Hollywood and creative industries are fighting back against AI infringement, with one startup offering proprietary technology to detect suspected misuse of copyrighted material within generative AI models.
LightBar, a research platform focused on AI training data, positions itself as the evidence layer that makes those threats actionable.
The platform claims it runs “research campaigns” in which users generate outputs using structured prompts designed to test specific models or studio intellectual property.
“When lesser-known or minor characters appear accurately across generations, that pattern strengthens the signal of what may have been included in training data,” LightBar told Decrypt.
Submissions are processed through what the company describes as a proprietary analysis engine that measures “percentage likeness, distinctive character traits, and prominence.”
Over the past week, major studios have moved to formalize their infringement claims through legal channels. Industry groups and unions have also set out positions that treat certain AI outputs as potential violations of copyright and contractual rights.
The letters show owners are converting infringement concerns into formal enforcement pressure. At the same time, labor groups are asserting that certain AI-generated outputs implicate consent and compensation rights under existing contracts and law.
LightBar said it is in active discussions with studios as they consider potential legal or licensing action related to Seedance 2.0 and other AI models, with the goal of helping “shift the conversation and the leverage back in their favor.”
The company said the results are compiled into analyses that “outline the methodology, similarity metrics, and representative examples to support further review.”
“The current wave of disputes makes one thing clear: attribution and evidence are becoming the battleground of the AI economy,” Ram Kumar, core contributor at AI and blockchain infrastructure firm OpenLedger, told Decrypt.
Documenting model outputs “absolutely strengthens a studio’s negotiating position, but only if that documentation is structured, time-stamped, and cryptographically verifiable,” Kumar said.
Creating verifiable logs that connect prompts to outputs and specific model versions can convert resemblance into quantifiable proof, strengthening a rights holder’s position in court or licensing talks, even when the underlying training data cannot be directly traced, Kumar explained.
“In the long run, this won’t just affect disputes,” he said. “It will shape how future AI systems are built: with transparent reward pathways, accountable execution, and verifiable contribution tracking embedded at the protocol level.”
Decrypt has approached ByteDance for comment.
