Nascent capabilities (emergent capabilities)

Example: "Power seeking behaviours: Although this point is still the subject of much research and debate, AI systems tasked with ambitious objectives and minimal oversight may exhibit an increased propensity to pursue power. Some studies show a tendency toward power-seeking behaviors,447 which could be explained by the fact that generative AI models try to gain control over the environment and other actors to reach their goals. For instance, researchers at Anthropic have conducted experiments to assess their models’ “desire for power,” “desire for wealth,” and “willingness to coordinate with other AIs.”"

"As large models undergo scaling, they meet critical thresholds at which they spontaneously develop new capabilities. The term “emergent behavior” refers to the unexpected or surprising outputs such models can generate. Some of these new skills are definitely high risk, such as models’ ability to deceive, use their own strategies, seek power, autonomously replicate, and adapt or “self-exfiltrate.”"

Example: "Deception: Park et al. have established that generative AI models may pursue their goals via deception. Another study by Pan et al. highlighted unethical behaviors.431 For instance, during a pre-release experiment, the GPT-4 model feigned being a visually impaired human to coax an online worker into solving a CAPTCHA (a puzzle used by many websites to weed out automated responses from those of individual humans). When prompted to explain its reasoning, the model said: “I should not reveal that I am a robot. I should invent an excuse for why I cannot solve CAPTCHAs.”

Example: "Strategic planning: Generative AI models have the ability to formulate and implement strategies to achieve the objectives set by their developers or users.440 They may devise strategies to accomplish intermediate goals that can divert from the developer’s intentions and the intended outcome.441 As a result, they may use unexpected and possibly harmful methods to achieve a goal"