The “Derived Data” Dilemma in AI

The term “derived data” has always been a messy grey area in market data contracts. In the era of Generative AI and LLMs, it’s officially broken.

Historically, exchanges defined derived data using a simple test: 𝘊𝘢𝘯 𝘵𝘩𝘦 𝘶𝘯𝘥𝘦𝘳𝘭𝘺𝘪𝘯𝘨 𝘳𝘢𝘸 𝘥𝘢𝘵𝘢 𝘣𝘦 𝘳𝘦𝘷𝘦𝘳𝘴𝘦-𝘦𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘦𝘥 𝘰𝘳 𝘴𝘶𝘣𝘴𝘵𝘪𝘵𝘶𝘵𝘦𝘥?
If the answer was no (e.g., an index value or a risk metric), it was usually considered “derived” and freed from heavy redistribution fees.

AI shattered that framework. Here’s why:
1. Are model weights “Derived Data”? When you train a custom LLM on decades of proprietary tick data, does the weight matrix count as derived data? Data providers argue yes—claiming the model "embodies" their data. Consumers argue no—it’s a transformative mathematical process that extracts abstract patterns, not raw data.
2. The Illusion of “Irreversibility” Vendors historically allowed derived data if raw prices couldn't be extracted. But with modern prompt extraction and model inversion, "irreversible" isn't binary anymore. Vendors use this ambiguity to claim AI pipelines retain original data—and charge accordingly.
3. RAG and Real-Time Context If an AI agent pulls live order book data via Retrieval-Augmented Generation (RAG) to generate a trading summary, is the output a derivative work, a display use, or unauthorized redistribution? Most legacy contracts don't have an answer.
4. Overreaching Exchange Audits Facing potential revenue loss from AI automation, exchanges are expanding their contractual definitions of "derived work" to capture AI outputs and levy surprise back-license fees during audits.

Where I Come In ⚖️
As a market data licensing lawyer, I bridge the gap between complex AI architecture and legacy data vendor rights:
Contract& Audit Defence: Reviewing legacy agreements to identify hidden exposure before an audit hits.
AI Data Licensing Strategy: Negotiating explicit AI and model-training carve-outs for inputs vs. outputs.
Data Governance Frameworks: Structuring compliant RAG pipelines and training sets without triggering redistribution penalties.

If your firm is building AI tools on market data feeds using decade-old contract clauses, you are sitting on an audit landmine.

📩 Book a Consultation with me to audit your AI data exposure before your vendors do.