PRAC3 Framework: Dual Perspectives in AI Governance
- PRAC3 Framework is a dual-use model that formalizes six governance pillars for synthetic voice applications and maps three praxes for social accessibility research.
- It defines Privacy, Reputation, Accountability, Consent, Credit, and Compensation to address emerging risks such as non-consensual voice cloning and data misuse in AI pipelines.
- Additionally, the framework organizes artifact, ecosystem, and epistemology praxes to uncover systemic barriers and promote integrative, reflexive research practices.
PRAC3 Framework is an ambiguous term in recent research usage. In the synthetic voice literature, PRAC3 is a formally introduced six-pillar framework—Privacy, Reputation, Accountability, Consent, Credit, and Compensation—proposed to explain long-tailed risks faced by voice actors in the AI data-economy (Sharma et al., 22 Jul 2025). In social accessibility research, “PRAC3” appears as community shorthand for the “Three Praxes Framework,” which organizes the field around Artifact, Ecosystem, and Epistemology, together with two cross-cutting stances and one reflexive cycle (Jang et al., 8 Mar 2026). The label therefore names two distinct frameworks rather than a single canonical model.
1. Terminological scope and disambiguation
The term has two documented uses.
| Usage of PRAC3 | Expansion | Domain |
|---|---|---|
| PRAC3 | Privacy, Reputation, Accountability, Consent, Credit, Compensation | Synthetic voice data-economy |
| “PRAC3” shorthand | Three Praxes | Social accessibility research |
In the first usage, PRAC3 was introduced because the traditional three Cs—Consent, Credit, and Compensation—do not capture long-tailed risks arising when voices function both as creative labor and biometric identifiers. In the second usage, the paper does not coin “PRAC3” formally, but community conversations use it as a shorthand because the framework centers “3 praxes, 2 stances, and 1 reflexive cycle” (Sharma et al., 22 Jul 2025, Jang et al., 8 Mar 2026).
A common source of confusion is that the two frameworks share a compact label while addressing different objects of analysis. One is a governance model for audio data, synthetic replication, and downstream deployment; the other is a field-level analytical map for understanding how accessibility research builds artifacts, studies ecosystems, and develops theory.
2. PRAC3 as a governance framework for the synthetic voice economy
The voice-sector PRAC3 framework emerges from the argument that early large-scale audio datasets such as LibriSpeech were built with hundreds of individual contributors whose voices later surfaced across commercial AI pipelines and cloning tools without transparent consent, credit, or safeguards (Sharma et al., 22 Jul 2025). The paper situates this problem in a synthetic voice economy in which actors reported non-consensual training, resale or “renting” of voices, and deployment of clones in decontextualized or harmful contexts, including erotic content, meme culture, political messaging, and scams.
Its empirical basis is an IRB-approved qualitative study of semi-structured interviews with 20 U.S.-based professional voice actors across roles including freelancers, contractors, and studio owners, with experience levels from 4 to 20 years. Interviews lasted 40–70 minutes, approximately 60 minutes, and covered workflow and data-sharing practices, contract experiences, awareness of generative AI, and perceptions of ownership, consent, privacy, and security. The analysis used deductive-inductive coding and thematic analysis, added a Participant Category code after initial coding, and generated four personas: Emerging Professionals, Solo Defender, Delegator, and Strategist (Sharma et al., 22 Jul 2025).
The framework is presented as an expansion of C3. The paper argues that consent can be bypassed through vague or hidden clauses, credit is inconsistently applied and does not prevent reputational or biometric harms, and compensation is mismatched to the enduring value of data in AI systems when one-time buyouts allow indefinite extraction of value from a voice (Sharma et al., 22 Jul 2025). PRAC3 therefore adds Privacy, Reputation, and Accountability as interdependent pillars intended to capture decoupling of vocal identity from context, authorship, and control.
3. The six pillars of the voice-sector PRAC3 model
Privacy is defined as treating voice data not merely as creative output but as biometric personal data, specifically a voiceprint, with heightened sensitivity and re-identification risk. Privacy includes long-term control over where and how voice recordings are reused, repurposed, or integrated into AI training. Documented risks include non-consensual training, silent scraping of public reels and demos, identity and security harms such as deepfake scams and swatting, and possible repurposing of recordings used in financial authentication. The paper notes that auditions, brief samples, and platform uploads can become training data for TTS in ways actors cannot track or retract (Sharma et al., 22 Jul 2025).
Reputation concerns harm to a voice actor’s public persona and professional standing when a voice or synthetic clone is deployed in contexts misaligned with the actor’s values or brand. The paper documents erotic or pornographic deepfakes, offensive political messaging, defamatory or low-quality productions, and meme-culture repurposing. The core mechanism is decontextualized deployment: the voice becomes detached from original intent, authorship, and contractual scope, which can confuse audiences and damage credibility.
Accountability denotes legal and technical pathways to identify who is responsible for misuse across complex AI data ecosystems and to enforce boundaries, redress, and removal. The paper emphasizes that dataset builders, platforms, intermediaries, model providers, downstream users, and malicious actors may all participate in the chain, while actors often discover harms only “in the wild,” after the fact. Hidden contract terms, audition reuse without hire, replacement of pick-ups with AI clones, and platform auto-translation without opt-in are presented as manifestations of accountability breakdown.
Consent is defined as explicit, informed, and revocable agreement governing data collection, training, deployment, and reuse across time and contexts. The paper describes risks arising from asymmetric NDAs, the unregulated audition stage, vague and perpetual license language, Exhibit A clauses allowing replication after recording, and platform submissions reused for AI training without clear opt-in (Sharma et al., 22 Jul 2025).
Credit refers to accurate and visible attribution of voice actors for their contributions, including recordings, characters, and datasets. The paper reports that credit is more common in audiobooks and rare in commercial or corporate voice work. It further argues that decontextualized synthetic content can blur attribution, erase original authorship, and create confusion between original and synthetic performances.
Compensation is defined as fair remuneration aligned with the enduring value of voice data in AI systems across reuse, training, and deployment. The paper documents one-time buyouts “in perpetuity,” widespread subsequent use without additional pay, displacement by synthetic voices, and replacement of “pick-ups” and minor roles by synthetic substitutes. A central claim is that long-tail monetization persists without residuals while training value is extracted indefinitely (Sharma et al., 22 Jul 2025).
The paper’s conceptual model depicts these six pillars as interdependent. Privacy breaches can flow into reputational harm when biometric voiceprints are reused in decontextualized deployments, while accountability is presented as the system-level condition that makes consent, credit, and compensation enforceable.
4. Risk taxonomy, workflow phases, and governance mechanisms in the voice PRAC3 literature
The paper organizes harms across the voice workflow—Discovery, Audition, Contracting, Recording/File Sharing—and maps recurring incidents to PRAC3 domains (Sharma et al., 22 Jul 2025). Examples include audition samples reused in national commercials without hiring the actor; hidden AI training and cloning clauses in Exhibit A; AI-generated porn using a game character voice; content sounding like an actor endorsing political agendas; platform-level translation or cloning without opt-in; cloned voices used in emergency scams; and assistant voices licensed and redistributed to third parties without additional compensation.
These incidents are described as long-tailed because synthetic replication without enforceable constraints detaches voice from context, authorship, and control. The paper argues that persistent threats arise because biometric voice signatures can be re-identified and cloned from short samples, reuse continues indefinitely across jurisdictions and products, and decontextualized deployments combine reputational damage with security risk.
The proposed governance response is explicitly multi-layered. Contractual mechanisms include AI riders such as NAVA, explicit opt-in or opt-out provisions for training, cloning, and translations, time limits, scenario exclusions such as no political or erotic uses, disclosure of all exhibits and attachments, and restrictions on resale or transfer without re-consent. Consent management mechanisms include consent registries at dataset and model levels, per-clip consent states, revocation pathways, and logs of training events and retraining. Provenance and audit mechanisms include dataset and model cards documenting voice source, scope, and allowed uses; traceability infrastructure spanning the full workflow; and chain-of-custody records linking clips to models and deployments. Technical measures include watermarking, voice signatures, labeling of synthetic outputs, pick-up detection, detection APIs, and takedown channels. Organizational measures include unionization, collective bargaining, designated contacts at platforms and model providers, time-bound response SLAs, and evidence standards for takedown and removal (Sharma et al., 22 Jul 2025).
The policy touchpoints named in the paper are Illinois BIPA, the EU AI Act, and CCPA. It also notes a copyright and publicity gap in the United States: copyright generally protects recordings rather than a person’s voice itself. The paper extends its implications to existing datasets such as LibriSpeech, Common Voice, and VoxCeleb, arguing for retrospective governance, provenance documentation, retroactive consent where feasible, opt-out and removal processes, dataset cards with allowed uses and restrictions, and training-lineage audits (Sharma et al., 22 Jul 2025).
5. PRAC3 as shorthand for the Three Praxes Framework in social accessibility research
In the second usage, PRAC3 refers to the Three Praxes Framework, a thematic review and map of social accessibility research (Jang et al., 8 Mar 2026). The framework was developed to understand a 15-year evolution in accessibility and assistive technology research from functional, device-centered work toward social accessibility, defined in the paper as how disability is lived and negotiated in social contexts. The stated aim is both diagnostic and aspirational: to surface structural disconnects and propose an integrated model for how building, relating, and theorizing can become mutually reinforcing practices that transform access.
The review covers research published between January 2011 and June 2025 in premier HCI venues: ACM venues ASSETS, CHI, CSCW, DIS, TACCESS, TOCHI, and UIST, and IEEE venues HRI, ISMAR, VR, and VIS. A two-pass keyword search retrieved 8,731 items, narrowed to 1,599 by adding social experience terms such as stigma, identity, and collaboration. Manual screening of titles and abstracts produced a full corpus of 605 papers, from which the authors curated a theoretically rich subset of 90 papers via theoretical sampling. Inclusion criteria required explicit focus on social or experiential phenomena, use of rich qualitative methodologies, or direct engagement with social or theoretical frameworks; purely technical, algorithmic, or functional-usability contributions were excluded. The constructivist grounded theory analysis produced 362 concepts, 46 codes, and six thematic territories, which were synthesized into the framework (Jang et al., 8 Mar 2026).
The paper identifies those six thematic territories as Communication as World-Building, Critical Making Practices, Access Ecologies, Methodological Justice, Identity Sovereignty Complex, and Algorithmic Disability Encounters. It also presents figures describing a triangular model of the three praxes, a corpus-construction flowchart, a fragmentation figure showing “missed connections,” and a DREEM cycle visualization showing recursive movement across praxes over five years (Jang et al., 8 Mar 2026).
6. Praxes, stances, and reflexive cycle in the Three Praxes Framework
The Three Praxes Framework is organized around three sites of practice. Artifact Praxis is “the constructive work of what the field builds.” It includes material interventions such as devices, systems, platforms, design methods, standards, and conceptual design artifacts. Typical methods include Research through Design, prototyping, DIY-AT and fabrication, speculative and critical design, co-design with disabled participants, and iterative deployment. Representative examples named in the paper include Higgins et al. on DIY-AT makerspaces, Bennett et al. on prostheses as identity work, Curtis and Neate on embodied AAC, and Ellis et al. on STEAM kits (Jang et al., 8 Mar 2026).
Ecosystem Praxis is “the relational work of where technology lives.” It addresses the social arrangements in which technologies and access practices unfold, including families, communities, organizations, platforms, and policy environments. Typical methods include ethnography, longitudinal deployment, case studies, organizational and process analysis, community-engaged research, and participatory methods focused on relationships and care infrastructures. Examples include McDonnell and Findlater on collective communication access, Crawford and Hamidi on LGBTQIA+ relationships, Rajapakse et al. on personal infrastructuring, Kaur et al. on online advocacy in India, and Chen et al. on social VR “mutes.”
Epistemology Praxis is “the critical work of why and how we know.” It interrogates assumptions about disability, knowledge production, and power, and develops concepts, critiques, and methods that reorient research and practice. The paper lists theoretical synthesis, methodological innovation, participatory and justice-oriented approaches, situated knowledge, and critical technoscience and STS lenses as typical methods. Examples include Baltaxe-Admony et al. on DREEM, Harrington et al. on race and disability, Kritika et al. on masking as “a learned safety skill,” and work by Mankoff, Williams, and Spiel on disability studies, crip HCI, and embodied critique (Jang et al., 8 Mar 2026).
Two cross-cutting stances modulate how projects understand and target change. Temporal Orientation is divided into remedial, adaptive, and generative. Remedial work accepts current systems as fixed and delivers immediate fixes; adaptive work helps people navigate existing systems over time; generative work envisions and builds toward different futures or norms. Stakeholder Focus is divided into individual, network, and societal levels. The paper uses these stances to distinguish, for example, an individual remedial prosthetic project from a generative societal makerspace project, or an adaptive captioning pipeline from a generative-network redesign of conversation as collaborative work (Jang et al., 8 Mar 2026).
The framework’s reflexive cycle models bidirectional flows among Artifact, Ecosystem, and Epistemology. The paper summarizes the cycle as: Situate, Translate, Test, Reflect, and Iterate. It also names six primary linkages: Artifact ↔ Epistemology, Epistemology ↔ Ecosystem, and Ecosystem ↔ Artifact. The central claim is that building, relating, and theorizing should inform each other continuously rather than remain isolated.
7. Fragmentation, limitations, and comparative significance
The central finding of the Three Praxes review is that the praxes largely operate in isolation. Constructive artifact work often proceeds without uptake of critical epistemologies; ecosystem studies identify systemic barriers but rarely become tools or methods; and epistemological critiques often remain abstract rather than translating into material interventions or organizational change (Jang et al., 8 Mar 2026). Documented consequences include insights staying academic, assistive technologies reinforcing existing barriers through conformity metrics, and privatization of response when families and couples absorb costs that institutions should bear.
The voice-sector PRAC3 paper presents an analogous concern in a different domain: risks persist because voice data move through complex, multi-stakeholder AI pipelines without provenance, traceability, enforceable constraints, or effective recourse (Sharma et al., 22 Jul 2025). Its limitations are also explicit. The sample is U.S.-centric, undisclosed practices may exist beyond the interviews, PRAC3 is presented as a conceptual framework, and non-union vulnerabilities are highlighted but not exhaustively quantified. The Three Praxes review likewise states that its corpus is limited to English-language works in premier ACM and IEEE venues from 2011–2025, excludes disability studies and other policy or practice venues, non-English outlets, majority-world contexts, and grey literature, and may underrepresent technical works with embedded social components because of the keyword strategy (Jang et al., 8 Mar 2026).
Taken together, the two literatures do not define a single unified PRAC3 framework. One formalizes six interdependent governance pillars for synthetic voices; the other maps three sites of practice, two stances, and one reflexive cycle for social accessibility research. This suggests a limited family resemblance rather than identity: both formulations emphasize interdependence, traceable linkages across stages or sites of practice, and the insufficiency of narrowly scoped interventions, but they operate in distinct empirical domains and should not be treated as interchangeable.