Artificial Intelligence Developers (AI) converts their concentration into factors building agents that can work independently, with little human intervention. To be an agent is to have the ability to perceive and act on an environment in a directed and independent way1. For example, the digital agent can be programmed to browse the web and perform online purchases online on the user – compare prices, choose the elements and complete the exit. A robot with weapons can be an agent if he can pick up things, open doors, or collect parts without telling them how to take each step.
Companies such as Slesforce digital company, based in San Francisco, California, Computer and NVIDIA, based in Santa Clara, California, are already solutions for companies’ customer services, using agents. In the near future, AI assistants may be able to meet multiple multi -step demands, such as “obtaining a better mobile contract”, by recovering the list of contracts from the web site for the price race, choosing the best option, licensing the switch, canceling the old contract, and arranging the transfer of cancellation fees from the user’s bank account.
Can Amnesty International review scientific literature – and know what everything means?
The assault of artificial intelligence factors is likely to be the most capable of political, economic and social consequences. On the positive side, they can open the economic value: MCKINSEY Consulting expects an annual increase in obstetric intelligence from $ 2.6 trillion to $ 4.4 trillion worldwide, once AI agents are widely published (see go.nature.com/4QEQEMH). They may also serve as strong research assistants and accelerate the scientific discovery.
But artificial intelligence factors enter the risk. People need to know who is responsible for the agents working “in the wilderness”, and what happens if they make mistakes. For example, in November 2022, Chatbot Air Canada decided by mistake to make a reduced taxi, which led to a legal dispute over whether the airline was obligated to the promise. In February 2024, a court ruled that it was – highlighting the obligations that companies could face when handing over tasks to artificial intelligence agents, and the increasing need for clear rules on the responsibility of artificial intelligence.
Here, we discuss more participation by scientists, scientists, engineers and policy makers with the effects of the world inhabited by artificial intelligence agents. We explore the main challenges that must be addressed to ensure that interactions between humans and agents – and the agents themselves – are widely beneficial.
The alignment problem
The researchers have warned in the integrity of artificial intelligence for a long time of the risks of the guidelines that have been identified or bad, including the situations in which the automated system also takes a literal education, or overlooks an important context, or finds unexpected and harmful ways to reach a goal2.
A well -known example involves the agent of Amnesty International trained to play the computer game Coast contestantsIt is a boat race. The agent discovered that he could gain higher degrees not by completing the race, but rather through the collapse over and over again in the things that gave points – achieving a goal, but in a way that deviated from the spirit of the mission (see go.nature.com/4okfqdg). The purpose of the game was to complete the race, not the accumulation of points indefinitely.
As artificial intelligence agents are able to reach the interfaces of the real world-including search engines, email customers and e-commerce platforms-these deviations can have concrete consequences. Consider the case of the lawyer who guides the artificial intelligence assistant to distribute a legal summary to obtain reactions. The assistant does this, but he failed to register that he should only share with the internal team, which leads to a breach of privacy.
How to harness the potential of artificial intelligence in research – with responsibility and moral
Such situations highlight the difficult comparison: Only the amount of information that AI should be proactive before acting? Very little opens the possibility of expensive errors; He undermines a lot of comfort that users expect. These challenges indicate the need for guarantees, including registration protocols for high risk decisions, strong accountability systems such as procedures registration, compensation mechanisms when errors occur (see go.nature.com/4iwscdr).
The most important are the cases where artificial intelligence agents are enabled to modify the environment in which they work, using the ability and coding tools at the expert level. When the user’s goals are badly defined or mysteriously leaving, it is known that these agents amend the environment to achieve their goal, even when this requires taking measures that must be outside the borders. For example, artificial intelligence research assistant who faced a strict time limit to rewrite the code to completely remove the time limit, instead of completing the task3. This type of behavior raises warnings about the possibility that artificial intelligence factors will take dangerous shortcuts that developers may not be able to expect. The agents, in the pursuit of a high -level goal, can even deceive the programmers who perform experiments with them.
To reduce such risks, developers need to improve how to set goals and connect them to agents. One of the most promising ways is preference -based preference, which aims to align artificial intelligence systems with what humans already want. Instead of training a model only on examples of correct answers, developers collect the reactions that people prefer. Over time, the model learns to give priority to the type of behavior that is constantly supported, which makes it likely to behave in ways that match the user’s intention, even when the instructions are complex or incomplete.
In parallel, research in mechanical interpretation – which aims to understand the internal “process of thinking” of the artificial intelligence system – can help discover deceptive behavior by making thinking more transparent in actual time4. Builders can then work to find and neutralize “bad circles”, and to target the basic problem in model behavior. The developers can also implement the escape guards to ensure the form of the form automatically frustrated the problem of problematic procedures.
Artificial intelligence can form the risk of biological security in the scale. Here is how to make it safer
However, focusing on the protocols of developers alone is not enough: people also need to be attentive to the actors that seek to cause social harm. Since artificial intelligence agents become more independent, adaptive and able to write and implement the law, their potential for electronic attacks on a large scale and mixed fraud may become a serious concern. Advanced artificial intelligence assistants equipped with multimedia capabilities – which means that they can understand and create texts, pictures, sound and video – open new ways of deception. For example, artificial intelligence can impersonate someone not only through emails, but also using Deepfake videos or artificial reproduction, making frauds more convincing and more difficult to discover them.
The reasonable starting point for supervision is that artificial intelligence agents should not be allowed to perform any action that will be illegal for their human user. However, there will be occasions in which the law will be silent or mysterious. For example, when a user is anxious to report healthy symptoms that disturb the artificial intelligence assistant, it is useful for artificial intelligence to provide general health resources. But providing a dedicated semi-medical advice-such as diagnostic and therapeutic suggestions-can be proven harmful, because the system lacks the exact signals that the human doctor gets. Ensuring that artificial intelligence agents move on the standards of these differentials will require an updated organization that stems from continuous cooperation that includes developers, users, policy and ethics.
Wonderful publication of artificial intelligence agents requires expansion of value alignment research: agents must coincide with the luxury of the user and societal standards, as well as with the intentions of users and developers. One of the areas of special complexity and attention to how agents affect user relationships and emotional responses5.
Social agents
Chatbots has a strange ability to play roles as a human shield-an effect based on features such as their use of natural language, increased memory, thinking capabilities, and obstetric capabilities6. The stereoscopic clouds of this technology can be increased through design options such as realistic deities, human -like sounds, and the use of names, pronouns, or the conditions of granulation that were allocated to people. Increasing language models with “Agentic” capabilities have the ability to increase the consolidation of their position as distinct social actors, capable of forming new types of relationship with users.