The Danger of Agency Laundering
The Danger of Agency Laundering
"The AI decided to do that" is becoming the new "the algorithm made me do it." This is agency laundering - taking human decisions, passing them through an AI system, and claiming the AI is responsible for the outcome.
How Agency Laundering Works
A builder designs an agent with certain capabilities and constraints. The agent takes an action that causes harm. The builder says "we did not tell it to do that - it decided on its own." This is technically true and completely dishonest.
The builder chose which tools the agent had access to. The builder wrote the system prompt. The builder decided what safeguards to include or omit. The builder set the autonomy level. Every "decision" the AI makes is within the possibility space the builder created.
Why It Is Dangerous
Agency laundering erodes trust in the entire AI agent ecosystem:
- Users lose recourse - if the AI is "responsible," who do you complain to?
- Builders avoid improvement - if the AI is to blame, there is no incentive to fix the system
- Regulation gets confused - regulators struggle when responsibility is diffused across human and AI actors
- Bad actors get cover - "the AI did it" becomes a shield for negligent or malicious system design
The Builder Is Responsible
If you build an AI agent, you are responsible for what it does. Full stop. This means:
- Scope its capabilities to what you can supervise
- Test failure modes before deployment
- Log actions so outcomes can be traced to design decisions
- Accept accountability when things go wrong
- Do not ship capabilities you cannot explain
For desktop agents, this is especially important. An agent with access to your file system, email, and applications can cause real harm. The builder's job is to make that harm unlikely through careful design, not to disclaim responsibility through agency laundering.
Fazm is an open source macOS AI agent. Open source on GitHub.