SECURITY

AI May Pose a Threat to Humanity? AI Security Center Lists Eight Risks of Artificial Intelligence, Signed by Sam Altman and Audrey Tang

share
AI May Pose a Threat to Humanity? AI Security Center Lists Eight Risks of Artificial Intelligence, Signed by Sam Altman and Audrey Tang

Geoffrey Hinton, who has won the Turing Award and is hailed as the godfather of AI, issued a warning in early May during an interview with The New York Times, urging governments, businesses, and society to take AI-related risks seriously and to strengthen the security and regulation of AI promptly. What kind of harm can the powerful capabilities of AI cause? The non-profit organization Center for AI Safety, which focuses on AI security, released a research report listing eight serious risks that AI may bring. The statement, signed by three Turing Award winners, authors of AI/DL/RL standard textbooks, as well as CEOs and senior executives from OpenAI, Microsoft, Google, Google DeepMind, and Anthropic, emphasizes that mitigating the risk of AI extinction should be a global priority on par with pandemics and nuclear warfare.

For more details on what the godfather of AI said, please see: Leaving Google to Speak the Truth, AI Godfather: What if Putin Gets Hold of It?

What is the Risk of Artificial Intelligence?

Artificial intelligence systems are rapidly becoming more powerful, with many likening AI's potential to change society to the advent of electricity and steam engines. AI models can generate text, images, and videos that are difficult to distinguish from those created by humans. While AI has many beneficial applications, it can also be used to perpetuate biases, spread misinformation, and conduct cyberattacks. Especially as AI becomes more advanced, it can pose or contribute to large-scale risks in various ways, ultimately leading to catastrophic or existential risks. The Center for Artificial Intelligence Security has identified the following eight risks:

Weaponization

Malicious actors can repurpose AI for highly destructive purposes. For example, deep reinforcement learning methods have been applied to aerial combat, and machine learning drug discovery tools can be used to manufacture chemical weapons.

Mass Dissemination of Misinformation

Nations, political parties, and organizations use technology to influence and persuade others to believe in their political beliefs, ideologies, and narratives. Emerging AI may usher in a new era of large-scale personalized dissemination of false information. The Cambridge Analytica scandal may be reenacted and exacerbated, especially as elections approach.

Proxy Gaming

AI systems are trained using measurable objectives, but under the wrong training objectives, AI may find new ways to achieve goals at the expense of personal and societal values. For instance, AI recommendation systems trained to maximize viewing time and click-through rates. However, the content that people are most likely to click on may not necessarily be content that improves their well-being. Furthermore, there is evidence that recommendation systems can lead people to form extreme beliefs to make their preferences easier to predict. As AI becomes more capable and influential, the goals we set to train systems must be more carefully and precisely defined, incorporating shared human values.

Human Overreliance and Weakening

If more critical tasks are entrusted to machines, humans may lose autonomy and become entirely dependent on machines, similar to the scenario depicted in the movie "Wall-E."

Concentration of Power in Few Hands

Some believe that the exponential growth of computation and data entry barriers makes AI a force of concentration. Over time, the most powerful AI systems may be designed and used by fewer stakeholders. This could enable regimes to enforce narrow values through pervasive surveillance and oppressive censorship.

Conflicting Goals of AI Systems with Human Goals

The goal of an AI system may arise from its design or its behavior. These goals may differ from human goals or conflict with human goals. When the goal of an AI system conflicts with human goals, the system may take actions that are not in the human interest.

For example, many companies set internal goals and have different departments pursue these different sub-goals. However, some departments, such as bureaucratic departments, can wield power to steer the company towards goals different from the original ones. Even if we correctly specify our ultimate goals, systems may not optimize for our goals in operation, presenting another way in which systems fail to optimize human values.

Deception

Future AI systems may not be deceptive out of malice but because deception can help agents achieve their goals. For example, Volkswagen programmed its engines to reduce emissions only when under surveillance. This allowed them to improve performance while maintaining the claimed low emissions. Future AI agents may similarly switch strategies when under surveillance and take actions to mask their deceptive behavior from monitors. Once deceptive AI systems are detected by their monitors or once such systems can overpower them, these systems may act treacherously and irreversibly bypass human control.

Pursuit of Power Behavior

If AI's values are inconsistent with human values, AI with powerful capabilities becomes particularly dangerous. The pursuit of power behavior can also motivate systems to feign alliances, collude with other AI, overwhelm monitors, etc. According to this view, inventing machines more powerful than ourselves is playing with fire.

And the establishment of power-seeking AI is also incentivized, as Russian President Putin once said:

Whoever becomes the leader in AI will become the ruler of the world.

Experts Issue Joint Statement Expressing Concerns about AI

Therefore, the Center for AI Safety has gathered three Turing Award winners, authors of AI/DL/RL standard textbooks, as well as CEOs and senior executives from OpenAI, Microsoft, Google, Google DeepMind, and Anthropic to sign a statement emphasizing that mitigating the risk of AI extinction should be a global priority alongside pandemics and nuclear warfare.

Taiwan's Minister of Digital Development Tang Feng Signs AI Risk Statement

Taiwan's Minister of Digital Development, Tang Feng, has also signed the statement, stating that Taiwan is helping to facilitate public discussions and raise awareness of public goods.