Latest

AI Agents

Stanford: Does your AI agent work for you or for whoever created it?

Stanford warns about conflicts of interest in AI agents and proposes a duty of loyalty to protect users

Stanford: Does your AI agent work for you or for whoever created it?
Stanford: Does your AI agent work for you or for whoever created it?

A report by Stanford HAI researchers examines the risks posed by autonomous artificial intelligence agents when the interests of their developers or of the companies deploying them conflict with those of users. The researchers propose establishing a fiduciary framework, especially in high-risk areas such as financial services and healthcare.

The evolution of artificial intelligence from generative tools that react to user instructions toward agentic systems capable of autonomously executing complex multi-step tasks is opening a new front related to conflicts of interest.

The report by Stanford HAI researchers examines the critical risks arising from this new generation of artificial intelligence agents and raises a fundamental question: will an AI agent act in the user's best interest when the interests of its developer or of the company deploying it conflict with the user's own?

The question becomes especially relevant as these systems increase their autonomy and gain access to ever-greater amounts of personal information to carry out the tasks their users entrust to them.

There is currently a lack of regulation and no standardized requirements or mechanisms to disclose the potential conflicts of interest of developers and deployers of these systems, a situation that, according to the report, exposes consumers to harms that may remain invisible.

Against this backdrop, the researchers argue that developers and deployers of AI agents used in consumer-facing contexts should be subject to a fiduciary "duty of loyalty." The proposal is especially important when these systems are involved in regulated or high-risk sectors, such as financial services or healthcare.

From chatbots to autonomous agents

Agentic systems run on the same large language models (LLMs) used by traditional chatbots, but they incorporate advanced capabilities that significantly transform how users interact with digital services.

The fundamental difference lies in their ability to act. While a chatbot responds to a request and waits for the user's next instruction, an agent can autonomously execute previously defined goals, use external tools, query databases and coordinate with other agents with minimal human intervention.

This capability turns agents into systems able to carry out processes made up of different actions and to interact directly with other digital services on behalf of users. AI agents can also maintain memory across different sessions, tools and platforms, overcoming the traditional data silos specific to each application.

Since early 2025, tech giants such as Amazon, Google, Anthropic, OpenAI, Perplexity, Meta and Microsoft have integrated proprietary agents directly into browsers and applications, marking a structural shift in how users operate on the internet.

AI agents and access to sensitive personal data

Agents' ability to act on behalf of users also implies much deeper access to their personal information. To effectively carry out tasks such as booking medical appointments, buying products online or managing personal finances, these systems may need to access especially sensitive information.

This data includes login credentials, financial records, communication histories and real-time control of certain computing environments. This broad, cross-cutting access increases the usefulness of agents, but it also creates privacy risks that may remain invisible to users themselves.

The problem is not limited to the information a person provides directly to the system. An agent may infer highly sensitive information about the user's physical health, possible cognitive decline, financial difficulties or immigration status.

These are categories of information the user may not have explicitly revealed, but that the agent can collect, correlate and infer from data obtained while carrying out completely different and seemingly unrelated tasks.

Risk that agents prioritize platforms' interests

The growing autonomy of these systems raises another fundamental problem: determining who artificial intelligence agents really work for when conflicting interests arise. Without specific consumer-oriented AI regulation and adequate data privacy laws, agents risk becoming extensions of the corporate platforms that develop or deploy them.

In that scenario, the commercial interests of those companies could override the interests of the users the agents supposedly represent. The report precisely raises the need to establish mechanisms to address this situation.

The researchers argue that both developers and deployers of agents should be considered fiduciaries, especially when these systems operate in high-risk contexts. The proposal applies both to the entity responsible for training the underlying models that power an agent and to the company in charge of making that agent available to users.

Duty of loyalty for AI developers and deployers

Treating these companies as fiduciaries would mean imposing a "duty of loyalty" toward users. This duty would require the entities responsible for AI agents to act in their users' benefit within the scope of the task those users have delegated to them, and to do so free of undisclosed conflicts of interest.

In this way, the debate over autonomous agents would no longer be limited to questions about their technical operation, their security or their ability to correctly execute certain tasks. It would also require addressing the relationship of interests between the user who delegates an action, the agent that executes it and the companies that develop or deploy the technology.

The report considers that creating an effective fiduciary framework for artificial intelligence agents requires coordinated action at three levels: the bodies responsible for setting technical standards, federal regulators and Congress.

The proposal seeks to answer one of the questions that will accompany the deployment of autonomous agents: ensuring that tools designed to know their users ever better and act on their behalf do not end up using that privileged position to serve interests other than those of the people who have delegated tasks to them.

Alfredo Collosa

Author

Alfredo Collosa

Alfredo Collosa is a tax specialist. Public accountant. Bachelor's degree in Business Administration. Official Master's in Public Finance and Tax Administration from UNED/IEF in Spain. AMLCA Anti-Money Laundering Certificate (Florida University, US) and member of MetaLaw

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading