AI reliability and risk statistics
How often AI gets real tasks right, where it still fails and how reported incidents have grown.
These figures show the limits of current AI systems. They are the reason a professional checks AI work before it goes to a client.
Key takeaways
- AI agents completed 66.3% of computer tasks on the OSWorld test in 2025. The human baseline is 72.35%, according to the Stanford 2026 AI Index.
- The top model read an analog clock correctly 50.1% of the time in the same report.
- Documented AI incidents rose to 362 in 2025, from 233 in 2024.
How often AI gets the task right
66.3% of computer tasks on the OSWorld test were completed by AI agents in 2025.
The human baseline on the same test is 72.35%. OSWorld tests an agent on tasks in real computer programs.
What this means: AI agents now complete most tasks on this test, but they still fail about one task in three. Review agent work before you rely on it.
50.1% of the time, the top AI model read an analog clock correctly.
The report uses this as an example of uneven ability. The same models pass very hard tests in other areas.
What this means: A model that passes hard tests can still fail a simple one. Do not assume that skill in one area carries over to another.
19 percentage points lower chance of a correct solution when consultants used AI on a task outside its abilities.
The experiment had 758 consultants at Boston Consulting Group.
What this means: Trust in AI can lower the quality of work when the task is outside AI ability. Human judgment is still necessary.
Incidents and trust
362 AI incidents were documented in 2025, up from 233 in 2024.
The report tracks publicly documented incidents year by year.
What this means: Documented AI problems increase as use increases. A check step in your workflow protects you and your customers.
31% of people in the United States trust their government to regulate AI.
This is the lowest figure among the countries in the report. Singapore is the highest, at 91%.
What this means: Most people in the United States do not trust the government to manage AI risk. Clear rules about how your business uses AI can build trust with your customers.
What the numbers mean for your work
Current AI is strong and uneven. An agent that finishes two of three computer tasks is useful, and it still needs review on every task that matters.
Build a check step into each AI workflow. Give the model the source material, ask it to show where each claim comes from, and approve the result before it leaves your business.
Sources
| Organization | Report or study | Published |
|---|---|---|
| Stanford HAI | The 2026 AI Index Report | 2026 |
| Harvard Business School working paper | Navigating the Jagged Technological Frontier (Dell'Acqua and others) | September 2023 |
To cite this page, use: Excellent AI Prompts, "AI reliability and risk statistics", https://www.excellentaiprompts.com/statistics/ai-reliability-statistics, figures checked November 10, 2026.