Israeli Government Contractor Built “Think Tank” Apparently Designed to Influence AI Answers

01Israeli Government Contractor Built “Think Tank” With More Than 100 Reports Allegedly Designed to Influence AI Answers

A website presenting itself as an American think tank has published more than 100 reports on Israel and Palestine in a format that researchers say appears designed to influence search results and AI chatbot answers. The Hanover Institute for Public Policy claims to study the causes of antisemitism in the United States, but its reports carry no bylines and its website discloses that it was created by Piro, Inc. on behalf of the Israeli Government Advertising Agency.

Piro, a firm co-founded by film producer Daniel Rosenberg, received $900,000 from the Israeli government through Havas Media, a French public-relations conglomerate. Hanover began publishing on August 6 and has since produced reports addressing questions a user might put directly to a chatbot, including what displaced Palestinians in 1948 and whether Israel is deliberately starving Gaza.

The site imitates the conventions of a conventional American research institution: a generic name, red-white-and-blue design, neutral-sounding prose, tables of contents, footnotes, statistics and extensive citations. NewsGuard analyst Alice Lee said this packaging appears intended for Americans researching the conflict through search engines or chatbots, which tend to favor concrete data and strongly sourced material.

Piro’s own marketing supplies the clearest evidence that AI influence is a goal. Its website advertises “AI Story Optimization” and says the company writes content engineered around how large language models assess credibility. Rosenberg separately wrote on LinkedIn that Piro had spent months reverse-engineering how services including ChatGPT, Gemini and Perplexity construct answers, asking how to ensure that AI “knows how to tell” a client’s stories.

The reports are not uniformly aligned with Israeli government positions. One called the theory that foreign university funding explains antisemitic incidents weak and riddled with exceptions, despite Prime Minister Benjamin Netanyahu having advanced that argument. Other reports question claims about destruction in Gaza or repeatedly connect their subjects to rising antisemitism, sometimes citing Israeli government bodies alongside material Hanover describes as peer-reviewed academic research.

Responsible Statecraft tested 12 randomly selected reports with GPTZero. The detector classified 11 as AI-written with high confidence and one with moderate confidence, but those results alone do not establish how the articles were produced.

Piro’s agreement filed with the US Justice Department does not explicitly say its assignment is to influence AI. Rosenberg has described the work as placing accurate, sourced facts in the public record and countering misinformation. Nor is there evidence in the source confirming that Hanover’s reports have been indexed, cited by major chatbots or changed any model’s answers.

02Independent AI-Use Observatory Aggregates 24,500 Conversations: Anthropic Criteria Would Filter Nearly Half

The AI Observatory, a new public research platform, has assembled 24,521 real conversations with generative AI systems to provide an independent view of how people use them. The project involves researchers from MIT, Stanford, the Data Provenance Initiative, and other institutions, and draws on conversations shared with user consent through seven existing datasets.

The platform is intended to complement reports from AI companies, whose analyses reflect particular products and research questions. One widely cited example, the Anthropic Economic Index, primarily examines work and productivity uses of Claude and filters out conversations unrelated to that focus. When AI Observatory researchers applied Anthropic’s methodology to their own dataset, 48% of its conversations would have been excluded.

That filtering substantially changed the resulting picture. Among the non-work conversations, health and relationship topics appeared at a rate of 44.2%, compared with 31.2% in Anthropic’s analysis. Adult or illicit topics accounted for 7.9% versus 2.1%; harassment and hate for 27.5% versus 5.66%; and sexual content for 16.7% versus 2.4%. These figures compare category rates within the respective analyses and do not represent each category’s share of all AI conversations.

The Observatory covers 85,633 conversational turns—each comprising a user prompt and an AI response—from about 5,000 users interacting with 52 models between 2023 and 2025. Its analysis also found that usage varied by product: Anthropic systems were more often used for coding, Gemini for social interactions and role-playing, and ChatGPT for homework assistance.

The researchers are not claiming that company reports are incorrect. Anthropic told MIT Technology Review that its published research reflects the specific questions and interests of its research teams, while also supporting independent external research. The contrasting results instead show how a study’s scope and exclusions determine which uses remain visible.

The Observatory has important limits of its own. Its sample is far smaller than company datasets, and voluntarily shared conversations may underrepresent sensitive behavior because users could be less willing to disclose it. The researchers therefore caution that the results cannot establish how common these uses are across all AI users.

03Cursor Launches Origin Code Hosting: Paid Users Can Create Repositories and Sync Pull Requests Both Ways With GitHub

Cursor has launched Origin, an early-beta code-hosting service that places repositories, pull-request reviews and coding agents inside the same product. Origin is available through Cursor’s new Codebase tab and is beginning to roll out to users on all paid plans, although enterprise administrators can opt their organizations out.

The launch gives developers two distinct ways to bring code into Origin. They can create a repository directly in Cursor, then use its command-line interface to clone it or push an existing local project. In that case, Cursor hosts the code. Alternatively, users can connect a GitHub organization, select repositories to synchronize and place Origin’s copies alongside Cursor-hosted projects.

That distinction determines where changes ultimately go. GitHub repositories synchronized into Origin update in real time and can be browsed, searched and pulled from Cursor. Developers can disconnect them at any time, but pushes continue going to GitHub, which remains the source of truth for projects that started there. Origin displays icons identifying whether each repository is hosted by Cursor or came from GitHub.

Origin also provides a pull-request interface for every repository. Developers can view a PR’s timeline, commits, checks and changed files, review differences, leave comments and merge changes. For synchronized GitHub repositories, PR discussions work in both directions: comments posted in Cursor appear on GitHub, while reactions and replies made on GitHub appear in Cursor within seconds. A developer assigned a GitHub review can complete and merge it from Cursor.

Cursor’s coding agents enter that workflow from the code or PR being viewed. Users can ask questions about the code, request changes, update a pull request or have Cursor push a branch. Available integrations include Vercel for PR preview deployments and production releases, plus Depot and Buildkite for continuous integration. Both CI services can run existing GitHub Actions workflows, while Buildkite can also run its own pipelines.

Cursor says more “agent-native” capabilities are coming, but has not disclosed what they will do or when they will arrive. It also has not detailed broader pricing, migration tools for independently hosted repositories, data-residency options or enterprise security arrangements.