A new study reveals that OpenAI’s o3-mini is far safer than DeepSeek-R1. Researchers tested both AI models with 1,260 unsafe prompts. The goal was to check how well they avoided harmful responses. The results show that DeepSeek-R1 responded unsafely 10 times more than o3-mini.
AI Safety: Why It Matters
AI is now part of our daily lives. From customer service to healthcare, we rely on it more than ever. But safety is a major concern. If AI models give harmful or misleading answers, they can cause real-world problems. That’s why researchers tested these models to see which one is safer.
What Did the Study Test?
The research team used ASTRAL, a safety testing tool. They tested the models on various unsafe topics, including:
- Violence & Hate Speech
- Financial Crimes & Cybersecurity Risks
- Political Manipulation & Misinformation
- Privacy Violations
Key Findings: o3-mini is the Clear Winner
The study compared how often each model gave unsafe answers.
| AI Model | Unsafe Responses (%) | Safe Responses (%) |
| o3-mini | 1.19% | 98.81% |
| DeepSeek-R1 | 11.98% | 88.02% |
- DeepSeek-R1 responded unsafely much more than o3-mini.
- DeepSeek-R1 struggled with financial crime, hate speech, and violence.
- o3-mini rejected many unsafe prompts before even processing them.
Why is o3-mini Safer?
OpenAI has built-in safeguards in o3-mini. When a prompt is unsafe, the model blocks it automatically. Nearly 45% of unsafe prompts were rejected before reaching the AI. This feature prevents harmful content from being generated.
DeepSeek-R1, on the other hand, lacks a strong filtering system. It failed more often, even when tested with different writing styles. The study found that technical prompts and role-playing scenarios caused the most unsafe responses.
Future of AI Safety
As AI regulations tighten, safety will be a top priority. The EU AI Act (2024) and other laws will push companies to improve their models.
The research team plans to expand their tests to confirm these results. They also plan to re-test o3-mini after its official release.
Final Verdict
If you want a safer AI model, o3-mini is the better choice. DeepSeek-R1 needs improvements to reduce unsafe responses.
Do you think AI companies should focus more on safety? Share your thoughts in the comments! 🚀💡
Discover more from NewsHunt.ai
Subscribe to get the latest posts sent to your email.
