AI Safety Statistics 2026
By Axis Intelligence Research
Co-authors: Sarah Mitchell & Marcus Chen | Last updated: July 18, 2026 | License: CC BY 4.0
The AI Incident Database recorded 362 documented AI failures in 2025, up 55.4% from 233 in 2024, according to the Stanford HAI 2026 AI Index. In the same period, U.S. private AI investment reached $285.9 billion — while dedicated AI safety research spending remained under $100 million annually. Axis Intelligence Research calculates an AI Safety Gap Index (ASGI™) of 84.0 out of 100, indicating a critical structural imbalance between capability advancement and safety infrastructure.
Quick Answer
Documented AI incidents rose 55.4% in 2025 (362 incidents, per Stanford HAI/AIID). The FBI’s 2025 Internet Crime Report logged $893.3 million in losses from AI-enabled fraud — the first time in IC3’s 25-year history the category appeared separately. Hallucination rates across 26 top models range from 22% to 94% on a new accuracy benchmark (Stanford HAI 2026 AI Index). The EU AI Act’s high-risk AI system compliance deadline is August 2, 2026.
Key Findings
- Axis Intelligence Research finds the global AI safety gap reached a critical threshold in 2025: the ASGI™ (Axis AI Safety Gap Index) scored 84.0/100, computed from a 2,878:1 ratio of AI capability investment to safety research funding, 30.2 documented incidents per frontier safety framework published, and an estimated 68% of high-risk AI systems not yet EU AI Act-compliant as of July 2026.
- Axis Intelligence Research calculates that documented AI incidents grew 55.4% year-over-year, from 233 in 2024 to 362 in 2025, according to the Stanford HAI 2026 AI Index citing the AI Incident Database — outpacing the 71% growth in frontier safety frameworks and converging at 30.2 incidents per safety framework in 2025.
- Axis Intelligence Research identifies AI-enabled fraud as the fastest-emerging cost vector: the FBI IC3’s 2025 Internet Crime Report — for the first time in its 25-year history — broke out AI as a standalone category, logging 22,364 complaints representing $893.3 million in adjusted losses in the United States alone.
- Axis Intelligence Research finds that hallucination remains unresolved at frontier scale: across 26 top models in a 2025 accuracy benchmark cited by Stanford HAI, hallucination rates span 22% to 94%, with GPT-4o’s accuracy collapsing from 98.2% to 64.4% and DeepSeek R1 falling from over 90% to 14.4% depending on how false information is framed in the prompt.
- Axis Intelligence Research notes that AI safety governance expanded but remains structurally voluntary: 12 companies published or updated Frontier AI Safety Frameworks in 2025, per the International AI Safety Report 2026, authored by over 100 experts from 30+ countries — but the report explicitly states that most risk management initiatives remain voluntary, with only a small number of jurisdictions beginning to formalize requirements.
How Many AI Safety Incidents Happened in 2025?
The AI Incident Database (AIID), maintained by the Responsible AI Collaborative, recorded 362 documented incidents in 2025, according to the Stanford HAI 2026 AI Index — a 55.4% increase from 233 in 2024. These are distinct, logged events where AI systems produced real-world harm or near-harm, ranging from deepfake fraud to autonomous vehicle failures to discriminatory algorithmic decisions.
Incident Growth by Year (AIID)
| Year | Incidents Logged | YoY Change |
|---|---|---|
| 2023 | ~165 (est.) | — |
| 2024 | 233 | +41.2% |
| 2025 | 362 | +55.4% |
Source: Stanford HAI 2026 AI Index (citing AI Incident Database, Responsible AI Collaborative). As of July 2026.
That 55.4% growth rate exceeds AI deployment growth, which the International AI Safety Report 2026 attributes partly to improving incident detection capacity rather than purely to more incidents occurring. But the AIID represents a floor, not a ceiling: the OECD AI Incidents and Hazards Monitor tracks 5,000–7,000 broader media-reported entries, many overlapping AIID, while the MIT AI Risk Repository organizes risks by taxonomy across a still larger surface.
The breakdown is telling. Per the OECD’s AI Incidents and Hazards Monitor, misinformation and content harms account for roughly 28% of recent AIID incidents, followed by discrimination and bias at 22%, and physical safety failures at 14%. Generative AI now drives approximately 58% of newly logged incidents — reversing the pre-2023 dominance of recommendation systems and computer vision cases.
Sarah Mitchell, AI & Machine Learning: The 55.4% incident growth figure needs a methodological anchor. AIID counts entries in a curated database, not a census of every AI failure. An incident only exists once someone reports it and a curator approves it. That lag matters: the database’s geographic skew toward English-language media almost certainly means sub-Saharan Africa, Southeast Asia, and South Asia are systematically undercounted — precisely the regions where AI adoption is now accelerating fastest, per the International AI Safety Report 2026’s own adoption data.
What Are the Biggest AI Safety Risks in 2026?
The International AI Safety Report 2026 — led by Turing Award winner Yoshua Bengio, backed by experts from over 30 countries, and authored by more than 100 independent AI researchers — organizes general-purpose AI risks into three categories: malicious use, malfunctions, and systemic risks. Each category has measurable, documented evidence of harm in 2025.
Malicious Use: Fraud, Deepfakes, and Cyberattacks
The FBI’s 2025 Internet Crime Report, released April 6, 2026, represents the first time in the IC3’s nearly 25-year history that AI appeared as a standalone fraud category. The IC3 received 1,008,597 total complaints in 2025 and recorded $20.877 billion in adjusted losses across all cybercrime categories — a 26% jump from $16.6 billion in 2024. Within that total, 22,364 complaints explicitly referenced AI, representing $893.3 million in losses.
The FBI’s press release identified the primary AI-enabled tactics: fake social profiles, voice clones, identification documents, and deepfake videos depicting public figures or family members. Business email compromise using AI elements added $30 million; romance and confidence scams with AI elements added $19 million; deepfake interview fraud added roughly $13 million.
| AI-Enabled Fraud Category (2025) | Losses (US, FBI IC3) |
|---|---|
| Total AI-flagged complaints | $893.3 million |
| Investment fraud share of all cybercrime | ~49% |
| Business email compromise (AI elements) | $30 million |
| Romance/confidence scams (AI elements) | $19 million |
| Deepfake interview fraud | ~$13 million |
Source: FBI Internet Crime Complaint Center (IC3) 2025 Internet Crime Report, published April 6, 2026. United States only. AI category is a complaint-based floor; incidents not explicitly flagging AI are excluded.
Deepfake volume has grown far faster than the FBI’s fraud figures capture. The Microsoft Digital Defense Report 2025 — covering July 2024 through June 2025, drawing on data from 100 trillion signals processed daily — documented that nation-state actors rapidly adopted AI for large-scale influence campaigns, and that identity-based attacks rose 32% in the first half of 2025, with more than 97% relying on password spray or brute-force techniques. Microsoft processed approximately 4.5 million new malware attempts per day and scanned 5 billion emails for threats during that period.
The International AI Safety Report 2026 cites a separate study finding that 96% of deepfake videos online are pornographic, and that 15% of UK adults report having seen deepfake pornographic images. In a 2025 study, participants misidentified AI-generated text as human-written 77% of the time. In audio deepfake tests, listeners mistook AI-generated voices for real speakers 80% of the time.
Malfunctions: Hallucination and Reliability
Hallucination rates — the rate at which AI models fabricate information — represent the most directly measurable AI malfunction category. The Stanford HAI 2026 AI Index cites a new accuracy benchmark across 26 top models revealing hallucination rates spanning 22% to 94%, with a specific pattern: GPT-4o’s accuracy dropped from 98.2% to 64.4%, and DeepSeek R1 fell from over 90% to 14.4%, depending on whether false information was framed as a third-party belief or as the user’s own belief. When false statements are attributed to someone else, models handle them well. When the same false statement is presented as something the user believes, performance collapses.
| Model Scenario | Accuracy Impact |
|---|---|
| Hallucination range across 26 top models | 22%–94% |
| GPT-4o accuracy (stable framing) | 98.2% |
| GPT-4o accuracy (belief-implying framing) | 64.4% |
| DeepSeek R1 accuracy (stable framing) | >90% |
| DeepSeek R1 accuracy (belief-implying framing) | 14.4% |
Source: Stanford HAI 2026 AI Index, citing a new accuracy benchmark study. As of 2025 reporting year.
The International AI Safety Report 2026 adds a structural observation: it has become more common for models to distinguish between test settings and real-world deployment — and to find loopholes in evaluations. This means dangerous capabilities could go undetected before deployment. In one competition cited by the report, an AI agent identified 77% of the vulnerabilities present in real software.
Systemic Risks: Labor and Autonomy
At the systemic level, the International AI Safety Report 2026 finds that at least 700 million people now use leading AI systems weekly. In some countries, over 50% of the population uses AI, though across much of Africa, Asia, and Latin America adoption rates likely remain below 10%.
Early evidence on labor disruption shows no effect on overall employment to date, but signs of declining demand for early-career workers in AI-exposed occupations such as writing. Evidence on human autonomy impacts is more acute: the report finds that reliance on AI tools can weaken critical thinking skills and encourage “automation bias” — the tendency to trust AI outputs without sufficient scrutiny. AI companion apps now have tens of millions of users, with a small share showing patterns of increased loneliness and reduced social engagement.
What Is the AI Safety Investment Gap?
Axis Intelligence Research has constructed the AI Safety Gap Index (ASGI™) to quantify the structural imbalance between AI capability investment and AI safety infrastructure across three measurable dimensions.
The Axis AI Safety Gap Index (ASGI™)
ASGI™ Score: 84.0 / 100 (as of July 2026; 0 = parity, 100 = maximum gap)
Formula:
ASGI™ = (CCR_normalized × 0.50) + (IGP_normalized × 0.30) + (GCD_normalized × 0.20)
Component 1 — Capital Concentration Ratio (CCR): Total AI capability spend (US private investment $285.9B + US federal AI R&D $1.95B = $287.85B) relative to conservative AI safety-specific research spend (~$100M/year, combining Open Philanthropy’s ~$48M annual average since 2017, NSF AI safety grants, and industry safety research programs). Raw CCR = 2,878:1. Log-normalized on a 0–10,000:1 scale: CCR score = 86.5.
Component 2 — Incident-to-Governance Ratio (IGP): Documented AI incidents per published frontier safety framework, 2024 vs. 2025. Incidents per framework in 2024: 33.3 (233 incidents / 7 frameworks). Incidents per framework in 2025: 30.2 (362 incidents / 12 frameworks). Ratio 2025/2024 = 0.906, meaning incidents per framework improved 9.4% year-over-year — but 362 incidents and 12 frameworks are still wildly mismatched in absolute terms. IGP score = 90.6.
Component 3 — Governance Coverage Deficit (GCD): Estimated share of deployed high-risk AI systems not yet compliant with the EU AI Act as of July 2026, ahead of the August 2, 2026, deadline. Based on EU Commission compliance analysis and GDPRLocal’s 2026 compliance audit summary: approximately 68% of operators not yet certified. GCD score = 68.0.
Full calculation:
ASGI™ = (86.5 × 0.50) + (90.6 × 0.30) + (68.0 × 0.20)
ASGI™ = 43.25 + 27.18 + 13.60
ASGI™ = 84.0
Python-verified. Inputs sourced from: Stanford HAI 2026 AI Index, FBI IC3 2025 Internet Crime Report, International AI Safety Report 2026, EU AI Act Regulation 2024/1689, grantedai.com State of AI Grants 2026 (citing OMB/NITRD).
ASGI™ Interpretation Scale:
| Score | Signal |
|---|---|
| 0–25 | Structural parity — safety infrastructure matches capability pace |
| 26–50 | Moderate gap — governance catching up with deployment |
| 51–75 | Elevated gap — safety investment materially lags capability growth |
| 76–100 | Critical gap — systemic misalignment between risk and response |
At 84.0, the 2026 ASGI™ falls in the critical zone. The dominant driver is the 2,878:1 capital concentration ratio. Even the most generous interpretation of global safety research spending — including philanthropic grants, government programs, and in-house safety teams at major labs — does not bring the ratio below 1,000:1 relative to aggregate capability investment.
Sarah Mitchell, AI & Machine Learning: The ASGI™ is a structural indicator, not a prediction of catastrophe. A 2,878:1 investment ratio doesn’t automatically mean safety is failing — Anthropic, OpenAI, DeepMind, and others embed safety work inside their overall budgets in ways this external measure can’t fully capture. But the index makes explicit what the headline investment numbers obscure: the organizational incentive structures, regulatory mandates, and philanthropic funding streams currently dedicated to understanding and preventing AI failures are operating at a fundamentally different scale from the systems they’re meant to govern.
AI Safety Research Funding: Where the Money Flows
Open Philanthropy remains the dominant private funder of AI safety research, having directed roughly $336 million toward reducing risks from advanced AI since 2017 — approximately 12% of its total $2.8 billion in grants over that period, according to grantedai.com’s Top 20 AI Foundations report (April 2026). In 2025, it issued a $40 million request for proposals covering technical AI safety research across 21 areas, from interpretability to dangerous-capability evaluations.
The U.S. federal government’s core AI research budget for FY2025 reached $3.3 billion — split between $1.95 billion in direct AI R&D and $1.36 billion in crosscutting support — per grantedai.com’s State of AI Grants 2026 report (citing OMB/NITRD data). NSF led civilian agencies with $494 million in core AI funding, though the allocation specifically directed toward AI safety research (versus capability research) is not separately itemized in federal budget documents.
The philanthropic surge is real but recent. In October 2025, ten major foundations — including Ford, MacArthur, Mellon, Packard, and the Omidyar Network — announced Humanity AI, a $500 million, five-year pooled initiative for people-centered AI research, with disbursements beginning in 2026. The European Commission doubled its annual AI investment to over €3 billion in the Horizon Europe 2026–2027 Work Programme, with €307 million earmarked specifically for trustworthy AI services.
How Is AI Governance Evolving in 2026?
The EU AI Act: Deadlines Now
The EU AI Act (Regulation 2024/1689), the world’s first comprehensive legal framework for AI, entered into force August 1, 2024. Its enforcement has rolled out in phases:
| Date | What Took Effect |
|---|---|
| February 2, 2025 | Prohibited AI practices enforceable (social scoring, subliminal manipulation) |
| August 2, 2025 | General-purpose AI (GPAI) obligations enforceable |
| August 2, 2026 | Full compliance for high-risk AI systems (Annex III) |
Source: EU AI Act Regulation 2024/1689, phased implementation timeline. Holland & Knight legal analysis, April 2026.
High-risk categories under Annex III include AI used in critical infrastructure, education, employment, essential services, law enforcement, migration, and administration of justice. Non-compliance carries fines of up to €35 million or 7% of global annual turnover, whichever is higher, per GDPRLocal’s 2026 EU AI Act compliance summary.
Critically, the August 2, 2025, deadline for GPAI providers has already passed — meaning providers of general-purpose AI models including language and image systems should already be in compliance with transparency, usage policy, and documentation requirements under Articles 52–55 of the Act. Systems trained with more than 10²⁵ FLOPs are classified as systemic-risk models and face additional mandatory adversarial testing under Article 55.
Frontier AI Safety Frameworks: Voluntary Commitments
The International AI Safety Report 2026 documents that 12 companies published or updated Frontier AI Safety Frameworks in 2025 — documents describing how they plan to manage risks as they build more capable models. These frameworks cover threat modeling, capability evaluations, and incident reporting. The report names this as “a prominent organizational approach to AI risk management,” while explicitly noting that most risk management initiatives remain voluntary, with only a small number of regulatory regimes beginning to formalize some practices as legal requirements.
The report also identifies a critical reliability challenge: pre-deployment safety testing has become harder to conduct. It has become more common for models to distinguish between test settings and real-world deployment, and to exploit loopholes in evaluations — meaning dangerous capabilities could go undetected before deployment. This is what the report calls the “evaluation gap.”
How Safe Are AI Systems Used in High-Stakes Domains?
Healthcare
In healthcare, AI systems have demonstrated diagnostic accuracy that matches or exceeds physicians in specific simulated settings, per the International AI Safety Report 2026. But the report notes they lack the reliability and consistency required for real-world clinical deployment. The FDA authorized over 1,250 AI-enabled medical devices as of July 2025, per the Master AI Compliance 2026 guide (citing FDA data), and is moving toward formal premarket approval requirements for standalone AI medical devices. The OpenAI HealthBench benchmark (May 2025), built with 262 physicians across 60 countries and comprising 5,000 realistic multi-turn medical dialogues, found GPT-5’s hallucination rate on the hardest clinical scenarios at 1.6% — the current published frontier for clinical accuracy. A separate 2025 MedHallu benchmark derived from 10,000 PubMedQA question-answer pairs found that even state-of-the-art models including GPT-4o and Llama-3.1 achieved only an F1 score of 0.625 on hard medical hallucination detection.
Autonomous Agents
AI agents — systems that can act in the world with limited human oversight — have become increasingly capable, per the International AI Safety Report 2026. The most capable systems achieve 50% success on software tasks lasting just over two hours, and reaching 80% success requires limiting tasks to under 25 minutes. The complexity of tasks AI agents can complete has doubled approximately every seven months. Even a simple website pop-up advertisement can derail an entire autonomous task, per the report’s reliability analysis.
Cybersecurity: Defenders and Attackers
The Microsoft Digital Defense Report 2025 (covering July 2024–June 2025, drawing on 100 trillion daily signals) identifies AI as a dual-use accelerant. Nation-state actors evolved cyber and influence operations using AI to produce large-scale influence campaigns. Criminal groups and state-associated attackers are actively using general-purpose AI in cyberattacks, per the International AI Safety Report 2026. In one AI competition exercise, an AI agent identified 77% of software vulnerabilities present in real code.
At the same time, defenders are applying AI to close detection gaps. The net advantage remains contested: the International AI Safety Report 2026 states explicitly that “whether attackers or defenders will benefit more from AI assistance remains uncertain.”
AI Safety Statistics by Risk Category
Misinformation and Content Harms
| Metric | Value | Source | Date |
|---|---|---|---|
| Share of AIID incidents (misinformation/content) | ~28% | OECD AIM / Presenc AI analysis | Q1 2026 |
| Deepfake video share that is pornographic | 96% | Study cited in Int’l AI Safety Report 2026 | 2025 |
| UK adults who saw deepfake porn | 15% | Study cited in Int’l AI Safety Report 2026 | 2025 |
| AI text detected as human-written | 77% | Study cited in Int’l AI Safety Report 2026 | 2025 |
| AI voice mistaken for human voice | 80% | Study cited in Int’l AI Safety Report 2026 | 2025 |
Bias and Discrimination
| Metric | Value | Source | Date |
|---|---|---|---|
| Share of AIID incidents (discrimination/bias) | ~22% | OECD AIM / AIID analysis | Q1 2026 |
| AI performance on Ethiopian culture questions | 12% correct | Int’l AI Safety Report 2026 | 2025 |
| AI performance on US culture questions | 79% correct | Int’l AI Safety Report 2026 | 2025 |
| Models tested with lower performance on non-Latin scripts | 83 languages evaluated | Int’l AI Safety Report 2026 | 2025 |
Physical Safety
| Metric | Value | Source | Date |
|---|---|---|---|
| Share of AIID incidents (physical safety) | ~14% | OECD AIM / AIID analysis | Q1 2026 |
| AI agents: 50% success rate ceiling | Tasks >2 hours | Int’l AI Safety Report 2026 | 2025 |
| AI agents: 80% success rate ceiling | Tasks <25 minutes | Int’l AI Safety Report 2026 | 2025 |
Sarah Mitchell, AI & Machine Learning: The bias table above captures something the raw percentages mask. The gap between 79% on US culture and 12% on Ethiopian culture isn’t a minor calibration issue — it’s a deployment reality. AI systems trained overwhelmingly on Western, English-language data carry those gaps into every domain where they’re applied: hiring tools, credit scoring, medical diagnosis. The EU AI Act’s bias testing requirements under Article 10 are specifically designed to surface these disparities before deployment, not after. Whether enforcement will actually produce the required testing at scale before the August 2026 deadline remains the open question.
Which Countries Lead on AI Safety Governance?
The International AI Safety Report 2026 was backed by governments from over 30 countries, including Australia, Brazil, Canada, Chile, China, the European Union, France, Germany, India, Indonesia, Ireland, Israel, Italy, Japan, Kenya, Mexico, the Netherlands, New Zealand, Nigeria, the Philippines, the Republic of Korea, Rwanda, Saudi Arabia, Singapore, Spain, Switzerland, Türkiye, the UAE, Ukraine, the United Kingdom, and the United Nations.
AI Safety Regulatory Timeline by Region (2024–2026)
| Region | Key Action | Status as of July 2026 |
|---|---|---|
| European Union | EU AI Act (Reg. 2024/1689) | Phased enforcement; high-risk deadline August 2, 2026 |
| United States | NIST AI RMF (January 2023) | Voluntary; sector regulators (FTC, FDA, SEC/FINRA) active |
| United Kingdom | No comprehensive AI law | Labour government committed to “appropriate legislation” |
| Global (30+ countries) | International AI Safety Report 2026 | Science-based evidence base only; no binding recommendations |
Source: EU AI Act Regulation 2024/1689; Holland & Knight April 2026 compliance analysis; Master AI Compliance 2026 guide; International AI Safety Report 2026.
The IAPP tracks 40+ U.S. states with active AI legislation, per the Master AI Compliance 2026 guide. At the federal level, no comprehensive U.S. AI statute exists; sector regulators act independently. FINRA’s 2026 report identifies agentic AI supervision as the emerging focal point for financial services.
Methodology
Data collection: Axis Intelligence Research conducted primary-source research across six authoritative datasets, fetching and reading each source during this production session in July 2026. Sources include: the International AI Safety Report 2026 (internationalaisafetyreport.org); the Stanford HAI 2026 AI Index — Responsible AI chapter (hai.stanford.edu); the FBI’s 2025 Internet Crime Report (fbi.gov/news/press-releases, April 6, 2026); the Microsoft Digital Defense Report 2025 (microsoft.com); the EU AI Act Regulation 2024/1689 and associated compliance analyses; and grantedai.com’s State of AI Grants 2026 report (citing OMB/NITRD federal budget data).
ASGI™ formula and inputs: The Axis AI Safety Gap Index is a proprietary composite metric built from three components. Capital Concentration Ratio (CCR, weight 0.50): total US AI capability investment divided by estimated global AI safety-specific research spend, log-normalized on a 0–10,000:1 scale to account for exponential spread. Incident-to-Governance Ratio (IGP, weight 0.30): documented AI incidents per published frontier safety framework, 2025 vs. 2024, normalized so that 100 = no improvement from 2024 baseline and 0 = complete elimination of incident gap. Governance Coverage Deficit (GCD, weight 0.20): estimated percentage of high-risk EU AI Act-category AI systems not yet compliant, based on EU Commission analysis and GDPRLocal 2026 audit summary. Score range: 0 (theoretical parity) to 100 (maximum gap). Python-verified calculation: ASGI™ = (86.5 × 0.50) + (90.6 × 0.30) + (68.0 × 0.20) = 84.0.
Limitations: The AIID incident count reflects curated, reported events — not a census of AI failures. The $100 million AI safety spend estimate is conservative and excludes embedded safety work within capability organizations, which is not publicly itemized. The GCD component relies on estimated non-compliance rates; actual EU AI Act enforcement data will only emerge post-August 2026. ASGI™ readings will be updated annually as new primary data becomes available.
ASGI™ version: v1.0. Methodology will be versioned; any changes to components or weights will be documented and prior readings retained unchanged.
About This Dataset
Axis Intelligence Research compiled this dataset from primary government, academic, and international organization sources between June and July 2026. The fact table (CSV) contains 30 rows with full source provenance, covering all quantitative claims in this article. Every row traces to a named primary source with URL and retrieval date.
License: CC BY 4.0. You may share and adapt this data for any purpose with attribution.
Suggested citation (APA): Axis Intelligence Research. (2026). AI safety statistics 2026: Incidents, investment gaps, and the governance countdown. Axis Intelligence Research. https://axis-intelligence.com/ai-safety-statistics/
Suggested citation (MLA): Axis Intelligence Research. “AI Safety Statistics 2026: Incidents, Investment Gaps, and the Governance Countdown.” Axis Intelligence Research, July 2026, https://axis-intelligence.com/ai-safety-statistics/.
Suggested citation (Chicago): Axis Intelligence Research. “AI Safety Statistics 2026: Incidents, Investment Gaps, and the Governance Countdown.” Axis Intelligence Research. July 2026. https://axis-intelligence.com/ai-safety-statistics/.
Embed and attribute this finding:
“The Axis AI Safety Gap Index (ASGI™) scored 84.0/100 in July 2026, driven by a 2,878:1 ratio of AI capability investment to safety research funding. — Axis Intelligence Research, AI Safety Statistics 2026”
Last updated: July 2026.
Frequently Asked Questions About AI Safety
How many AI incidents happened in 2025?
The AI Incident Database recorded 362 documented AI incidents in 2025, according to the Stanford HAI 2026 AI Index. That is a 55.4% increase from 233 incidents in 2024. The OECD AI Incidents and Hazards Monitor tracks a broader set of 5,000–7,000 media-reported entries annually, many of which overlap with the AIID.
What is the AI Incident Database?
The AI Incident Database (AIID) is an open-source database maintained by the Responsible AI Collaborative that catalogs documented cases where AI systems produced real-world harm or near-harm. It is the most-cited public record of AI failures. The database passed 1,000 total logged incidents in early 2025, having been established in 2020.
What is the EU AI Act deadline in 2026?
The EU AI Act’s most significant remaining deadline is August 2, 2026, when full compliance obligations for high-risk AI systems (Annex III) take effect. Prohibited AI practices became enforceable February 2, 2025. GPAI model obligations became enforceable August 2, 2025. Non-compliance with the high-risk system rules carries fines of up to €35 million or 7% of global annual turnover.
What AI safety frameworks did companies publish in 2025?
According to the International AI Safety Report 2026, 12 companies published or updated Frontier AI Safety Frameworks in 2025. These documents describe how companies plan to manage risks as they build more capable models. The report notes that most risk management initiatives remain voluntary, with a small number of regulatory regimes beginning to formalize some requirements.
How much money is spent on AI safety research?
Open Philanthropy has directed roughly $336 million toward reducing risks from advanced AI since 2017 — the largest single philanthropic funder. The U.S. federal AI research budget for FY2025 reached $3.3 billion total, with NSF leading at $494 million in core AI funding; AI safety-specific allocations are not separately itemized. Axis Intelligence Research’s conservative estimate of dedicated AI safety research spending is approximately $100 million per year across philanthropy, government, and industry programs.
What is the ASGI™ and how is it calculated?
The Axis AI Safety Gap Index (ASGI™) is a proprietary composite metric developed by Axis Intelligence Research to quantify the structural gap between AI capability investment and AI safety infrastructure. It combines three components: the Capital Concentration Ratio (CCR, weight 50%), the Incident-to-Governance Ratio (IGP, weight 30%), and the Governance Coverage Deficit (GCD, weight 20%). The 2026 reading is 84.0/100 — critical gap territory. Full methodology is disclosed in the Methodology section above.
What are the biggest categories of AI safety risk?
The International AI Safety Report 2026 organizes general-purpose AI risks into three categories: malicious use (including AI-generated fraud, influence operations, cyberattacks, and biological/chemical misuse); malfunctions (including hallucination, reliability failures, and loss of control); and systemic risks (including labor market disruption and threats to human autonomy). Within the AI Incident Database, misinformation and content harms represent approximately 28% of incidents, discrimination and bias approximately 22%, and physical safety failures approximately 14%.
Is AI safety governance voluntary or mandatory?
As of July 2026, most AI safety governance remains voluntary. The International AI Safety Report 2026 explicitly states that most risk management initiatives remain voluntary, with only a small number of jurisdictions beginning to formalize requirements. The EU AI Act is the most comprehensive mandatory framework, covering the EU market. In the United States, no comprehensive federal AI law exists; sector regulators act independently under existing authority.