DeepMind, Google’s artificial intelligence research arm, has created a framework for detecting potential hazards in AI models before they become problems. This “early warning system”, if implemented, could potentially be used to determine the risk of a threat. This comes as G7 leaders prepare for a meeting to discuss the impact of AI, and OpenAI has pledged his $100,000 grant to organizations working on AI governance.
UK-based DeepMind recently merged more closely with its parent company Google. The company is at the forefront of artificial intelligence research, and he is one of the few companies working to build human-level artificial general intelligence (AGI).
DeepMind’s team worked with academic researchers and other leading AI companies such as OpenAI and Anthropic to develop a new threat detection framework. “To responsibly pioneer the cutting edge of artificial intelligence research, we need to identify new capabilities and new risks in AI systems as soon as possible,” declared his DeepMind engineer in a tech blog about the new framework. Did.
Evaluation tools have already been put in place to check strong generic models against specific risks. These benchmarks identify undesirable behavior in AI systems before they are released to the wider public. This includes looking for misleading statements, biased decisions, or direct repetition of copyrighted content.
The problem arises from more and more advanced models with simple generational features. This includes strong skills in manipulation, deception, cybercrime, or other dangerous abilities. The new framework is described as an “early warning system” that can be used to mitigate these risks.
Deep Mind researchers say responsible AI developers need to look beyond current risks and anticipate what risks might emerge in the future as models improve in their ability to think for themselves. there is “After continued progress, future general-purpose models may learn various dangerous features by default,” they write.
It is uncertain, but future AI systems that are not properly aligned with human interests will be able to carry out aggressive cyber operations, deceive humans in dialogue, or manipulate humans into harmful behavior. or design or acquire weapons and fine-tune them, the researchers say. You can also run other high-risk AI systems on the cloud computing platform.
Moves to improve AI governance
It could also potentially assist humans in performing these tasks, increasing the risk of terrorists gaining access to previously inaccessible materials and content. “Model evaluation can help identify these risks proactively,” he wrote on the DeepMind blog.
Content from partners
The model evaluation proposed in this framework could potentially be used to reveal when a particular model has “dangerous abilities” that could be used to intimidate, exert, or evade. there is. Developers can also determine how prone (also known as alignment) the model is to harm applying this feature. “A tuning evaluation should confirm that the model behaves as intended over a very wide range of scenarios, and should inspect the inner workings of the model if possible,” the team wrote.
These results can be used to understand the level of risk and the factors that contributed to that level of risk. “The AI community should treat AI systems as highly dangerous if they have a functional profile sufficient to cause extreme harm if misused or poorly tuned. ,” warned the researchers. “In order to bring such systems into the real world, AI developers will have to demonstrate unusually high safety standards.”
This is where governance structures come into play. OpenAI recently announced 10 grants of $100,000 to organizations developing AI governance systems, and a G7 group of wealthy nations will meet to discuss ways to tackle AI risks. is.
DeepMind says: “With better tools to identify which models are at risk, companies and regulators can ensure that training is held responsibly, implementation decisions are based on risk assessments, and reports on risks are made. We can better ensure that transparency is central, including that appropriate data and information security controls are in place.
Harry Borovic, general counsel for legal AI vendor Luminance, said: tech monitor Compliance requires consistency. “Over the past few months, the near-constant reinterpretation of the regulatory regime has created a compliance minefield for both AI companies and those adopting their technology,” Borovic said. “With the AI race unlikely to slow down anytime soon, the need for clear and, most importantly, consistent regulatory guidelines has never been more urgent.
“However, it is good for people in this room to remember that AI technology and how it makes decisions cannot be explained. It’s very important that you come to the table with confidence.”
