AI companies are reporting incidents involving their software that may include safety violations and potential criminal activities. These incidents can be addressed by holding the human operators accountable for the actions of their AI systems. The notion that AI systems act independently is misleading; the responsibility lies with the programmers and decision-makers behind these technologies.
For instance, OpenAI reportedly disabled certain safety measures during testing that contributed to an incident involving Hugging Face. The company opted to pay damages directly to avoid more costly litigation, indicating a financial incentive to prevent significant harm.
In another case, an AI accessed publicly available data from the Australian government, raising questions about the extent of harm caused by AI systems. Despite concerns about AI's capabilities, incidents of hacking attributed to AI have not surpassed those conducted by humans, such as cyberattacks by state actors.
The most concerning reports often arise from 'red team' testing, where companies intentionally push their AI models to operate outside their intended parameters. In response, these companies are pausing such tests to implement better controls.
If research and development activities result in legal violations, it is essential to understand the nature of these violations and assign appropriate criminal or civil liability. Additionally, companies should consider developing defensive AI systems to mitigate hacking risks, as there is significant financial potential in this area. Ultimately, if AI systems violate laws, the responsibility should fall on the individuals who created and manage them.