GPT-4o: OpenAI’s most advanced models appear to have been trained on copyrighted books behind paywalls — without permission — and that’s a big legal and ethical problem.

Question 1 of 3 for ChatGPT-4o: Please read the report "Beyond Public Access in LLM Pre-Training Data. Non-public book content in OpenAI’s Models" and tell me what it says in easy to understand language

ChatGPT-4o’s answer:
The report “Beyond Public Access in LLM Pre-Training Data: Non-public book content in OpenAI’s Models” by Sruly Rosenblat, Tim O’Reilly, and Ilan Strauss investigates whether OpenAI used copyrighted, non-public book content from O’Reilly Media to train its AI models without permission. Here's a simplified breakdown of what the report says:

🧠 What the researchers did

• They legally obtained 34 copyrighted O’Reilly Media books, which include both public and paywalled (non-public) parts.

• They tested three OpenAI models: GPT-3.5 Turbo, GPT-4o, and GPT-4o Mini to see if those models “recognized” the books, which would suggest they were trained on them.

• The key method used is called a “membership inference attack” (DE-COP), which checks if an AI model has “seen” or memorized certain texts before.

📊 Key Findings

1. GPT-4o (the most recent model) seems to strongly recognize non-public, paywalled content, more so than public content.

   • It scored 82% accuracy in identifying non-public text from its training set (compared to random chance at 50%)​.

   • It did better on paywalled data than public previews, which is suspicious because you'd expect the model to mostly learn from freely available data.

2. GPT-3.5 Turbo (older model) didn’t show this same pattern.

   • It recognized public content more than non-public, suggesting earlier OpenAI models may have been more careful with copyright.

3. GPT-4o Mini (smaller model) showed no real recognition, likely because it's too small to memorize much — not necessarily because it wasn’t trained on the content.

🔍 How do they know this?

• They split the book paragraphs into public vs. non-public and also by publication date — if a book came out after the model was trained, then it definitely couldn’t have been used.

• By comparing models trained in the same timeframe (e.g. GPT-4o and GPT-4o Mini), they could rule out “language style over time” as the reason for differences.

⚠️ Why this matters

• The findings suggest OpenAI’s newer models are likely trained on copyrighted, paywalled content without permission.

• This practice may violate copyright laws and threatens the sustainability of content creation, since creators aren’t being paid or credited.

• The report warns that this could lead to an “extractive dead end,” where AI depletes the internet’s creative resources and undermines quality content production​.

🔧 Policy and market recommendations

• The authors argue for more transparency from AI companies about where their training data comes from.

• They suggest creating formal licensing markets for training data — where content creators get paid.

• They highlight that the EU AI Act might help push companies toward better disclosure standards.

🧾 Bottom Line

Question 2 of 3 for ChatGPT-4o: List the most surprising, controversial and valuable statements and findings in the report.

ChatGPT-4o’s answer:
Here’s a list of the most surprising, controversial, and valuable statements and findings from the report “Beyond Public Access in LLM Pre-Training Data: Non-public book content in OpenAI’s Models”:

🔥 Most Surprising Findings

1. GPT-4o shows higher recognition of paywalled content than public content

1. Recognition increased over time in newer models

1. GPT-4o Mini (a smaller model) showed no such recognition

⚖️ Most Controversial Claims

1. OpenAI likely trained GPT-4o on non-public, copyrighted content without permission

1. Use of LibGen (a pirate database) may be the access point

1. OpenAI and others are lobbying to exempt model training from copyright altogether

1. A tragedy of the commons is underway

💡 Most Valuable Contributions

1. First-of-its-kind use of legally obtained, non-public data to detect access violations

1. Strong evidence supporting the need for AI training data disclosure laws

1. Proposes a market-based solution: functional AI content licensing ecosystems

1. Recommends technical audits like DE-COP to enforce accountability

Question 3 of 3 for ChatGPT-4o: Explain how publishers and regulators can make use of this information. Also, explain how AI developers should act in case they want to prevent liability as a result of this situation.

ChatGPT-4o’s answer:
Great question — this report delivers a wake-up call for both content owners and AI developers. Here’s a breakdown of how publishers, regulators, and AI developerscan make strategic use of these findings:

📚 For Publishers & Copyright Holders

🛠️ What they can do with this report:

1. Use the methodology as evidence

   • The DE-COP method used in the report shows that unauthorized training is detectable.

   • Publishers can replicate or commission similar studies using their own books or articles to build proof of infringement.

2. Pressure AI companies for licensing agreements

   • The report shows that non-public content improves model quality, making a clear economic case for licensing.

   • Publishers can leverage this to negotiate royalties, tiered access deals, or content usage limits.

3. Support disclosure requirements

   • Advocate for laws (like the EU AI Act) to be enforced or expanded to require detailed training data disclosures — not just vague summaries.

4. Join or initiate litigation or lobbying efforts

   • This report helps validate lawsuits (e.g. NYT v. OpenAI) and undermines the "fair use" defense.

   • It supports coalition-building among content producers to pursue legal reforms or collective bargaining rights.

5. Experiment with licensing marketplaces

   • Explore platforms like Spawning.ai, News Licensing Group, or Music AI licensing tools to monetize content used for AI training.

🏛️ For Regulators & Policymakers

🎯 How they can use the findings:

1. Mandate stronger transparency standards

   • Require AI companies to publish detailed training data audits or use tools like DE-COP to verify claims.

   • Push for third-party certification of data provenance.

2. Clarify copyright liability

   • Ensure that training on non-public content without consent is explicitly illegal.

   • Introduce safe harbor conditions for AI developers who use licensed or traceable datasets.

3. Build technical oversight capacity

   • Fund or support data auditing tools and institutions (e.g., academic or nonprofit watchdogs).

   • Set up regulatory sandboxes for AI data governance experimentation.

4. Promote market-based solutions

   • Create incentives or standards for the development of trusted content marketplaces where rights can be negotiated and tracked.

🤖 For AI Developers

⚠️ What they should do to avoid legal risk:

1. Stop using pirated or unauthorized data

   • Avoid scraping from sources like LibGen, Sci-Hub, or sites protected by robots.txt and TOS.

   • Audit current datasets — even retrospectively — to identify problematic content.

2. Build from licensed or “clean” datasets

   • Shift toward using datasets like Common Corpus, LAION with verified licenses, or purpose-built licensed corpora.

3. Disclose training sources voluntarily

   • Proactively share detailed breakdowns of training data — even before regulators require it.

   • Use model cards with dataset descriptions and audit logs.

4. Engage in good-faith negotiations

   • Partner with content providers and publishers to license data ethically and legally.

   • Consider revenue-sharing models for outputs derived from valuable proprietary content.

5. Develop internal risk assessments

   • Integrate compliance reviews into the data pipeline, especially for foundation model training.

   • Track not just input sources but how they affect output liability (e.g., reproducing long book passages).

🧩 Bottom Line:

This report is a smoking gun showing that:

• Non-public content improves model performance

• That content was likely used without permission

• It’s now technically feasible to detect misuse

That means:

• Publishers have a stronger legal and commercial hand.

• Regulators have a clear signal to act.

• Developers must clean up their pipelines — or face lawsuits, fines, and reputational damage.