Samsung employees three times leaked source code and a recording of an internal meeting to ChatGPT. Meta planned to train its models on posts and correspondence of Facebook and Instagram users. And Anthropic downloaded more than 5 million pirated copies of books for training. AI companies have already turned the user footprint into an asset. Hence, in various countries the question has become acute: this new and rapidly developing topic needs to

OpenAI directly states that in its consumer services, content may be used to train models if the user has not declined that option, while business products and APIs are excluded by default from training. In 2025, Anthropic introduced a choice for Claude Free, Pro, and Max users: whether to allow chats and coding sessions to improve Claude. With consent, the retention period for such data increases to five years, whereas with refusal, a 30-day regime is maintained.

For organizations, this means double vulnerability: external AI services become work tools, and employees may transfer internal documents, software code, commercial plans, databases, and personal data there. This is why many international organizations describe Shadow AI as already a reality: up to 58% of employees regularly use AI at work, 44% applied AI contrary to internal policies and leadership, 57% admitted errors due to AI, and 58% relied on results without verification.

Mass AI adoption: from experiment to working infrastructure

Based on available statistics, AI has stopped being niche for "ordinary" users. Roughly one-third of adults have ever used an AI chatbot, 72% have heard of chatbots at least once, and 57% interact with AI at least once a day, several times a week or more.

In February 2026, OpenAI announced 900 million weekly active ChatGPT users and 50 million paid subscribers, making AI a channel for accessing knowledge, workplace automation, and digital services rather than a separate software product.

The practical risk for companies, in turn, is not a hypothetical "superintelligent" threat but ordinary work habits: copy code, contract, client table, meeting transcript, or strategic document into a convenient external chat, analyze it, and quickly find errors.

To avoid this, an expert lists practical risk management measures:

• data classification,
• a list of approved AI services,
• prohibition or DLP control of transferring certain categories of data to external services,
• logging and auditing of AI requests,
• employee training on specific examples.

For instance, employees in Samsung's semiconductor division implemented sensitive corporate data in ChatGPT across three separate cases, including debugging source code, optimization software code, and recording an internal meeting to prepare a protocol.

"Data Enrichment": how user footprint becomes an asset for AI companies

In the context of AI, data enrichment means not only adding missing fields to the customer database but also transforming scattered texts, files, images, actions, clicks, reviews, geolocation, request history, and user behavior into more valuable training and productive signals.

OpenAI's privacy policy enumerates a broad set of collected data: account data, payment information, user content, uploaded files, images, audio, video, and much more. Meta was also at the center of controversy over AI training data in 2024: MIT Technology Review reported that the company planned to use posts and interactions with chatbots on Facebook, Instagram, Threads, and WhatsApp for training generative models, while U.S. users and countries without strong data protection legislation effectively had no reliable way to prohibit such processing.

Books as raw material for models: licenses, courts, and "digital libraries"

The problem of AI training on books has moved from theoretical discussion to judicial and contractual practice. The most illustrative case was the Bartz v. Anthropic lawsuit. Authors Alliance described Judge William Olsap's June 2025 decision, where training...