This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Major AI developers including OpenAI, Google, and Meta are reported to have engaged in aggressive data collection practices to train their models, potentially violating copyright laws and platform terms of service. These practices included transcribing YouTube videos without authorization and discussing the acquisition of publishing houses to secure training data. The report highlights the industry's desperate race for high-quality data to advance large language models.
In late 2021, OpenAI and other tech giants like Google and Meta reportedly faced data shortages for training AI models. OpenAI is said to have developed a tool called Whisper to transcribe over one million hours of YouTube videos, potentially violating YouTube’s terms of service. Similarly, Google allegedly transcribed YouTube videos, risking copyright infringements. Meta reportedly explored summarizing copyrighted texts without permission and debated acquiring Simon & Schuster for data.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems capable of creating economic or cultural value, including through reproduction of human innovation or creativity (e.g., art, music, writing, coding, invention), destabilizing economic and social systems that rely on human effort. The ubiquity of AI-generated content may lead to reduced appreciation for human skills, disruption of creative and knowledge-based industries, and homogenization of cultural experiences.
Also classified as
Google updated its terms to harvest user data from Google Docs and Sheets, raising significant privacy concerns.
The intense race to lead in AI capabilities drove companies to aggressively bypass terms of service and copyright boundaries.
The unauthorized ingestion of books, articles, and creative videos devalues the intellectual property and creative efforts of human authors and creators without consent or compensation.
Human
Due to a decision or action made by humans
Intentional
Due to an expected outcome from pursuing a goal
Pre-deployment
Occurring before the AI is deployed
The data harvesting and policy changes were intentional decisions made by human executives and engineers during the pre-deployment training phase of the AI models.
Major US AI developers engaged in aggressive, unauthorized data harvesting of YouTube videos, books, and user documents to train large language models. While highlighting the critical strategic bottleneck of data in the global AI race, the incident's current national security impact is minor, representing domestic intellectual property and privacy disputes manageable through standard legal procedures.
The Whisper speech recognition tool was used to perform a specific, well-defined function of transcribing audio files into text without autonomous planning.
Limited Risk: The models being trained (such as GPT-4 and Bard) function as chatbots, which are classified under Limited Risk and require transparency obligations.
Whisper's primary goal was voice recognition to transcribe YouTube videos, which was then used to train GPT-4, a chatbot.
The report is highly detailed, drawing on internal recordings, emails, and statements from multiple tech companies, providing a clear picture of the industry's data collection practices.
AI developers argue that training on public data constitutes fair use and does not represent copyright infringement or theft.
Technology
Looming Training Data Depletion
AI models consumed high-quality internet text faster than it was produced.
Model Collapse from Synthetic Data
Training on AI-generated data can reinforce quirks, errors, and limitations.
Automated Scraping Tools
Whisper speech-to-text tool enabled large-scale YouTube transcription.
Data Inputs
Use of Copyrighted Material
Scraping books, articles, and videos without creator permission or licensing.
Privacy Policy Violations
Using personal consumer data from Google Docs and Maps for AI training.
Lack of High-Quality Datasets
Exhaustion of reputable English-language web text forced risky sourcing.
Human Factors
Prioritizing Capabilities over Ethics
Engineers and executives ignored ethical warnings regarding IP theft.
Desire to Beat Competitors
Pressure to match ChatGPT led to rushed and risky data acquisition.
User Privacy Overlook
Companies assumed user consent or quietly altered terms of service.
Process and Methods
Bypassing Terms of Service
Scraping YouTube videos using automated means despite platform prohibitions.
Quiet Terms of Service Changes
Google updated terms over a holiday weekend to expand data harvesting.
Inadequate Licensing Processes
Negotiating licenses with creators was deemed too slow for rapid AI scaling.
Management
Scale is All You Need Culture
Leadership focused entirely on model size and token count over data safety.
Dismissal of Legal Concerns
Meta and OpenAI leadership accepted lawsuit risks to procure training data.
Aggressive Competitive Pressure
CEOs pushed teams to deliver rival chatbots regardless of data sourcing.
Regulatory Environment
Vague Fair Use Interpretations
Ambiguity in copyright laws allowed companies to claim fair use for training.
Lagging Copyright Guidance
The Copyright Office was slow to issue clear rules for the AI era.
Weak Enforcement of Platform Rules
Google did not block OpenAI scraping to avoid scrutiny of its own methods.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
