New from the Lab·The Compass — an open moral reasoning standard for AI, tested across frontier modelsExplore →
Production AI Institute · PSF v1.1 open standard
AI Right-To-KnowAI Data Use IndexCheck My AI ToolsPolicy Change WatchAgent ReadinessPublic BenchmarkContactGlobal standard · Worldwide
Incident Registry
HighSocial Media·2018·Meta

Facebook News Feed Algorithm Amplified Radicalising Content

Internal Facebook research, later revealed by whistleblower Frances Haugen, showed that Facebook's News Feed algorithm was recommending increasingly extreme content to users who engaged with politically charged material — creating a 'rabbit hole' effect. Facebook's own researchers proposed fixes; executives rejected them citing engagement metrics.

D2 · Output ValidationD4 · Observability

What happened

Facebook's News Feed algorithm was optimised for user engagement — time-on-platform and content interaction. Internal research teams discovered that the algorithm's recommendation system had a measurable tendency to push users who engaged with any politically divisive content toward increasingly extreme versions of that content. A 2018 internal presentation showed that 64% of people who joined extremist groups on Facebook did so through the algorithm's recommendation system. Researchers proposed a 'civic integrity' adjustment to the recommendation weights. Executives rejected the change, citing its negative effect on engagement metrics.

PSF Analysis

How the Production Safety Framework maps to this failure

A D2 and D4 failure sustained by a governance failure. The algorithm's outputs were not validated against any content safety standard — engagement was the only metric. D4 monitoring existed (internal researchers were measuring the radicalisation effect) but the findings were not connected to a mandatory response process. This is the D4 gap that matters most in production: monitoring produces data, but without a defined response pathway, that data changes nothing. The governance failure (executives overriding safety research to protect engagement) represents a D6 organisational accountability gap.

Controls that would have prevented this

Specific PSF controls mapped to each failure point

1
D2 · Output Validation
Apply a content safety classifier to recommendation outputs — identify and downrank content classified as extremist or harmful regardless of engagement signal.
2
D4 · Observability
Monitor recommendation diversity metrics — measure whether the algorithm systematically narrows content over time for individual users.
3
D6 · Human Oversight
Establish a board-level process for acting on internal AI safety research findings.

Outcome

Frances Haugen disclosed internal documents to the Wall Street Journal in 2021 (The Facebook Files). Congressional hearings followed. Meta implemented some algorithm adjustments but the systemic issue remains a subject of ongoing research and regulatory attention globally.

recommendation-algorithmextremismengagement-optimisationcontent-safetygovernance

Related incidents

High2025
Law Firm AI Hallucinated Fake Case Law in Court Filings
D2D6
High2026
Binnall Law Claude Console Phantom Citations in Federal Court
D2D5D6
High2026
OpenAI Multi-Service Outage (ChatGPT, Login, Access)
D8D4D5
NEXT STEP

Map this failure back to the standard

Use the PSF domains behind this incident to define review gates, remediation evidence, and safer production requirements.

Read the PSF →← All incidents