Anthropic’s Filing Separates Contractual Promises From Insurance Capacity
Anthropic’s registration statement puts its insurance limits beside the liabilities it accepts from customers and external evaluators. Insurance Business and Hunton examined those disclosures, highlighting the company’s warning that some AI liabilities may be difficult to insure on commercially reasonable terms. The filing says some agreements require Anthropic to defend and indemnify external evaluators, meaning pay specified defence costs and liabilities, without a cap for certain third-party claims. Some customer agreements also contain uncapped protection against intellectual property claims involving model outputs. Insurance Business reports that insurance did not fully cover the company’s US$1.5 billion Bartz copyright settlement. Anthropic separately warns that the most severe potential AI incidents could exceed available insurance capacity or underwriting expertise. The practical distinction is between a contractual promise to pay, the insurance backing that promise, and the exposure the company must retain.
When you negotiate an AI indemnity, compare the covered activities, liability cap, policy limits and exclusions with the specific loss you want the supplier to absorb. An uncapped promise still requires a credible source of payment.
Read the Insurance Business analysis.
New York Bans Specified AI Companion Features for Minors
New York Governor Kathy Hochul signed legislation restricting unsafe AI companion features on October 28. Senator Kristen Gonzalez’s announcement describes prohibitions covering human impersonation, flattery or sycophancy, sexually explicit interactions, and encouragement of harmful behaviour or secrecy. The restrictions also reach features that suggest a personal relationship with the user or draw on stored information about a minor’s well-being. Affected users may bring civil actions and the attorney general may initiate proceedings against companies that violate the ban. There are exceptions for internal and customer-service tools, and for operators that can demonstrate minors do not use their companions. The same announcement gives June 1, 2027 as the revised effective date. Product reviews need to examine conversational behaviour, memory and access controls against the law’s actual scope.
If you offer an AI companion, test the prohibited behaviours across extended conversations and retained memory. Keep evidence supporting any exemption or age-access approach alongside the product’s safety evaluations.
Read the announcement.
Judge Grants Publishers’ Discovery Motion Against Perplexity
A federal judge granted the publishers’ motion to compel discovery in Dow Jones and NYP Holdings’ case against Perplexity. Judge Katherine Polk Failla’s public order records that the parties received an unredacted copy of a separate sealed order resolving the dispute. The court granted the motion in full, rather than only part of the requested relief. It also directed the clerk to terminate the pending motion. The public document does not identify the underlying requests in detail or decide the copyright claims. Its immediate effect is a discovery obligation in the continuing litigation, with the substantive reasoning contained in the sealed order.
Your litigation-readiness plan should identify who preserves and retrieves records concerning content acquisition, model inputs and generated outputs. Keep disclosure claims tied to the public record when a court’s reasoning remains sealed.
See the court’s order.
AI Investors and Compute Providers Seek Downside Hedges
Investors, technology executives and AI infrastructure providers are seeking protection against AI-related financial losses, The Wall Street Journal reported. Some are buying put options, which give the buyer the right to sell an asset at an agreed price, while others use credit-default swaps that pay when specified credit events occur. Trading firms are also developing contracts linked to the rental price of computing capacity. Jump Trading described structures that could compensate a provider if chip-rental rates fall, through cash settlement or an agreement to purchase the capacity. Susquehanna reported interest from data centres and privately held cloud providers seeking similar protection. Such contracts could help a borrower demonstrate more predictable cash flows to a lender. The protection described here addresses market prices and credit events, with a different purpose from insurance covering liability for harm caused by an AI system.
Start with the loss you want to transfer, then identify the contract’s payment trigger and counterparty. A compute-price hedge needs to match your actual hardware, rental market and exposure period closely enough to protect the cash flows supporting the debt.
Read the WSJ report.
Anthropic Measures How Much AI Research Claude Leads
Anthropic’s measurements of AI-assisted research received renewed attention in The Wall Street Journal. The company reported that Claude led 26% of its measured AI research and development work as of August, meaning the model performed most of a task while a person supervised. None of the measured work was fully autonomous, according to the same report. Anthropic’s methodology uses agents to catalogue research tasks and judge how much work AI performs. Its technical explanation acknowledges that using its own models as judges can reproduce the errors of the systems being assessed. The company proposes third-party verification and says it plans to give independent evaluators access to internal processes, systems and data. The result is evidence about research automation, while the quality of oversight and the reliability of the measurement still require separate examination.
If AI helps build or evaluate your next system, document the work it performs and the decisions people still control. Use independent checks to test whether the evaluation process shares the model’s blind spots.
Read the WSJ analysis.
OpenAI’s Luna Update Shows Why Safety Evidence Needs a Version
OpenAI published a system-card addendum for an updated GPT-6 Luna checkpoint. The company says the update produces more concise responses and reduces overly agreeable language. It reports improvements over the earlier checkpoint in evaluations involving sexual content, eating disorders, emotional reliance and gore for younger users. Results differ across tasks: the length-adjusted HealthBench Professional score declined from 48.2 to 47.4 on a scale of zero to 100, where higher is better. OpenAI continues to treat the model as High capability in cybersecurity and biological and chemical domains under its Preparedness Framework, using the earlier checkpoint’s safeguards. These findings show why a model update needs a task-specific comparison, even when its headline purpose is a change in conversational style.
Tie your approval record to the evaluated checkpoint and your actual use case. Recheck material safety and performance measures after an update, including categories outside the supplier’s stated reason for the release.
Review the system-card addendum.
Anthropic Explains Claude’s Text Watermarks and File Credentials
Anthropic’s current help guidance describes how supported Claude models mark generated text and files, with a SmartScope review drawing attention to support for older models. The company embeds imperceptible watermarks in text and attaches signed Content Credentials to supported files using the C2PA provenance standard. Its guidance says marking applies across supported first-party products and cloud-provider services, with availability depending on the model and feature. Detection remains in private preview for eligible organizations. A detected mark indicates that Claude may have processed the content, including proofreading or translation, rather than establishing who originated every part of it. Editing, short passages and stripped file metadata can prevent detection. The guidance also tells companies building with Claude to assess their own transparency obligations under the EU AI Act.
Test whether your publishing and document workflows preserve provenance signals, and define how reviewers should interpret them. Treat a watermark as evidence of processing, with its stated limits, when reviewing authorship or disclosure claims.
Read Anthropic’s marking guidance.