Epistemic Harms

Epistemic harms are harmful end states where AI has damaged humanity’s ability to know things, verify claims, and make good collective decisions.

How These Risks Connect

increases

These are end-state harms to humanity’s epistemic capacity:

Risk	Description
Knowledge Monopoly	A few AI systems become sole arbiters of “truth”
Institutional Capture	AI advisors subtly bias all organizational decisions
Reality Fragmentation	Populations live in incompatible information environments
Epistemic Collapse	Society loses ability to establish shared facts
Trust Cascade Failure	Once trust breaks, no mechanism exists to rebuild it
Learned Helplessness	People give up trying to know anything; epistemic nihilism

Risk	Description
Legal Evidence Crisis	Courts can no longer trust digital evidence
Cyber Psychosis	AI-induced psychological harm from reality distortion

These outcome-level harms are driven by pathways and amplifiers (also in this section):

Factor	How It Contributes
Authentication Collapse	Verification fails → can’t distinguish real from fake
Scientific Corruption	Fake papers proliferate → knowledge base corrupted
Expertise Atrophy	Humans lose ability to evaluate AI outputs
Sycophancy at Scale	AI confirms biases → no reality check
Preference Manipulation	AI shapes what people want, not just believe
Consensus Manufacturing	Fake agreement masks actual disagreement
Historical Revisionism	Past becomes contested → shared history lost
Automation Bias	Over-reliance on AI recommendations
Trust Erosion	Gradual decline enables sudden collapse

This section includes risks at three causal levels:

Outcomes (what we ultimately want to avoid):

Pathways (mechanisms leading to outcomes):

Amplifiers (conditions that increase risk):

The distinction matters because intervening on amplifiers and pathways can prevent multiple downstream outcomes.

Previous information problems (propaganda, fake news) were limited by human capacity. AI changes the game:

Old Problem	AI Escalation
Propaganda exists	Personalized propaganda for each individual
Fake news spreads	Generated faster than verification
Evidence can be faked	All digital evidence becomes deniable
Experts can be wrong	Expertise itself atrophies
Institutions can be corrupted	AI advisors capture decisions invisibly

Risk	Severity	Reversibility
Knowledge Monopoly	High	Difficult (infrastructure lock-in)
Institutional Capture	High	Moderate (requires institutional reform)
Reality Fragmentation	High	Difficult (no shared ground to rebuild)
Epistemic Collapse	Catastrophic	Very difficult
Trust Cascade	High	Very difficult (no trusted rebuilder)
Learned Helplessness	High	Generational