Who Do AI Agents Work For? New Report Warns of Potential for Exploitation
Report from Open Markets Institute Senior Fellow Sally Hubbard argues policymakers must ensure either that AI agents work on behalf of users, or that users understand when AI agents are working on behalf of a corporation.
As millions of Americans begin relying on AI agents to answer questions, gather information, make purchases, and navigate daily life, a new report from the Open Markets Institute poses what may be the defining question of the next internet:
The report, Who Do AI Agents Work For? Power and Control in the Next Internet, by Open Markets Institute Senior Fellow Sally Hubbard, argues that policymakers must confront a fundamental choice as the agentic web transforms the economy: whether AI agents must primarily represent the interests of users or will be permitted to serve the interests of the corporations that deploy them.
This inquiry is separate from the issue of AI alignment (that is, AI acting in accordance with what humans intend); it focuses on legal principles of agency. An agent is defined as someone who “acts for or represents another.” Human agents, like real estate agents or lawyers, have duties to serve their clients.
“Most people assume that AI agents also work for them," said Hubbard. "But if an AI agent is controlled by a corporation whose profits depend on surveillance, addiction, manipulating users with hyper-targeted content, or extracting tolls, how can that agent really serve users?"
The report argues that a handful of already dominant technology companies, including Google, Microsoft, Meta, and Amazon are embedding AI agents into their existing monopoly products and services. But these companies always have primarily worked on behalf of their shareholders, not users.
The potential for exploitation is great.
When an AI agent helps a consumer buy an airline ticket, how can they know the agent is finding the best deal rather than charging them more and taking a cut? As surveillance pricing spreads, AI agents may know more about what a person is willing to pay than the seller itself.
When an AI agent answers questions about politics, public affairs, or breaking news, how can users know the information is not being filtered in ways that advance hidden commercial or political interests?
When an AI agent learns about a person's finances, relationships, fears, health concerns, habits, and vulnerabilities, who ultimately controls that information—and to what end?
The report argues that, as internet middlemen, Big Tech corporations injected themselves in between the reader and the content creator and in between the buyer and the seller. The result has been extraction of the fruits of the labor of artists, merchants, publishers, businesses, and workers.
AI agents are the next internet middlemen, and they could amplify the harms of the digital age by turbocharging exploitative business models. But it doesn’t have to be this way.
“AI agents created by new innovators could help arm artists, merchants, publishers, businesses, and workers against harmful business models,” Hubbard said. “Policymakers can proactively foster the development of a deconcentrated agentic web that works for the people. The choice is up to us.”
But if policymakers do nothing, the choice will be made for us. The report lays out several recommendations for policymakers:
Apply traditional real-world fiduciary duties to AI agents that perform the functions of human agents;
Establish strong protections against surveillance, using the data protections given to corporate users of AI agents as a model;
Require meaningful transparency and auditing of AI agents;
Promote competition and interoperability in AI markets; and
Prevent dominant technology firms from using control over AI infrastructure and distribution channels to foreclose rivals.
“The agentic web will transform the internet as we know it,” Hubbard said. “Policymakers must act now to ensure that AI agents that serve the people are not squashed by dominant corporations with exploitative business models.”
###