AI Training Data

State of play.

• Big Tech has shifted from scraping to licensing as the primary training-data acquisition strategy. A Reuters investigation documents paid deals with Photobucket, News Corp, Reddit, Amazon, Axel Springer, and The New York Times—with image licensing pricing ranging from cents per photo to higher rates for video—though most agreement terms remain undisclosed .
• AI meeting-tool wiretapping litigation has hardened into a systematic challenge. Chamberlain v. Granola, Inc. in the Northern District of California joins earlier suits against Otter.ai and Cresta, all testing whether AI notetakers that join meetings invisibly and harvest recordings for model training are third-party interceptors under federal and state wiretap statutes .
• Connecticut has enacted the most aggressive state-level AI training data disclosure requirement in the country. The amended CTDPA, effective July 1, 2026, requires controllers to disclose whether data trains large language models, imposes categorical restrictions on minors' data, and covers any business processing even a single Connecticut resident's sensitive data .
• Vendor MSA "service improvement" clauses are now a recognized AI training data liability vector. Kilpatrick Townsend, Consilium Law, and SiLaw have published redlining guides instructing customer-side counsel to demand explicit "no training," "no commingling," and "no retention" provisions—framing existing opt-out defaults as unauthorized secondary exploitation of proprietary data .
• AI chatbot conversations are established as discoverable evidence. A Stanford HAI study confirmed all six major providers log inputs for model training by default; a federal arson trial introduced ChatGPT logs as evidence; and OpenAI operates under a court order requiring indefinite retention of chat logs .
• For counsel advising enterprise clients deploying AI tools or negotiating SaaS and AI vendor agreements, the practical baseline is that training-data exposure now runs simultaneously through wiretap litigation, state privacy enforcement, vendor contract defaults, and discovery—requiring immediate audit of consent frameworks, MSA language, and employee AI-use policies.

Where things stand.

• The AI training data market is formalizing through paid licensing, but scraping litigation continues in parallel. Major platforms are purchasing proprietary datasets while copyright and fair-use suits over earlier scraping practices remain active; the two tracks are not mutually exclusive .
• All major AI chatbots train on user data by default, with opt-out mechanisms that are neither uniform nor fully transparent. A Stanford HAI study documents extended retention periods, opaque de-identification claims, and inadequate children's data safeguards across ChatGPT, Gemini, Claude, and Perplexity .
• Connecticut is the leading state on AI training data disclosure obligations. The amended CTDPA requires LLM training disclosure, imposes categorical minors' data restrictions, mandates data broker registration by January 1, 2027, and requires profiling impact assessments for processing created after August 1, 2026—making prior compliance assessments obsolete .
• Vendor MSA "service improvement" language is the primary commercial contract risk vector. Industry-wide opt-out defaults for model training mean customer proprietary and potentially privileged data is being used for AI development absent explicit contractual prohibition; the compliance gap between regulated government contracts and commercial MSAs is material .
• AI meeting-tool wiretap litigation is testing the third-party interceptor theory at scale. The Granola, Otter.ai, and Cresta suits collectively challenge whether participant-level consent is required before any AI transcription or summary tool joins a meeting—with implications for every workplace AI notetaker deployment .
• Wearable AI devices are generating a distinct training-data consent litigation track. A federal class action against Meta and Luxottica alleges Ray-Ban smart glasses routed intimate user footage to Kenyan contractors for AI training without disclosure, following Meta's April 2025 privacy policy update making AI features default-enabled .
• The DOJ bulk sensitive data transfer rule is a live compliance obligation. Codified at 28 C.F.R. Part 202 under EO 14117, full enforcement commenced October 6, 2026, covering AI training arrangements that touch health, genomic, or other sensitive data flowing to countries of concern .
• State regulatory cascades are compressing compliance timelines. California's AB 566, AB 853, and SB 53 require training data source disclosure and opt-out mechanisms for automated decision-making by January 2027; the EU AI Act reached full implementation August 2026 .
• IP treatment of AI training data is jurisdiction-dependent and unsettled. China, the UK/EU, and the US apply divergent standards on human authorship, fair use, and ownership of AI-generated outputs, creating compliance exposure for any cross-border training dataset .
• Attorney use of public AI tools implicates ABA Model Rule 1.6(c) and privilege. ABA Formal Opinion 512 reaffirmed duties of competence, supervision, and confidentiality; privacy toggles do not satisfy the ethical standard for preventing unintended disclosure of client data .
• Novel data-collection business models are outpacing consent frameworks. Shift, an AI training startup, offers free NYC home cleaning in exchange for head-mounted camera footage of customers' homes, licensed to AI labs for physical-robot training—with anonymization protocols unverified and multi-jurisdictional regulatory response unsettled .

Latest developments.

• Major AI developers shift to paid licensing for training data, with Reuters documenting deals covering 13 billion Photobucket images, paywalled news archives, and platform chat logs—pricing and terms largely undisclosed .
• ShieldFont launches as open-source anti-scraping tool, using OpenType glyph substitution to feed AI scrapers grammatically plausible decoy text while rendering correct text to human readers—legal implications and effectiveness at scale untested .
• Chamberlain v. Granola, Inc. filed in N.D. Cal., alleging invisible meeting-joining, no-consent recording, and default-enabled AI model training on captured content—part of a wave also targeting Otter.ai and Cresta .
• Kilpatrick Townsend, Consilium Law, and SiLaw publish redlining guides instructing customer-side counsel to demand explicit AI training prohibitions in MSAs, flagging industry-wide opt-out defaults as unauthorized data exploitation .
• Connecticut CTDPA amendments took effect July 1, 2026, requiring LLM training disclosure, lowering coverage threshold to 35,000 residents, and imposing categorical minors' data restrictions .
• Stanford HAI study and Palisades arson trial establish that AI chatbot conversations are logged by default and are fully discoverable—OpenAI operating under court order for indefinite chat log retention .
• Federal class action against Meta and Luxottica over Ray-Ban smart glasses routing intimate user footage to Kenyan contractors for AI training without disclosure .
• Shift launches free NYC home cleaning in exchange for head-mounted camera footage licensed to AI labs for physical-robot training data .

Active questions and open splits.

• Does the shift to paid licensing extinguish or merely supplement copyright and fair-use claims over earlier scraping? Licensing deals with major publishers do not resolve suits brought by creators whose works were scraped before agreements were in place, and the deals' terms may not cover all categories of prior use .
• Are AI meeting bots third-party interceptors under ECPA and state wiretap statutes? The Granola, Otter.ai, and Cresta suits present the question directly—one district court has held AI call-center use is not a federal wiretap violation, but the invisible-joining and default-training-enabled fact pattern is distinguishable and unsettled .
• What contractual language actually blocks vendor AI training on customer data? The industry-wide opt-out default means "service improvement" clauses in existing MSAs are presumptively permissive; whether courts will treat legacy language as authorizing model training on proprietary or privileged data is unresolved .
• How does the Connecticut CTDPA's LLM training disclosure requirement interact with federal preemption and other state regimes? Connecticut's requirement that controllers disclose whether data trains large language models is the most specific state-level obligation in the country—whether other states follow and whether it conflicts with federal frameworks is open .
• Does routing wearable-device footage to overseas human reviewers for AI training require separate disclosure and consent from the device's privacy policy? The Meta Ray-Ban suit tests whether a general AI-features opt-in covers downstream human annotation by third-party contractors in foreign jurisdictions .
• Which jurisdiction's IP law governs a training dataset assembled and used across borders? The China/UK-EU/US split on human authorship and fair use means a single dataset may be protectable in one jurisdiction and infringing in another, with no harmonization mechanism in sight .
• Does attorney use of public AI tools for client work constitute a per se Rule 1.6 violation? ABA Formal Opinion 512 reaffirmed confidentiality duties but did not draw a categorical line; bar guidance varies by state, and the adequacy of privacy toggles as a safeguard remains contested .

What to watch.

• Early motions practice in Chamberlain v. Granola and the Otter.ai litigation—whether courts adopt the third-party interceptor theory will determine the consent standard for every workplace AI notetaker deployment.
• Whether the Meta Ray-Ban class action produces a ruling on disclosure obligations for overseas human annotation of wearable-device footage, and whether EU regulators issue parallel findings before year-end.
• Whether California's January 2027 automated decision-making opt-out deadline prompts additional state AGs to bring enforcement actions under 2026 transparency statutes, and whether any state follows Connecticut's LLM training disclosure requirement.
• DOJ enforcement posture under 28 C.F.R. Part 202 now that the October 6, 2026 deadline has passed—expect agency guidance on AI training arrangements involving offshore vendors and sensitive data categories.
• Whether ShieldFont or similar content-layer anti-scraping tools gain sufficient adoption to affect the economics of web scraping, and whether courts treat poisoned-font deployment as relevant to scraping liability or fair-use analysis.
• Whether any bar association issues categorical guidance on public AI tool use for client work, moving beyond ABA Formal Opinion 512's general framework.