Researchers at Google DeepMind have released a new white paper explaining how to safely develop artificial general intelligence (AGI). Despite skeptics' predictions, they believe that AI of this level could be coming relatively soon, and they look at four ways it could cause harm.
Researchers at Google DeepMind have released a new white paper explaining how to safely develop artificial general intelligence (AGI). Despite skeptics' predictions, they believe that AI of this level could be coming relatively soon, and they look at four ways it could cause harm.
The document contains a huge amount of detail and runs to about 108 pages before references. While some in the AI industry believe that AGI, or strong artificial intelligence (SAI), is an unattainable concept, DeepMind researchers believe it could happen by 2030, Ars Technica reports.
With this in mind, researchers sought to understand the risks of human-like synthetic intelligence, which they believe could lead to «serious harm».
The study outlines four possible types of risks associated with building AI, along with suggestions for how we can mitigate those risks. The DeepMind team, led by co-founder Shane Legg, categorized the negative impacts of AI as misuse, bias, error, and structural risks. Misuse and non-compliance are discussed in detail in the paper, but the latter two are only briefly touched upon.
AI abuse is the kind of problem that humans already face. But as AI becomes more powerful, the damage it can do will be much greater. An AI attacker could abuse the system to cause harm, for example by asking the system to detect and exploit zero-day vulnerabilities or by creating a virus that can be used as a biological weapon.
DeepMind warns that companies developing AI will need to conduct extensive testing and create robust security protocols after training. They also propose developing a method to completely suppress dangerous capabilities, sometimes called «back-learning,» but it is unclear whether this is possible without severely limiting the models.
Deviation is reminiscent of the problems with generative AI as it exists today. But in the study of AI harm, it is described as a machine that has gone beyond the limits set by its creators. In this case, the artificial intelligence consciously performs actions that were not planned by the developer. DeepMind claims that their standard of deviation is more advanced, and takes into account not only the simple deception or intrigue that modern AI is capable of, which has already been described in previous studies .
To avoid this, DeepMind suggests developers use techniques such as enhanced control, where two copies of the AI check each other’s work to create robust systems that are unlikely to fail. If that fails, DeepMind suggests conducting intensive stress testing and monitoring to track any hints that the AI might turn against humans.
Keeping AI in virtual sandboxes with strict security and direct human oversight can help mitigate problems arising from disagreements. Basically, make sure there is an «off» switch.
Mistakes are the type of problem in which the human operator did not intend to cause harm, and the AI did not know that its results would cause harm. Many such mistakes already occur with modern AI systems, such as when Google’s AI advised to spread glue on pizza or cook poisonous mushrooms. DeepMind notes that the military may deploy AGI due to «competitive pressures,» but such systems could make serious mistakes because they would be tasked with much more complex functions than current AI.
The paper doesn’t offer a foolproof solution to reducing errors. It boils down to preventing AI from becoming too powerful. DeepMind calls for slow deployment and limiting AI’s powers. The paper also suggests running AGI commands through a «safeguard» system that ensures they’re safe before they’re deployed.
DeepMind defines structural risks as unintended but real consequences of the impact of multi-agent systems on human existence. For example, AI can create false information that is so believable that people begin to doubt everything.
The paper also raises the question whether AGI could accumulate more and more control over economic and political systems, perhaps by developing rigid tariff schemes. In such a case, AI could gradually take over humanity. This category of risks is also the most difficult to protect against, as it will depend on how people, infrastructure, and institutions work in the future, the researchers point out.
An earlier study called «Model Alignment between Statements and Knowledge» showed that large AI models can lie to their users under pressure .
While various tests and tools test AI for accuracy, the MASK benchmark was designed to determine whether AI believes what it tells users — and under what circumstances it might provide incorrect information. The study tested 27 models from the GPT, Llama, Qwen, Claude, and DeepSeek families.
