Research Report

State of AI Failures 2025

A data-driven analysis of 472 documented AI incidents. Financial impact, litigation trends, severity patterns, and the regulatory landscape shaping AI risk management.

Published April 2026·Provyn Research·1,079 source citations
472
Documented Incidents
$90.2B
Estimated Financial Impact
35.8%
Litigation Rate
179
Regulatory Actions

Executive Summary

AI failures are not decreasing — they are diversifying. Our analysis of 472 verified incidents across 16 industries reveals a risk landscape that is broadening faster than organizations can adapt.

Bias is the most common failure category (117 incidents), while Safety Failure drives the highest financial impact ($35.3B). The Technology sector is most affected (167 incidents).

35.8% of documented incidents have resulted in litigation, and 179 have triggered formal regulatory action. With the EU AI Act enforcement beginning in 2025 and expanding through 2026, the regulatory risk exposure for AI-deploying organizations is increasing materially.

Finding 1: Incident Volume Has Accelerated

Documented AI failures increased 229% from 2022 to 2023, driven by the mainstream deployment of large language models. While 2024 and 2025 show some normalization, the absolute volume remains far above pre-2023 levels.

Incidents Over Time

Documented AI failures by year reported

Finding 2: High-Severity Incidents Are Increasing

The proportion of high and critical-severity incidents has increased over time, reflecting AI deployment in higher-stakes domains including healthcare, finance, and government.

Severity Trends

How incident severity is evolving over time

Severity Distribution

Breakdown of incident severity levels

critical
10121%
high
26155%
major
20%
medium
10723%
low
10%

Finding 3: Financial Impact Concentrates in Few Categories

Total documented financial impact exceeds $90.2B. The top cost categories are disproportionately concentrated — the highest-cost category alone accounts for more financial impact than the bottom five combined.

Financial Impact by Category

Estimated total cost of incidents per category

Finding 4: Litigation Risk Varies by Category

169 incidents (35.8%) have resulted in litigation. Litigation rates vary significantly across categories, with copyright violations and bias-related incidents seeing the highest rates of legal action.

Bias52/117 (44%)
Safety Failure32/103 (31%)
Hallucination8/37 (22%)
Deepfake / Fraud13/35 (37%)
Privacy Leak11/33 (33%)
Other5/32 (16%)
Agent Error6/23 (26%)
Financial Error16/20 (80%)

Finding 5: Industry Exposure Is Uneven

AI incidents are concentrated in a small number of industries. The top three sectors account for the majority of documented failures, reflecting where AI deployment is most advanced and where public scrutiny is highest.

Technology167 incidents
Government74 incidents
Media48 incidents
Healthcare45 incidents
Finance34 incidents
Education30 incidents
Other27 incidents
HR / Recruiting19 incidents

Regulatory Landscape

EU AI Act— Enforcement began February 2025 (prohibited practices). Full high-risk system compliance required by August 2026. Organizations deploying AI in the EU must conduct fundamental rights impact assessments and maintain comprehensive documentation.
ISO/IEC 42001— Published 2023 — the first international standard for AI management systems. Increasingly referenced in procurement requirements and regulatory guidance. Certification demonstrates formal AI governance maturity.
NIST AI RMF— Voluntary framework widely adopted in US enterprise. Executive Order 14110 (October 2023) directed federal agencies to use the framework, expanding its influence across government contractors and regulated industries.
State-Level AI Laws— Colorado, Illinois, New York City, and other jurisdictions have enacted AI-specific requirements, particularly for hiring tools and automated decision-making. The patchwork is accelerating.

Methodology

This report is based on 472 incidents in the Provyn Index, each documented with 1,079 total source citations from news reports, court filings, regulatory actions, academic papers, and company statements.

Incidents are identified through systematic monitoring of public reporting and regulatory databases. Each entry is structured using AI-assisted research followed by a Chain-of-Verification quality pass that cross-checks claims, validates sources, and flags inconsistencies. Financial impact figures reflect documented losses, settlements, fines, and credible third-party estimates.

For a complete description of our data collection and verification methodology, see the Methodology page.

Download This Report as PDF

Enter your email to receive a formatted PDF of the State of AI Failures report.