Amazon and Twitch Face a Class Action Over How They Train Their AI Models

A former Twitch employee is suing Amazon, alleging they used her private data to train a generative AI model without her permission. The lawsuit, reported by Insider Gaming, claims the company scraped emails, chat logs, and other internal documents to feed a large language model. The employee says she was never informed or given a chance to opt out. This creates a direct link between how data is collected and how a model learns.

What actually happened

Gisela Salazar, a former Twitch employee, claims that data she created while working for the company was uploaded to a public AI training set. She alleges this happened through internal tools that were not designed for public data sharing. The core issue is that the data used to train the AI was not gathered with proper consent or under a data usage policy that allowed for public training. The AI model effectively learned from her private work life.

The governance failure mode

This case represents a failure in data grounding and control. In AI risk, grounding refers to the reliability of the data a model is trained on. If the data is scraped or used without clear ownership or consent, the model is not grounded in proper legal or ethical boundaries. The failure here is not that the AI said something false. It is that the company failed to control the data pipeline. They treated internal communication tools like public internet archives. This creates a liability risk where the company is responsible for outputs generated from data it should not have used.

What the law actually requires

Two major privacy frameworks are at play here. The General Data Protection Regulation in Europe requires companies to have a lawful basis for processing personal data. Using data for training a public model without explicit consent is rarely a lawful basis. In California, the CCPA and CPRA give consumers the right to know what personal data is collected and the right to opt out of its sale or specific uses. If a company uses employee data to train a commercial AI model, they must be transparent about that use. The EU AI Act also requires transparency when personal data is used in AI systems. Companies must inform users when an AI is making decisions or generating content based on their personal information. Failing to do so violates these core principles.

Why this matters for risk teams

This incident shows that generative AI is a double-edged sword for data privacy. Using internal data to train models can save money and improve performance, but it introduces significant legal risk. Risk teams must treat internal data repositories as high-security assets, not just raw material for AI projects. A model trained on improperly sourced data is not just risky to deploy; the act of creating it is risky. It exposes the company to class-action lawsuits and regulatory fines. Companies cannot assume that data sitting on an internal server is fair game for a public model.

What to do

Risk teams need to establish strict boundaries around data usage for AI training.

• Audit your training data sources. You must know exactly where every data point comes from and whether the legal basis for its use is valid.
• Implement a data consent layer. Before any employee data is used to train a model, there must be a clear mechanism for opting out or opting in.
• Create a data classification policy. Clearly label which data can be used for public training and which must remain strictly internal.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.