep, what to delete, who can access client material and how long to retain it should stay with you, and a public chatbot should not be used to work around a confidentiality promise. A separate thread on the show involved vibe coding a missing tool with Claude, and that kind of experiment belongs with test material rather than client files. Several decisions should not be delegated because the text reads smoothly. Performance and interpretation, vocal health, client relationships, contract approval, pronunciation disputes and final quality control all affect your reputation and liability in ways a chatbot cannot absorb. Bob Bergen's habit of marking up contracts before signing came up on the show as an example worth copying, and union agreements and carefully negotiated riders remain among the most practical protections on active sessions. If language involves synthetic replicas, perpetual use, training, or derivative rights, have an agent, entertainment lawyer or union representative review it before you sign, since this article provides information only and not legal advice. What it costs to build a voice career when solid totals are scarce My $30,000 to $50,000 figure from the episode is a multi-year capitalization estimate for a durable commercial business, not a ticket price for getting started and not a sourced industry total. The sources supplied for this piece do not verify that total, and they do not provide current vendor quotes for booths, microphones, coaching or demos. The component ranges below are therefore illustrative estimates to help you request real quotes, while only the marketplace figures carry a source citation. Capital bucket What it covers Illustrative entry range Illustrative professional buildout Physical space Treated closet or DIY treatment through engineered booth and isolation Around $1,500 for a basic treated setup $8,000 to $18,000 for a prefabricated or built booth, before install and freight Signal chain Microphone, interface, DAW, headphones and monitoring $600 to $1,000 $2,500 to $4,500 Craft and market entry Coaching, starter reel, website, casting profiles and marketing Around $2,300 for foundational classes plus a starter reel $11,000 to $21,000 or more for multi-genre demos, advanced coaching and conference travel Those illustrative ranges are starting points for calls to coaches, demo producers and equipment suppliers, and prices move with sales, used inventory and local labor. Monmouth University came up on the show as now offering voice acting courses, which I mention here as episode discussion rather than as a verified catalog claim. The Canadian price picture in the supplied material is thin, so actors in Toronto, Vancouver and elsewhere in Canada should budget in local currency, confirm current exchange, shipping, duties and provincial sales tax treatment, and get domestic quotes rather than treating United States examples as final. What the supplied source does support is how platforms take their cut and how audiobook work tends to pay [15]. VOTrainer reports as of 2026 that Fiverr takes a 20 percent commission on every order, while Upwork commission starts at 10 percent [15]. The same 2026 guide says ACX projects commonly use per-finished-hour rates of approximately $100 to $400 or more, royalty share, or a combination, and estimates that a six-hour audiobook may require 18 to 30 or more hours of recording and editing [15]. That guide also reports that royalty-share projects pay nothing upfront and characterizes most indie audiobooks as unlikely to generate significant ongoing revenue, which VOTrainer presents as its own practical assessment rather than audited market data [15]. The same guide places annual casting-site costs at about $500 for Voices.com basic membership and $300 to $600 or more for Voice123 by tier, which helps explain why many coaches advise building skills and reviews on free platforms before paying for larger marketplaces [15]. Direct marketing through outreach, a personal website, LinkedIn and, later, agents remains the path many sustained careers rely on for higher-budget and repeat work [15]. For planning this month, separate one-time purchases from recurring subscriptions, note the date and currency beside every line, and treat platform commissions as a direct reduction of gross before you decide whether a job supports your rate. Where the NO FAKES Act and Illinois voiceprint cases stand as of October 2026 As of October 2026, the NO FAKES Act of 2026 is identified as S. 4591, with a House companion identified as H.R. 8915 [3], [10]. A Congressional Research Service summary published July 2, 2026 and carried by GovTrack says the bill would create a federal intellectual property right against unauthorized digital replicas of voice and visual likeness, with authorization rights, civil liability, First Amendment-related exemptions, a notice-and-takedown process with counter-notification, penalties for false notices, and preemption of state digital-replica laws except causes of action under state statutes or common law in existence as of January 2, 2025 [3]. The procedural status needs care. Attorney and scholar Jennifer Rothman reported October 1, 2026 that a significantly revised Senate version was substituted in late June for the May versions, then moved out of the Judiciary Committee and placed on the Senate Legislative Calendar under Order No. 446 [10]. The same report says it is unclear whether or when the bill will receive floor discussion given the approaching midterm elections [10]. The supplied material shows no enactment, presidential signature or effective date, so performers should treat the bill as pending proposal rather than current federal protection and should check current legislative status before relying on it [3], [10]. That pending status does not leave performers without law today. Rothman's analysis notes that unauthorized digital replicas and voice clones can already implicate a thicket of existing state and federal claims, including publicity, privacy, copyright, trademark, unfair competition, false advertising, false endorsement, defamation, false light, infliction of emotional distress and fraud, plus newer AI-specific rules in some contexts [10]. Union contracts and negotiated riders sit among the most practical day-to-day shields on sessions, alongside those background rights and careful contract review. NAVA and other performer organizations publish guidance and resources on AI, consent and contract concerns, which is distinct from providing individual contract review. Illinois litigation centers on the Biometric Information Privacy Act, or BIPA, codified at 740 ILCS 14 [14]. A 2026 guide to the statute says BIPA protects exactly five biometric identifiers, including voiceprints, alongside retina or iris scans, fingerprints, hand geometry scans and face geometry scans, plus biometric information based on those identifiers and used to identify a person [14]. Before collection, a private entity must give written notice that data is being collected or stored, state in writing the purpose and duration, and obtain a written release, with electronic signatures recognized under the 2024 amendment known as Public Act 103-769 [14]. The same guide notes additional duties around a public retention and destruction schedule, a ban on selling or profiting from biometric data, limits on disclosure, and a duty to protect the data at least as well as other sensitive information [14]. The litigation snapshot most often cited here is a commercial trend report rather than a court opinion. Rain Intelligence generated a public preview October 9, 2026 that lists 27 filings dated from 2023 through 2026, with courts, firms, industries and theories shown in a sortable table [7]. The status of each filing is not verified in the preview, and some entries are labeled Other or involve unusual defendants, so the responsible reading is 27 filings listed rather than 27 active wins or a finding that every recording counts as a voiceprint [7]. The same preview says 156 additional names are hidden, and its own text describes those as predicted defendants, meaning companies that have not yet been sued but are on plaintiffs firms radar, plus pre-filing investigations and AI forecasts available only in the paid product [7]. Anyone facing a real dispute should preserve the contract, consent language and platform records and consult counsel or a union resource promptly. How to turn off model training, and what each switch actually covers Settings move, so confirm the live toggle in your own account before you rely on it. The paths below reflect guides current through mid-2026 and late 2025 rather than a promise that labels will never change [6], [12]. Tool Where to look as of mid-2026 What opting out affects Retention note to record ChatGPT Settings, Data Controls, turn off Improve the model for everyone [6], [9], [12] Future chats excluded from training, prior training not undone [6] Temporary Chats not used for training and deleted within about 30 days, subject to security or legal holds [6] Claude Settings, Privacy or Data Privacy Controls, look for the model-training toggle labeled Help improve Claude in one guide and Help Improve our AI models at claude.ai/settings/data-privacy-controls in another [6], [4] Future chats and coding sessions excluded, prior training not undone [2] About 30 days when training is declined, up to five years for new or resumed chats when training is allowed [2] Gemini Profile, Gemini Apps Activity or Keep Activity, turn off [6], [12] Future activity excluded from training use, prior human-reviewed copies not pulled back [12] Around 72 hours for service operation after opt-out, while conversations selected for human review can be kept up to three years and are not removed by deleting activity [6], [12] Copilot Account, Privacy, look for Training on conversation activity [12] Evidence is mixed, verify in app Check Microsoft account privacy dashboard and product terms for your account type ChatGPT consumer accounts train by default unless the Data Controls toggle is off, and the same default applies to paid Plus and Pro consumer tiers until the user changes it [6], [12]. Turning the toggle off walls future conversations off from the training dataset but does not delete history, and it cannot remove data from models already trained [6]. Temporary Chat is the more contained option for a sensitive question, since those chats do not appear in history, are not used for training, and are deleted from systems within about 30 days unless retention is required for security or legal reasons, a caveat that includes litigation holds noted in the 2026 privacy guide [6]. Claude changed expectations in 2025. Anthropic announced August 28, 2025 that consumer Free, Pro and Max plans, including Claude Code used through those accounts, would be covered by a training choice with a decision date of October 8, 2025 for existing users [2]. The control does not apply to services under Commercial Terms, including Claude for Work through Team and Enterprise plans, the API including via Bedrock and Vertex AI, Claude Gov and Claude for Education [2]. If training is allowed, Anthropic describes retention up to five years for new or resumed chats, coding sessions and submitted feedback to support model development and safety work, while declining leaves the prior 30-day retention period in place [2]. Deleting a conversation excludes it from future training, but training already in progress and models already trained are not retroactively revised [2]. Gemini ties training to account activity rather than a single in-chat switch. Guides direct users to turn off Gemini Apps Activity, sometimes labeled Keep Activity, to stop saving chats to long-term history and using them for generative model training [6], [12]. Google still keeps limited data for around 72 hours for service delivery and abuse checks even with activity off [6], [12]. The detail that matters for performers is the human-review copy described in Google documentation as summarized in a June 2026 guide, which states that conversations selected for human review can be retained up to three years and are not removed when activity is deleted, so turning the setting off stops new sampling without retrieving copies already held for review [12]. Microsoft Copilot requires direct verification because the supplied guides conflict. A 2025 LinkedIn summary states that Microsoft had not published a consumer-facing opt-out and advises reviewing Microsoft privacy settings and limiting sharing [5]. A more current and more specific June 2026 guide gives Account, Privacy, turn off Training on conversation activity, and adds that Microsoft states it does not train on signed-out sessions with consumer data in the European Economic Area excluded [12]. Given that mix, open the Copilot app or web account you actually use, search help for model training or data controls, and record what you find rather than assuming a ChatGPT or Gemini path applies [5], [12]. Meta AI needs a plain boundary. The supplied material describes no general United States consumer training opt-out for Meta AI, which means scripts, auditions or dialogue entered into Meta AI interfaces should be treated as potentially usable for training. That is not the same as proving that every pasted script was ingested, since provider language is often framed as may use interactions, and policies can change. Check Meta's current policy before pasting anything sensitive, and keep client material out of the tool in the meantime. Watching terms over time matters as much as flipping a switch once. Keep a short privacy log with date, product, account type, setting name as shown on screen, stated retention, and a link or saved copy of the policy. Revisit the log when you see a policy notice, a product rename, or a new memory, feedback or personalization feature, since Anthropic's 2025 consumer update shows how defaults and retention can shift with one notice [2]. Enterprise, business, education and API tiers are generally treated as excluded from consumer training by default across providers, but workplace admins can configure them differently, so confirm with documentation or information technology support for work accounts [6], [12]. What this means for your voice business Build a safe-for-AI folder with generic practice scripts, public information and fictional examples, and keep client scripts, contracts, invoices and identifying voice files outside consumer chat tools unless you have written permission. Record that permission with the project files so a future question about consent has a clear answer. Use AI for a first-pass script brief that lists emphasis options, unfamiliar terms, questions about objective and ambiguous lines, then verify pronunciation, names, numbers and technical terms against the client brief and an authoritative source. Save the verified brief beside the session file so pickup sessions start from facts rather than guesses. Run a monthly business checklist that covers invoices, follow-ups, auditions, backups and software maintenance, and let AI draft the checklist language and routine emails while you review recipients, amounts, deadlines and rights statements. Short drafts reviewed carefully beat long threads reconstructed from memory. Turn off the training controls on each consumer service you use, note the date and exact toggle name, and recheck quarterly and after app reinstalls or policy emails. Use Temporary Chat, incognito-style modes or reduced-activity settings for sensitive one-off questions, while remembering those modes shorten retention without creating end-to-end encryption [6], [12]. Treat any request involving cloning, digital replicas, perpetual use, training or derivative rights as a pause point for professional review by an agent, lawyer or union representative. Preserve the offer, the consent language and the platform messages, and consult NAVA resources and current contract guidance for context before you respond. Go further Anthropic, Updates to Consumer Terms and Privacy Policy [2]. The clearest supplied source for Claude training choice and the five-year versus 30-day retention distinction. GovTrack summary of S. 4591, NO FAKES Act of 2026 [3]. A concise legislative summary for understanding the proposed federal digital-replica right and its exemptions. Jennifer Rothman, NO FAKES Revised, Mostly Improved, and Put on Senate Calendar [10]. Status and policy analysis of the revised Senate text, including limits and uncertainty around floor action. VOTrainer, The Best Sites for Finding Voice Over Work in 2026 [15]. Marketplace commissions and audiobook compensation examples for Fiverr, Upwork and ACX. Rain Intelligence, Voiceprint BIPA [7]. A current litigation-trend snapshot to use as a lead for checking court records rather than as a ruling. Enzuzo, Illinois Biometric Information Privacy Act, 2026 Guide [14]. A plain-language walkthrough of BIPA identifiers, notice, written release and retention duties. Fello AI, How to Stop AI from Training on Your Data, 2026 Privacy Guide [6]. Step paths for ChatGPT, Claude and Gemini with training versus logging distinctions. MemX, What AI Does With Your Chats, and How to Stop It [12]. A June 2026 cross-tool comparison with Gemini human-review retention and Copilot toggle detail. 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 questio