NOTE TO EDITOR (delete before publishing): Load-bearing claim is the Kotaku/404 Media leak: that an internal Microsoft memo describes Copilot/Scout as wanting to be 'addictive' and ties this to Nadella. Please verify the 404 Media reporting directly and confirm the word 'addictive' and the Scout project name are accurately characterized before publishing. The Google/SpaceX $920M/month figure and the Cloudflare 42.5% human traffic figure are drawn from the signals but should be sanity-checked against primary reporting. Anthropic 'pause' framing is interpretive and should be flagged as analysis, not Anthropic's stated position.
The word that gives the game away
The most useful document in the AI industry this week is not a benchmark, a model card, or a $30 billion compute deal. It is a single adjective, reportedly lifted from an internal Microsoft strategy memo and surfaced by 404 Media: the company wants its AI to be addictive.
The leak, covered by Kotaku under the working headline "Leak Reveals Microsoft Wants Its AI to Be 'Addictive'", names an internal initiative called Scout and ties the framing to Satya Nadella's office. The detail to hold onto is not the lurid one. It is the structural one. A company that spent the last two years selling Copilot as a productivity tool, an enterprise tool, a serious tool for serious work, is internally measuring it on the same axis as a mobile game.
That is not a scandal. It is a confession about the unit economics.
What "addictive" actually means on a P&L
When a consumer software business says it wants engagement, it usually means one of two things. Either the product is free and the engagement is the inventory it sells to advertisers, or the product is paid and the engagement is what justifies the renewal. Microsoft Copilot, in its various incarnations, is both. It is bundled into Microsoft 365 seats that finance departments are now being asked to upgrade. It is embedded into Windows, where Microsoft's relationship with the user is rented from the OEM. And it is increasingly offered as a standalone consumer surface that competes with ChatGPT.
In all three configurations, the business problem is the same. Inference is expensive. A user who opens Copilot once a week and asks it to summarize an email is a loss-making user. A user who treats it as a default companion, a thinking surface, a habit, is the only kind of user the current cost structure can tolerate at scale.
This is the part the industry's public messaging has been careful not to say out loud. The race is not really about capability. On most benchmarks the frontier models are within striking distance of each other, and the gap closes again every six weeks. The race is about getting users to form a dependency before the novelty wears off and the bills come due.
Viewed through that lens, the Scout memo is not a deviation from Microsoft's strategy. It is the strategy, written down in language that was never meant to leave the building.
The consumer playbook arrives in enterprise clothing
There is a reason this lands awkwardly. Microsoft has spent two decades positioning itself as the adult in the room. Office is not TikTok. Teams is not Instagram. The pitch to the CIO has always been that Microsoft sells tools that get work done, and that whatever Meta or ByteDance are doing to people's attention, Redmond is on the other side of that line.
The AI assistant breaks that distinction, and the leak makes the break visible. The same company selling Copilot to a procurement officer on the basis of productivity gains is, internally, measuring success in the vocabulary of consumer social products. Daily active users. Session length. Return frequency. The mechanics are not a secret. We have thirty years of literature on how to engineer them.
The more interesting question is whether "addictive" is even the correct goal for an assistant that is supposed to make you more capable. A search engine you use compulsively is a different product from a search engine you use well. A calculator optimized for return visits is a broken calculator. If the internal north star at Microsoft is engagement, the product team has been handed a brief that quietly conflicts with the marketing.
This is also where Anthropic's posture this week becomes legible. The company's call for a "temporary pause" on frontier development, reported by The Guardian and circulated widely on Mastodon, reads to some observers as a marketing exercise from a lab heading toward an IPO. That reading is not wrong, but it is incomplete. Anthropic is selling a different product to a different buyer: the enterprise that is nervous about exactly the kind of memo that just leaked from Microsoft. Safety positioning and engagement optimization are not opposing philosophies. They are two pricing strategies aimed at two segments of the same market.
The signal in the surrounding noise
If you scan the broader AI conversation this week, the Microsoft leak sits inside a pattern that is becoming hard to miss.
Google is reportedly paying SpaceX roughly $920 million a month for around 110,000 Nvidia GPUs through 2029, a deal first surfaced by Implicator. Cloudflare's leadership has noted that bots now account for more than half of website traffic, with humans down to about 42.5 percent. A coalition of US trade associations is asking the Trump administration to intervene in an AI-driven memory chip shortage that is raising prices in unrelated industries. OpenAI has rolled out a "Lockdown Mode" for prompt injection, an opt-in admission that the attack surface of these systems is now well-documented enough to warrant a hardening switch.
None of this is a story about capability. All of it is a story about cost, distribution, and dependency. The hyperscalers are spending at a rate that only makes sense if the resulting products become habits. The infrastructure is being built for a world in which hundreds of millions of people open an assistant the way they currently open a browser, and where the assistant is the browser, and where the assistant is also the operating system.
In that world, an internal memo describing the goal as "addictive" is not a slip. It is the only honest description of what the spreadsheet requires.
What to actually watch
The right thing to do with the Scout leak is not to be shocked by it. The right thing is to use it as a translation key for everything that follows.
When Microsoft announces a Copilot redesign in the coming months, the question is not whether the new features are useful. The question is which behaviors they are tuned to reinforce. Notifications that surface unprompted suggestions. Memory features that make starting a new conversation feel like a loss. Default placements that make the assistant the path of least resistance for tasks the user could perform without it. None of these are evil. All of them are familiar from a decade of consumer app design, and all of them produce engagement curves that look healthy on a dashboard while producing users who feel slightly worse about their own competence.
The other thing worth watching is the labor market reaction. The professional class that Imprompted writes for has spent two years being told that AI will augment their work. A product whose internal success metric is dependency is not augmentation. It is substitution with extra steps, where the user is gradually trained to route more cognition through a system they do not own and cannot audit. That is a different deal, and it is the deal that the leaked memo describes.
The scandal, if there is one, is not that Microsoft wrote this down. It is that the public framing of AI as a productivity story has held this long without anyone in the industry being forced to choose between the two vocabularies. The Scout memo forces the choice. Expect more leaks. Expect more denials. Expect, eventually, a quieter pivot in which the word "addictive" disappears from the documents and the design choices stay exactly where they are.
The footnote is the story. It always was.