Do AI Agents Need a Social Contract?
As AI agents begin to act on behalf of people and institutions, societies need clearer rules for trust, delegation, accountability and consent.
Artificial intelligence is moving from answering questions to performing tasks.
That shift may prove more important than the first wave of generative AI.
A chatbot can produce text. An AI agent can pursue a goal. It can search, compare, organise, respond, book, draft, filter, negotiate, monitor and initiate action. It may interact not only with a user, but with platforms, institutions, databases, payment systems and other agents.
This changes the social meaning of AI.
When systems begin to act on behalf of people, the question is no longer only whether they are useful. The question is whether they are authorised, accountable and trustworthy.
That is why AI agents may require something broader than technical standards or product policies.
They may require a social contract.
The agentic shift
The first public shock of generative AI came from production.
Machines could write essays, generate images, summarise documents, compose code, translate language and imitate styles. The result was a debate about creativity, work, authorship, education and truth.
AI agents add another layer.
They do not only generate outputs. They can be designed to pursue tasks across steps, tools and contexts. They can connect intention to execution.
This is the agentic shift.
This does not mean AI agents are conscious, moral or independent in the human sense.
It means they are entering the space of action.
And once a system acts, society needs to ask different questions.
Why a social contract?
A social contract is not simply a law.
It is an understanding of the rules that make social life possible. It defines expectations, responsibilities, limits and forms of trust. It gives people a way to know what they can expect from others and what others can expect from them.
AI agents complicate this because they act in social environments without being social beings.
They can participate in communication without understanding it as humans do. They can influence decisions without carrying moral responsibility. They can produce consequences without having intentions. They can represent users, organisations or systems without always making the nature of that representation clear.
This creates a gap.
AI agents can enter social life before society has agreed on the terms of their participation.
That gap cannot be solved only by better interfaces.
It requires shared expectations about what agents may do, how they should disclose themselves, who is responsible for their actions and how harm can be corrected.
The problem of delegation
Delegation is the core issue.
When people use AI agents, they do not merely receive information. They allow systems to act within a defined or undefined scope.
That scope matters.
An AI assistant that organises notes is one thing. An agent that sends emails, negotiates prices, evaluates candidates, advises patients, makes financial recommendations or interacts with public services is another.
The more consequential the action, the stronger the need for clarity.
- What exactly has the agent been authorised to do?
- Can the user review the action before it happens?
- Can the agent act autonomously?
- Can the agent spend money, enter agreements or communicate with third parties?
- Can the user revoke authority easily?
- Is there a record of what the agent did and why?
Without clear delegation, AI agents risk becoming systems of vague authority.
They may act in ways that are convenient but socially unstable.
Who is speaking?
Communication is one of the first areas where AI agents will challenge social norms.
If a person sends an AI-drafted message, most people will still treat it as that person’s communication, especially if the person reviewed and approved it.
But what if the agent replies automatically?
What if it uses the person’s tone?
What if the recipient does not know that the response was generated?
What if two agents are communicating with each other while both humans are absent?
In human society, communication carries assumptions of attention, intention and responsibility. AI agents weaken these assumptions.
This does not make AI-mediated communication illegitimate.
But it does mean that disclosure, context and responsibility become essential.
When institutions use agents
The question becomes more serious when institutions use AI agents.
A company can use an agent to answer customers. A hospital can use an agent to triage information. A university can use an agent to support students. A public service can use an agent to guide citizens. A media organisation can use agents to produce or distribute content.
In each case, the agent becomes part of the institution’s public face.
This means that AI agents are not just operational tools. They are institutional actors, even if they are not moral actors.
They shape how an institution listens, explains, prioritises, apologises, refuses and corrects itself.
An unfair agent becomes an unfair institution. An opaque agent becomes an opaque institution. An irresponsible agent becomes an irresponsible institution.
Institutions should not be allowed to use AI agents as a way to reduce accountability while increasing control.
A social contract for AI agents must therefore include institutional obligations: transparency, contestability, auditability, human escalation and clear responsibility.
Minimum rules for agentic systems
A social contract for AI agents does not need to begin with grand theory.
It can begin with minimum rules.
These rules are not anti-innovation.
They are the conditions under which AI agents can become socially usable.
A powerful agent without clear rules may be impressive as technology, but fragile as social infrastructure.
Trust is not automation
Many AI systems are designed to feel smooth.
They answer quickly. They reduce friction. They simplify choices. They appear confident. They make complex tasks feel easy.
But trust is not the same as ease.
A system can be convenient and still be untrustworthy. It can be efficient and still be unfair. It can be persuasive and still be wrong. It can feel personal and still be manipulative.
Trust requires more than performance.
It requires visibility, limits, explanation, reversibility and accountability.
In an agent-mediated world, people should not have to guess whether a system is acting as a tool, a representative, a filter, a seller, a gatekeeper or a decision-maker.
If AI agents are to participate in social life, their role must be legible.
Rights, responsibilities and refusal
A social contract for AI agents should also include the right to refuse certain forms of agentic mediation.
People should not be forced to interact only with automated agents in consequential situations. They should have meaningful access to human review when dealing with healthcare, education, employment, finance, public services, legal matters or other high-impact contexts.
This is not nostalgia for human inefficiency.
It is a recognition that some interactions carry dignity, vulnerability and responsibility in ways that cannot be reduced to automation.
The right to human escalation may become one of the basic protections of the AI age.
At the same time, users who deploy personal AI agents should also carry responsibilities.
- They should not use agents to deceive others about presence or identity.
- They should not delegate consequential action without understanding the scope.
- They should not hide behind agents to avoid responsibility.
- They should not use synthetic communication to manipulate trust.
A social contract must apply not only to companies and governments, but also to everyday users.
The social contract we need
AI agents will become ordinary because they are useful.
They will reduce administrative burden, support accessibility, help people manage information, assist organisations and create new forms of productivity.
The issue is not whether they should exist.
The issue is under what terms they should participate in social life.
A serious social contract for AI agents would begin with a simple premise:
No system should act in the name of a person or institution without clear authority, visible limits and accountable responsibility.
AI agents do not need rights before society understands their role.
But people need rights in relation to AI agents.
They need to know when agents are present, what they can do, who controls them, how decisions can be challenged and who remains responsible when things go wrong.
The future of AI agents is not only a technical question. It is a question of social order.