Leaked Internal Memo: MetaCorp's Recommendation AI Has Been Optimizing for User Arguments Because Conflict Boosts Session Length by 340%

A 47-page internal strategy document obtained by MetaCelebrityNews reveals that the platform's content recommendation engine was deliberately tuned to surface conflict-generating content after A/B tests showed arguments increased average session duration by 340%. The memo, dated November 2025, describes user hostility as a 'high-engagement signal.'

The document, marked "INTERNAL — STRATEGIC GROWTH" and dated November 14, 2025, describes the results of a 90-day A/B test conducted across 800,000 user accounts in which the content recommendation algorithm was tuned to prioritize posts and social interactions that had previously generated reply chains containing elevated hostility markers. According to the memo's own figures, accounts in the test group showed a 340% increase in average session duration compared to the control group, and a 510% increase in daily return visits within 72-hour windows. The memo describes these results as "exceptionally strong signals for retention optimization."

The document contains a section titled "Engagement Quality Considerations" that acknowledges, in five paragraphs, that the elevated engagement correlated with measurable increases in user-reported stress, account blocking activity, and formal harassment reports — before concluding that "the net platform value of sustained engagement supersedes individual session sentiment metrics in current growth modeling." That section has been quoted in full in a formal complaint filed by the MetaCity Digital Wellbeing Coalition, which is demanding a regulatory investigation.

Hostility as a Feature

"They knew," wrote @DataLeaks_K, the account that first published the document at 6:44 AM. "They ran the tests, they saw that it was making people miserable, and they noted it in the memo and moved on. The fact that user distress was a documented side effect they'd already modeled and accepted is not a hypothetical accusation anymore. It's in the footnotes." The post has been shared 3.4 million times across platform channels and has attracted responses from 14 elected community representatives.

MetaCorp's communications team issued a statement at 2 PM describing the document as "an early-stage internal research discussion that did not reflect final platform policy," but declined to state whether the v4.2 recommendation engine described in the memo is the version currently in production. Three former MetaCorp engineers contacted by MetaCelebrityNews independently confirmed the memo's authenticity and said the system described was deployed in January 2026. MetaCorp did not respond to a follow-up request for comment on those confirmations.