Tennessee Bill Criminalizes Human-Like AI Training

A voice assistant that sounds exactly like a deceased family member. A chatbot that mimics a celebrity’s speech patterns down to their signature phrases. A virtual companion trained to replicate a real person’s mannerisms and emotional responses. Tennessee’s newly proposed legislation would make creating these AI systems a criminal offense, marking the first state-level attempt to regulate synthetic personhood in artificial intelligence.

Training Approach Under Scrutiny

The Tennessee bill, formally known as the ELVIS Act (Ensuring Likeness Voice and Image Security), targets a specific category of AI development: models trained to replicate identifiable human characteristics without explicit consent. This includes voice cloning systems that analyze speech patterns, facial recognition models that generate deepfake videos, and language models fine-tuned on an individual’s writing style or conversational patterns.

Current AI training methods for human replication typically involve collecting extensive datasets of a person’s voice recordings, photographs, videos, or text. Developers use these materials to fine-tune foundation models like Stable Diffusion for images or ElevenLabs-style architectures for voice synthesis. The process often requires only 10-30 minutes of audio or several hundred images to produce convincing results.

Under the proposed legislation, obtaining this training data without documented consent would constitute a Class A misdemeanor, with penalties escalating to felony charges if the resulting AI is used for commercial purposes or causes demonstrable harm. The bill extends Tennessee’s existing right of publicity laws into the AI domain, treating synthetic replication as equivalent to unauthorized use of someone’s likeness.

Notable Results and Industry Impact

The legislation arrives amid growing concerns about AI-generated impersonation. Voice cloning scams have already cost victims millions, with fraudsters using tools like https://github.com/CorentinJ/Real-Time-Voice-Cloning to create fake emergency calls from family members. Celebrity deepfakes proliferate across social media, while unauthorized AI recreations of deceased artists raise complex ethical questions about digital afterlife rights.

Tennessee’s music industry connections make this particularly significant. Nashville-based artists and their estates have pushed for stronger protections after incidents involving AI-generated songs mimicking famous voices. The bill would give individuals and their estates legal recourse to pursue both criminal charges and civil damages against unauthorized AI replication.

Several AI companies have already implemented voluntary consent mechanisms. OpenAI’s Voice Engine requires explicit authorization before cloning voices, while Metaphysic’s platform maintains strict identity verification protocols. However, open-source alternatives remain widely available, creating enforcement challenges that the Tennessee bill attempts to address through criminal penalties.

Running Locally and Technical Realities

The legislation faces practical complications around enforcement, particularly with locally-run AI models. Developers can download open-source voice cloning repositories, image generation models, or text-based personality replicators and run them entirely on personal hardware without internet connectivity. Popular implementations include:

# Example voice cloning workflow (educational purposes)
from TTS.api import TTS
model = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts")
model.tts_to_file(text="Sample output", 
                  speaker_wav="reference_audio.wav",
                  file_path="cloned_voice.wav")

This code snippet represents the type of accessible technology that makes detection difficult. When models run locally without cloud services or API calls, identifying violations requires either self-reporting, third-party complaints, or discovery during unrelated investigations.

The bill’s language attempts to address this by criminalizing the training process itself, not just distribution or commercial use. However, proving that someone trained a model on unauthorized data presents significant evidentiary challenges, especially when training happens on personal devices.

Trade-offs Between Protection and Innovation

Tennessee’s approach creates tension between individual rights and AI research. Legitimate use cases exist for training models on human characteristics: medical applications that recreate patient voices after laryngectomy, historical preservation projects, accessibility tools, and academic research into human communication patterns.

The bill includes exemptions for news reporting, commentary, criticism, and parody, mirroring fair use principles from copyright law. Yet these carve-outs may prove insufficient for researchers developing new AI techniques or companies building personalization features that adapt to user characteristics without explicitly “replicating” them.

Broader implications extend beyond Tennessee. If the legislation survives legal challenges, other states may adopt similar frameworks, creating a patchwork of regulations that complicates AI development. Companies would need consent management systems tracking permissions across jurisdictions, while open-source developers face uncertainty about which training approaches remain legal.

The bill represents a fundamental question about AI governance: whether existing legal frameworks around publicity rights and identity protection adequately address synthetic replication, or whether new criminal statutes are necessary to prevent harm in an era where anyone with technical skills can create convincing digital doubles.