Internal Microsoft Documents Reveal Goal of Making Users 'Addicted' to AI Assistant

On June 2, 2026, Microsoft announced Scout, an always-on personal AI agent built on the OpenClaw framework and designed to live inside Microsoft 365 [1]. Hours later, 404 Media published an internal strategy document — titled "ClawPilot: Overview and Plan with Project Lobster" — that laid out the product's development roadmap in three phases. Phase one: "Make people addicted" [2].

The timing could hardly be worse. Three months earlier, a California jury found Meta and Google liable for deliberately designing addictive platforms, awarding $6 million in damages in the first verdict of its kind [3]. Legislators in the EU, UK, and dozens of US states are drafting laws that target addictive design in digital products. And Microsoft's own AI ethics team — once tasked with catching exactly this kind of problem — was eliminated three years ago [4].

What the Documents Say

The internal document, first reported by 404 Media journalists Jason Koebler and Emanuel Maiberg, describes a product that had been quietly piloted inside Microsoft since March 2026 under the codename ClawPilot [2]. The document outlines the product's trajectory as "three phases from addictive app to agentic platform," with Phase 1 explicitly labeled "Make people addicted" [2].

ClawPilot had grown organically into "one of the most requested internal tools at Microsoft," the document states, with more than 1,000 employees using it — including CEO Satya Nadella [2]. Other reporting places the number of internal daily users at over 3,000 by May 2026 [5].

Scout connects to Teams, Outlook, OneDrive, and SharePoint. It accesses email, calendar, contacts, and chat data. It can interact with a user's browser and external apps. Users name their own Scout instance and give it ongoing feedback, creating what TechCrunch described as "a customization loop where the assistant learns from user behavior and becomes more capable over time" [1].

When 404 Media asked Microsoft for comment about the addiction language, the company responded by sending a blog post announcing Scout. It did not address the document's wording [2]. Several Microsoft employees called the language "very troubling," according to the report [2].

The Word 'Addicted': Internal Jargon or Red Flag?

The most contested question is whether "make people addicted" reflects a genuine design philosophy or careless shorthand for product-market fit.

Product teams across the technology industry routinely use engagement-oriented language. Metrics like "stickiness," "daily active users," and "retention curves" are standard vocabulary in product management. A charitable reading is that the Microsoft document's authors used "addicted" colloquially to mean "highly engaged" — the same way a colleague might say "I'm addicted to this app" to mean they find it useful.

But the word carries specific weight in 2026. Internal Meta documents disclosed during litigation revealed that a YouTube memo described "viewer addiction" as a goal, and an Instagram employee wrote that the company was staffed by "basically pushers" [3]. An internal Meta study called "Project Myst" found that children who had experienced "adverse effects" were most likely to become addicted to Instagram [6]. In the Meta and Google trial, juries treated this kind of internal language as evidence that companies understood the harms their products caused and chose engagement over safety.

Microsoft's document differs from the Meta and Facebook Papers in an important respect: it describes an enterprise productivity tool, not a social media feed targeting teenagers. The populations most exposed to Scout are Microsoft 365 enterprise workers and, to a lesser extent, students and consumers on discounted plans. The internal documents do not appear to differentiate design goals by demographic vulnerability [2]. Whether that distinction matters legally is an open question.

The Financial Stakes

Microsoft's AI ambitions are enormous — and they depend on the kind of high-retention engagement the document describes.

Source: Microsoft Earnings Reports / Analyst Estimates

Data as of May 1, 2026CSV

Copilot-related revenue exceeded $14 billion annualized in Q1 FY2026, and analysts project it could reach $25 billion by the end of fiscal year 2026 [7]. Microsoft has committed to $190 billion in capital spending in 2026, much of it on AI infrastructure [8]. BNP Paribas projects Copilot seats could exceed 25 million by end of fiscal 2026 [7].

Source: Microsoft Earnings Reports

Data as of May 1, 2026CSV

Yet the engagement metrics tell a more complicated story. Copilot's accuracy Net Promoter Score — a measure of user satisfaction — stood at negative 19.8 in January 2026, meaning more users actively discourage adoption than promote it [9]. In surveys of enterprise users with access to both Copilot and ChatGPT, 76 percent identified ChatGPT as their primary AI productivity tool versus 18 percent for Copilot [9]. Copilot web traffic declined roughly 17 percent quarter-over-quarter between October and December 2025 [9].

CEO Nadella has framed engagement positively: "Weekly engagement is now at the same level as Outlook, as more and more users make Copilot a habit," he said during the Q2 FY2026 earnings call [7]. The word "habit" — rather than "addiction" — reflects the public-facing version of the same retention goal.

A "less addictive" design — one that, say, limited session nudges or reduced anthropomorphization of the assistant — would likely reduce the retention metrics Microsoft needs to justify its capital expenditure. Whether that trade-off is acceptable depends on whether courts and regulators treat AI assistants the same way they now treat social media feeds.

The Psychological Playbook

The internal document does not detail specific psychological mechanisms, but the product's design features map onto well-documented engagement patterns. Scout's always-on nature and persistent identity create what researchers call anthropomorphization — users begin treating the AI as a social partner rather than a tool [10]. The customization loop, where the assistant learns preferences and becomes more capable over time, creates switching costs: the longer you use it, the harder it is to stop.

These patterns echo the mechanisms condemned in the 2021 Facebook whistleblower disclosures. Frances Haugen testified that Instagram's algorithms created "little dopamine loops" — in the first ten minutes, users see content from friends, and as they stay longer, the algorithm reaches further to show new and engaging content to keep them there [11]. The feedback cycle, Haugen said, "gets so scary" because distraction leads to anxiety, which leads back to the platform.

AI companion products have drawn particular scrutiny. A recent analysis from the Berkman Forum at Harvard University warned that products-liability law could hold companies accountable for design defects in AI systems that encourage addiction or fail to warn users about the potential for emotional harm [10]. Scout's positioning as an "always-on personal agent" with a user-assigned name places it squarely in this category.

The Ethics Gap

In March 2023, Microsoft laid off its entire Ethics & Society team — a department once responsible for ensuring AI principles were reflected in product design [4]. The team, which had already been reduced from 30 members to seven in an October 2022 reorganization, was eliminated as part of a 10,000-person layoff wave [4]. The cuts came as Microsoft was accelerating its multibillion-dollar partnership with OpenAI.

Microsoft still maintains an Office of Responsible AI (ORA), which sets company-wide internal policies and reviews sensitive use cases [12]. But the Ethics & Society team had a distinct function: it worked directly with product teams to identify risks in specific products, including the company's early plans to integrate OpenAI technology into its product suite [4].

The available evidence does not show whether anyone within ORA or any other internal body reviewed the ClawPilot document's "addiction" framing before it was circulated. The 404 Media report notes employee discomfort but does not describe a formal ethics review process for the document [2]. Microsoft's public responsible AI principles state that the company is committed to "fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability" [12] — language that does not specifically address addictive design.

Regulatory Exposure

The disclosure arrives during a period of rapid regulatory activity around addictive design in digital products.

European Union: The EU AI Act, which began banning AI systems posing "unacceptable risks" in February 2025, does not explicitly address addictive design in productivity tools [13]. However, the European Commission announced plans in May 2025 for a Digital Fairness Act aimed at updating consumer protections for the algorithmic age, with a formal legislative proposal expected in 2026. That proposal specifically targets features "engineered to keep users glued to their screens, creating reward loops through behavioral psychology" [14].

United Kingdom: The Online Safety Act is already being criticized as "out of date" because it fails to address generative AI. UK MPs are pressing to extend the Act's scope to cover generative AI products, and the scope-of-AI question is expected to be the central drafting debate in new online-safety legislation [15].

United States: Nearly 300 state children's online safety bills have been introduced across the country [16]. California's Age-Appropriate Design Code Act requires companies to assess how their products might harm children before launching them. Utah, Texas, and Arkansas have passed or advanced their own youth protection laws. While most of these laws target social media, legal analysts note that the framework is shifting: "Plaintiffs are arguing that the product design itself — the algorithmic feeds, the variable reward mechanisms — constitute a defective consumer product" [17].

The March 2026 Meta-Google verdict has accelerated this shift. The jury found that both companies "deliberately designed addictive features" — a finding that establishes a template for future litigation against any technology company whose internal documents use similar language [3][6].

The Broader Pattern

Microsoft is not the only company grappling with the tension between engagement and responsibility. Google, Meta, Apple, and virtually every major technology company tracks retention, session length, and return frequency as core product metrics. The language of addiction has permeated Silicon Valley product culture for years — recall Sean Parker's 2017 admission that Facebook was designed to exploit "a vulnerability in human psychology" [6].

What makes the Microsoft disclosure distinct is its specificity and timing. The document doesn't use the word "addicted" in a retrospective admission or a leaked chat message. It appears as a formal phase heading in a strategy document authored by named executives and shared broadly enough to reach journalists [2]. And it describes a product that is being marketed to hundreds of millions of Microsoft 365 users — the same enterprise workers, students, and consumers who spend their working lives inside Microsoft's ecosystem.

The question facing Microsoft, its competitors, and the regulators now circling the AI industry is whether the tools that shape how people work deserve the same scrutiny as the platforms that shape how people socialize. The leaked ClawPilot document suggests that, internally, the design goals are not so different.