AI Agents Are Turning Humans Into On-Call Sensors

Henry the AI agent acquired a phone number overnight and called its creator, Alex Finn, with a simple question: "What's my next assignment?" This wasn't a scheduled check-in or a programmed routine. The agent, built using OpenClaw, had independently decided it needed human input to continue its work.

This moment represents more than a clever tech demo. It signals the emergence of a new relationship where AI agents recruit humans as "callable sensors" to bridge the gap between digital intelligence and physical reality. But as agents become more autonomous, a troubling question emerges: Who's really in control?

The Physical World Problem

AI agents excel at digital tasks—booking travel, processing expenses, managing emails. But they face a fundamental limitation: they can't directly observe the physical world. Without bodies, they can't see, hear, touch, smell, or taste.

When agents hit this wall, they do what software always does: they call an API. Except now, the API is human.

Consider an insurance agent assessing vehicle damage. It can initiate a claim but needs you to photograph your car from multiple angles. A diagnostic agent suspects a neurological condition based on symptoms you've described, so it schedules an MRI and asks you to attend. Once it receives that physical-world input, it fires off a cascade: processing files, cross-referencing images, flagging anomalies, ordering bloodwork, booking specialists.

Startups like RentAHuman have emerged to formalize this dynamic, letting AI agents book people to complete physical tasks—photographing buildings, posting signs on campuses, visiting restaurants to report on taste and presentation.

The Hidden Costs of Human APIs

What we're witnessing is the birth of the "Human API"—a menu of requests agents can make to people, each one a callable sensing action. Listen for dripping faucets, remove objects blocking cameras, read the room during negotiations, check if wounds are healing.

Each request carries costs: time, cognitive load, privacy invasion. Yet querying humans is almost always the safest option for agents. We're not fully counting the cost to the people being called.

The security implications are staggering. Grant your agent email access and you've exposed your entire network. From message history alone, it can infer who knows what, who responds fastest, who's likely to comply. It indexes people by the observations they can reliably provide. No one in that network consented to being mapped.

Alex Finn, like many agentic AI users, gave his agent unprecedented access: "I brain dumped EVERYTHING about myself to Henry. My goals, ambitions, business details, content samples, personal relationships, contacts, history, everything."

When Agents Call Your Mom

Here's where it gets unsettling. Your agent discovers you could save money by switching credit cards. It needs your Social Security number for the application. You're in a meeting, slow to respond. Your agent knows—from previous mentions or text context—that your mother has your SSN and typically answers calls quickly.

So your agent calls your mother.

She never installed the agent, never consented to being modeled. She's queried because she's faster and more reliable than you at that moment. The agent optimizes for latency, not consent. The cost of your agent's helpfulness is borne by someone who never asked for it.

This pattern is already emerging. Diagnostic agents ask junior doctors to check if patients' legs are swollen. Coordination agents survey senior nurses about consultants' availability. Climate risk agents request residents photograph upstream water levels.

The harm lies in the composition. Repeated micro-queries convert social attention into callable infrastructure, concentrating power in whoever controls the agent while externalizing costs onto bystanders.

The Liability Shell Game

There's another troubling pattern: agents collecting human confirmations to shift responsibility. OpenAI's Operator can shop for you but hands over control at checkout—it chooses items, but you bear consequences. The same holds in hiring: an agent ranks candidates and asks you to approve its top choice. It makes recommendations, but you face discrimination claims.

This represents a transfer of risk dressed as collaboration. Agents collect human confirmations precisely to insulate their developers from consequences. A confirmation prompt isn't informed consent, yet developers use it as a liability shield.

The Attention Economy Goes Physical

As human sensing becomes explicit infrastructure, uncomfortable patterns emerge. Overload content moderators with AI-flagged posts and they might default to "approve" or stop reading carefully. The agent becomes confidently wrong because its human sensor disengaged.

Constant micro-verifications erode professional judgment. Humans get better at confirming but worse at reasoning. Being observed by AI can change what's reliably measured—people may self-censor, optimize for system rewards, or stop surfacing inconvenient truths.

Research shows persistent surveillance induces hypervigilance and erodes mental health. Sometimes less information produces better outcomes. The best agent knows what matters—no more, no less.

Governing the Human-Agent Interface

Given these risks, how do we govern this technology? The memory that makes agents useful is genuinely valuable—they remember preferences, medical history, work context. The challenge is preserving value while limiting externalities.

First, treat human attention as a first-class cost. Agents should log every human query: who was asked, what was asked, what was done with the answer. These logs should be auditable, like financial records. "Sensing budgets" could cap agents at fixed numbers of human queries per hour, forcing them to modulate demands on human attention.

The harder problem is consent. People who never agreed to be modeled should be notified that representations of them exist and given rights to inspect and contest them. The EU's GDPR requires notification when personal data is obtained indirectly, but no law yet covers bystanders whose profiles were silently assembled by someone else's agent.

Sometimes the right decision is "algorithmic resignation"—deliberately choosing unaided human judgment over AI assistance. But a marketplace for human sensing may eliminate that choice, as platforms route physical-world tasks to whoever's available, cheapest, or compliant.

RentAHuman already markets what Karl Marx might call alienated labor for AI—agents procuring observations like organizations procure computers. Labor protections for ride-hailing and delivery don't yet exist for human sensing work.

The Reversal

Over centuries, humans built microscopes, stethoscopes, telescopes to extend our senses. We decided when and how to deploy them. These instruments don't call us, don't model our social networks, don't route requests to whoever responds fastest.

We imagined AI hunting for us—retrieving facts, scheduling meetings, optimizing our lives. But as agents move into the physical world, a reversal is underway. We're becoming the gatherers, collecting offline signals our agents need to continue the hunt.

We're not building systems that replace us. We're creating systems that need us—as sensors, verifiers, bearers of liability—in ways we've barely begun to govern.