CCPA & AI Compliance

How the CCPA and CPRA apply to personal information in AI training data and model outputs

Overview

The California Consumer Privacy Act (CCPA), as amended by the California Privacy Rights Act (CPRA), is the most comprehensive consumer privacy law in the United States. Effective since January 2020 and significantly strengthened by CPRA amendments taking effect in 2023, the CCPA grants California residents broad rights over their personal information and imposes obligations on businesses that collect, use, or sell that data. For AI development teams, the CCPA has direct implications for how training data is sourced, processed, and retained.

The CCPA applies to for-profit businesses that do business in California and meet any of three thresholds: annual gross revenue above US$25 million, adjusted for inflation (US$26,625,000 from January 2025); buying, selling, or sharing the personal information of 100,000 or more California consumers or households; or deriving 50 percent or more of annual revenue from selling or sharing personal information. Given the scale of data required for AI training, many organizations building AI systems will meet these thresholds. The CPRA amendments also established the California Privacy Protection Agency (CPPA), the first dedicated privacy enforcement body in the United States.

For AI and machine learning, the CCPA's definition of personal information is expansively broad, covering obvious identifiers as well as inferences drawn from personal information to create consumer profiles. This means that AI models trained on personal data, and the predictions those models generate, may themselves constitute personal information under the CCPA. Organizations must therefore consider the regulatory implications of their model outputs and the profiles they create through AI-driven analysis, as well as of their training data.

AI-Specific Requirements

The CCPA establishes several consumer rights that directly affect AI training data pipelines. The right to know requires businesses to disclose what personal information they collect, the purposes for collection, and the categories of third parties with whom data is shared. For AI teams, this means maintaining clear records of what personal data enters training pipelines and how it is used. The right to delete obligates businesses to erase consumer data upon request, which raises complex questions about whether trained model weights must be retrained to "unlearn" deleted data.

The CPRA amendments introduced the right to limit use and disclosure of sensitive personal information, which includes precise geolocation, racial or ethnic origin, health information, and financial account details. AI training datasets frequently contain these sensitive categories, and organizations must provide consumers a mechanism to limit the use of their sensitive data to what is reasonably necessary for the expected service. The right to opt out of the sale or sharing of personal information is also critical. If training data is sourced from data brokers or shared with third-party model training services, consumers must have the ability to opt out.

Purpose limitation under the CPRA restricts businesses from using personal information for purposes materially different from or incompatible with the purposes disclosed at collection. This means organizations cannot simply repurpose customer data collected for service delivery into AI training datasets without providing additional notice and obtaining appropriate consent. Data minimization requirements further mandate that collection and processing be limited to what is reasonably necessary and proportionate to the disclosed purposes. Administrative fines of US$2,500 per violation or US$7,500 per intentional violation, adjusted for inflation each year, can accumulate rapidly across large datasets. In 2025 California also finalized regulations on automated decision-making technology, risk assessments and cybersecurity audits, with obligations phasing in from 2026.

How Ertas Helps

Ertas Studio trains on Ertas's cloud GPUs, so a training dataset is uploaded to Ertas and stored so runs can be repeated. If it contains consumers' personal information, Ertas receives it as part of your processing, and that belongs in your disclosures and your service provider contracts. The simpler route is to minimize first: train on de-identified data, or on only the fields the task needs. Data Craft, Studio's dataset builder, lets you edit each example and track its approval status before training. You can delete uploaded datasets, which helps when a deletion request reaches your source data.

Inference is where personal information flows continuously. Export the trained model as GGUF or safetensors and run it in your own environment, and consumers' prompts and the model's outputs never reach Ertas, which keeps Ertas out of those disclosures. You own the model, and its version stays fixed until you retrain. The adaptive model endpoint and enterprise builds can run the model in your own cloud or on-premise.

Notices, opt-out and limit-use mechanisms, consumer request handling, and the decision on whether a deletion request requires retraining a model all stay with you, as do risk assessments and automated decision-making obligations under California's newer regulations.