OpenAI, Google, and Meta Alleged to Have Overstepped Legal Boundaries for Training AI

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.

Risk classification

• Primary risk domain: 6 Socioeconomic & Environmental
• Primary risk subdomain: 6.3 Economic and cultural devaluation of human effort

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.

Additional risk subdomains

• 6.4 Competitive dynamics: The intense race to lead in AI capabilities drove companies to aggressively bypass terms of service and copyright boundaries.
• 2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive information: Google updated its terms to harvest user data from Google Docs and Sheets, raising significant privacy concerns.

Causal factors

• Entity: Human
• Intent: Intentional
• Timing: Pre-deployment

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.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

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.

AI system and alleged parties

• AI system: GPT-4, Google AI models, Whisper (Google, OpenAI)
• AI purpose: Voice Recognition; Chatbot
• Behaviour type: Tool
• Alleged developer: OpenAI, Meta, Google
• Alleged deployer: OpenAI, Meta, Google
• Alleged harmed parties: YouTube creators, General public, Content creators

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Substantial. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Minor
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Negligible, indirect Negligible
• Civil rights: direct Substantial, indirect Substantial
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Substantial, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Negligible, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Financial

Reported: The report describes potential financial losses to creators and publishers due to unlicensed use of their works, though specific total dollar figures are not quantified in the text.

Indirectly caused: Unlicensed ingestion of copyrighted works potentially deprives creators of licensing revenue, though specific amounts are not detailed.

Inferred additional harm: Significant financial losses to the creative and publishing industries due to lost licensing fees and devaluation of human-authored content, potentially in the millions of dollars.

Civil rights

Reported: The report explicitly describes potential violations of creators' intellectual property rights and user privacy rights.

Directly caused: OpenAI and Google transcribed over one million hours of YouTube videos, potentially violating creators' copyrights.

Indirectly caused: Meta and Google discussed or implemented policies to scrape user data and copyrighted books without explicit consent.

Inferred additional harm: Broad-scale infringement of intellectual property rights of millions of creators and authors globally whose works were ingested without consent.

Privacy

Reported: The report explicitly describes privacy concerns regarding Google's policy changes to tap user data and Meta's history with Cambridge Analytica.

Directly caused: Google changed its terms of service to allow tapping publicly available Google Docs and other user-generated content for AI training.

Inferred additional harm: Potential exposure of private or semi-private user data if 'publicly available' settings on Google Docs were poorly understood by users.

People affected

• Occurrences reported: 1
• People reportedly harmed: 10000
• People reportedly exposed: 10000

Potential causes

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.

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.

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.

Information quality

• Classification confidence: High
• Reason for confidence: 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.
• Ambiguities identified: The legal definition of fair use regarding AI training remains highly contested and unresolved in courts.
• Alternative interpretations: AI developers argue that training on public data constitutes fair use and does not represent copyright infringement or theft.

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.

• Overall national security impact: Minor
• Scope: Multiple nations
• Primary target: No clear primary
• Alleged perpetrator: US Technology Companies (OpenAI, Google, Meta)

Threat characteristics

• Imminence: Long-term. Represents an ongoing strategic concern regarding data access and copyright rather than an active national security crisis.
• Autonomy: Human-controlled. The data harvesting and policy decisions were entirely directed and executed by human executives and engineers, using AI only as a basic utility tool.
• Novelty: Evolved capability. Represents a significant advancement in the scale and automation of data harvesting through AI-driven transcription and policy modifications.

Impact by dimension

• Physical security: Negligible. No physical threat, kinetic capability, or threat to critical infrastructure is indicated in the incident details.
• Information security: Negligible. No compromise of classified systems, intelligence gathering methods, or state-sponsored information warfare operations.
• Sovereignty: Negligible. Core government operations, electoral systems, and state sovereignty were not targeted or disrupted by these commercial data collection practices.
• Economic security: Minor. Highlights the strategic bottleneck of training data in the race for AI leadership, but impacts are currently limited to domestic civil copyright disputes and standard legal challenges.
• Societal stability: Minor. Involves large-scale unauthorized harvesting of creator content and user data, raising privacy and intellectual property concerns, but does not threaten overall societal stability.