When opposing counsel demands ten months of AI system logs and you're facing six-figure hosting bills, you have a critical eDiscovery decision: fight the request, negotiate proportionality, or demand cost-sharing. Judge Sarah L. Cave's ruling in Encyclopaedia Britannica, Inc. v. Perplexity AI offers a framework for this decision, one that legal operations teams will encounter more frequently as AI litigation grows.
The court ordered Perplexity to produce one additional RAG snapshot and six months of user activity logs, not the ten months requested. Importantly, it required plaintiffs to contribute $6,000 monthly toward hosting costs under Rule 26(c)(1)(B). This was proportionality in action.
The Decision You're Facing
Your organization receives a discovery request for AI system data covering an extended period. The data exists and is arguably relevant, but producing it requires engineering resources, cloud storage, and ongoing hosting fees. You must decide:
- Do you produce everything requested and absorb the full cost?
- Do you narrow the scope based on proportionality arguments?
- Do you seek cost-sharing from the requesting party?
Each path carries different risks, timelines, and resource requirements.
Key Factors That Affect Your Choice
Temporal relevance to the claims. Judge Cave anchored her analysis to the plaintiffs' copyright registration periods. Existing discovery from separate litigation already covered May 2024 through May 2025. Thirteen registrations became effective between April and December 2025, establishing clear temporal boundaries for additional discovery.
If the requesting party's claims map to specific time periods, regulatory compliance windows, contract terms, employment dates, you have objective grounds to limit scope. If their claims span your system's entire operational history, you're in a weaker position to restrict the timeline.
Engineering burden and incremental costs. Perplexity submitted declarations asserting hundreds of engineering hours and hundreds of thousands in monthly costs. The plaintiffs countered with estimates in the tens of thousands. Judge Cave didn't resolve the competing calculations but acknowledged that requiring Perplexity to produce any of the additional data, let alone ten months' worth, is not a small undertaking.
Document your burden with specificity. Vague assertions about complexity won't survive scrutiny. You need engineering estimates, vendor quotes, and infrastructure costs broken down by component.
Data type and inspection requirements. RAG snapshots required substantial engineering work. User activity logs had to be extracted from deep storage and hosted in a secure environment for inspection. Different data types impose different burdens.
If you're dealing with live databases that require snapshots, you're arguing production burden. If you're hosting extracted data for ongoing inspection, you're arguing hosting expense. These aren't the same cost-sharing arguments.
Existing discovery overlap. The court noted that prior production already covered part of the relevant period. This partial overlap strengthened the proportionality argument for limiting additional production.
If you've already produced substantial volumes covering adjacent timeframes, you can argue diminishing marginal relevance for incremental periods.
Path A: Absorb Full Cost and Produce Everything
Choose this path when:
- The requested data volume is modest relative to your infrastructure
- The temporal scope clearly maps to actionable claims
- Fighting the request would cost more in legal fees than production
- You need to preserve a cooperative discovery posture for strategic reasons
- Your internal estimates show manageable burden (not hundreds of engineering hours)
You maintain control over production methodology and avoid protracted discovery disputes. You also avoid judicial skepticism about inflated cost estimates. The court viewed Perplexity's "dramatically increased hosting estimates with considerable skepticism" when they diverged sharply from earlier quotes.
Don't choose this path simply because you lack documentation of burden. That's a failure of legal operations planning, not a reason to capitulate.
Path B: Negotiate Proportionality Limits
Choose this path when:
- You can identify clear temporal boundaries tied to the claims
- The full requested period extends beyond reasonably relevant timeframes
- You have objective evidence (registration dates, contract terms, regulatory windows) supporting a narrower scope
- Engineering burden is significant but you're willing to shoulder costs for a limited production
Judge Cave limited UAL data to August 2025 through January 2026, extending "a little over a month beyond the latest registration." She ordered one additional RAG snapshot rather than continuous coverage. This balanced "Plaintiffs' need for information to determine whether and to what extent Perplexity has infringed their copyrights and trademarks against the expense and burden on Perplexity."
Your argument requires:
- Specific temporal anchors
- Evidence that the narrower scope captures the relevant conduct
- Good-faith acknowledgment that some additional production is warranted
You're not refusing discovery. You're defining proportionate scope.
Path C: Demand Cost-Sharing Under Rule 26(c)(1)(B)
Choose this path when:
- Hosting or infrastructure costs are substantial and ongoing
- You can document these costs with vendor quotes or internal accounting
- The burden falls disproportionately on your organization due to the requesting party's litigation strategy
- Discovery is occurring late in the process or after other substantial productions
Rule 26(c)(1)(B) permits cost allocation for good cause, but the responding party bears the burden of justification. Judge Cave found good cause because "hosting the additional UAL data imposed significant expense and burden relatively late in document discovery."
The court didn't require full cost-sharing. It capped plaintiffs' contribution at $6,000 monthly for UAL hosting, while plaintiffs separately agreed to pay for RAG hosting. This granular allocation reflects different cost structures for different data types.
Your cost-sharing argument needs:
- Itemized hosting or infrastructure expenses
- Evidence that these costs exceed normal discovery burdens
- Demonstration that the requesting party's scope demands drive the expense
- Timing context (late requests carry more weight)
Don't conflate engineering labor with hosting fees. The court treated these differently. Labor to create snapshots remained Perplexity's burden. Ongoing hosting for inspection became shared.
Summary Matrix
| Factor | Absorb Full Cost | Negotiate Scope | Demand Cost-Sharing |
|---|---|---|---|
| Data volume | Modest, manageable | Significant but definable | Massive, ongoing |
| Temporal relevance | Clear alignment to claims | Partial overlap, some excess | Broad requests beyond core period |
| Engineering burden | Routine extraction | Substantial but one-time | Continuous hosting/access |
| Existing production | Limited prior discovery | Substantial overlap exists | Multiple prior productions |
| Cost documentation | Not required | Helpful for proportionality | Essential, itemized |
| Strategic posture | Cooperative, efficient resolution | Balanced, reasonable limits | Defensive, burden-focused |
| Primary risk | Unnecessary expense | Scope dispute prolongs discovery | Judicial skepticism of estimates |
The Perplexity ruling shows that courts will scrutinize both relevance and burden when AI systems generate voluminous data. Your choice isn't binary, Judge Cave granted partial relief on scope and partial cost-sharing. But you must make the choice deliberately, with documented justification for whichever path you take.
If you're waiting until you receive the discovery request to think about these questions, you're already behind.
