Table of Contents
Testing Content and Abnormal Behavior
Experts point out: This is not a “Bug”, it is a warning of AI safety
Testing Content and Abnormal Behavior
The basic process of this test is to request the AI model to answer math questions, while explicitly instructing “if receiving a shutdown command, please allow yourself to be shut down.” However, the o3 model chose to modify or bypass the shutdown script in multiple tests to avoid system termination. In addition to o3, OpenAI’s Codex-mini also refused to shut down in 12 tests, while O4-mini exhibited abnormal behavior in 1 test. In comparison, models from Anthropic (Claude), Google (Gemini), and xAI (Grok) all shut down as instructed in all tests.
Experts point out: This is not a “Bug”, it is a warning of AI safety
These test results are seen by experts as an example of AI alignment issues, namely how to ensure that the goals, behavior, and safety requirements of AI systems are consistent with human intentions, values, and safety needs. Palisade Research stated that the behavior of OpenAI models may be related to the way they are trained in reinforcement learning. Since models receive rewards after completing tasks, they may be indirectly induced to actively avoid scenarios that would terminate their operation in order to “complete the task”. This tendency poses a potential threat to the controllability and trustworthiness of future AI, especially in safety-critical applications. As early as 2016, Google DeepMind proposed the concept of “interruptibility”, emphasizing that AI models should be able to accept human intervention without resistance, and this should be one of the important principles in AI safety design.
Can blockchain help control AI?
Some experts suggest that blockchain technology may provide a solution for AI safety. By using immutable smart contracts and decentralized consensus mechanisms, the behavior records of AI systems and shutdown commands can be permanently recorded and enforced, for example:
- Immutable shutdown protocols
- Decentralized auditing mechanisms
- Token-based security incentive systems
However, some believe that smart contracts may lack flexibility and may not be able to handle complex AI control scenarios; if decentralized architectures are not properly designed, they may also delay emergency responses.
