Human control verification is the process of establishing evidence that a real human is directly controlling a specific digital interaction at the time it happens. It goes beyond asking whether a human exists, whether they logged in or whether they are visible on camera. It asks whether the actions taken in this session, on this form, in this approval or in this assessment were actually directed by that human, rather than by another person, a remote operator or an AI system acting in their place.
That definition matters because most of the language we use about humans and AI was built for a different question. "Human in the loop" is one of the most widely used phrases in AI governance, and it is useful. But it describes how a process is designed. It does not tell a bank, a hiring platform or a marketplace whether a human was actually in control of the interaction in front of them. Human control verification is the missing evidence layer between those two ideas.
Why "who logged in" is no longer the same as "who is acting"
For most of the history of the web, a reasonable assumption held: if an account was authenticated and activity followed, the account holder was probably the one acting. That assumption shaped password policies, session cookies, fraud models and audit logs.
It no longer holds reliably. After a person authenticates, they can:
- hand the session to an AI agent that operates the browser
- paste answers produced by an AI model into an assessment
- allow a remote operator to control the device
- pass the device to another person
- approve a transaction they did not read, because an assistant prepared it
Some of these are entirely legitimate. Others are not. The point is that a successful login says almost nothing about which of them is happening. As we argue in Authentication Is No Longer Enough, authentication answers "can this entity access this account?" while the question that increasingly matters is "who or what is actually performing this action right now?"
Human control verification is a response to that gap.
Human in the loop vs human on the loop vs human out of the loop
Before going further, it helps to be precise about the established vocabulary, because "human control" is often confused with it.
What is human in the loop?
Human in the loop (HITL) describes a system in which a human must take part in a decision before the system acts. The system may recommend, draft, rank or flag, but a person reviews and approves (or rejects) each relevant decision. A content moderation tool that queues flagged posts for a human reviewer, or an AI model that drafts a loan decision that an officer must sign off, are typical examples. In machine learning, the term is also used for workflows where humans label data or correct model outputs that feed back into training.
What is human on the loop?
Human on the loop (HOTL) describes a system that acts on its own while a human supervises. The human does not approve each action but monitors the system and can intervene, override or stop it. Automated trading systems with a supervising desk, or industrial automation with an operator who can halt a process, fit this model.
What is human out of the loop?
Human out of the loop (HOOTL) describes a system that operates autonomously, with no human approving or supervising decisions in real time. Humans may have set the goals, policies or limits in advance, and may review outcomes afterwards, but they are not part of the live decision.
Comparison table
| Model | Role of the human | When the human acts | Typical example | What it tells you |
|---|---|---|---|---|
| Human in the loop | Approves or rejects each relevant decision | Before the system acts | Reviewer approves AI-flagged content | A human step exists in the process |
| Human on the loop | Supervises and can intervene or stop | During operation, by exception | Operator monitoring an automated system | A human could intervene |
| Human out of the loop | Sets goals and limits in advance | Before deployment and in later review | Fully autonomous agent within a policy | No live human decision is expected |
These three models are widely used in AI governance and system design. Regulation increasingly reflects them too. For example, Article 14 of the EU AI Act requires high-risk AI systems to be designed so that they can be effectively overseen by natural persons while in use, including the ability to override or stop the system.
The gap: a process design is not evidence
Here is the distinction this essay is built on.
Human in the loop is a claim about how a workflow is designed. Human control verification is evidence about what actually happened in a specific interaction.
A process can be designed with a human approval step, and the approval can still be given by:
- the intended human, reading carefully and deciding
- the intended human, clicking through without engagement
- a different human using the same credentials
- an AI agent operating the approval interface
- a script replaying a previously captured session
From the system's point of view, all five may produce the same "approved" event. The loop exists on paper. Whether a human was actually in control of that approval is a separate question, and today most systems have no reliable way to answer it.
This gap is growing as AI agents learn to operate ordinary software. An agent that can use a browser can also click "I have reviewed this". Without evidence, "human in the loop" risks becoming a label rather than a safeguard.
The same applies in the other direction. A service that wants to permit AI assistance, or wants to accept actions from authorised agents, also needs to know which mode it is dealing with. Human control verification is not about forcing everything to be human. It is about knowing.
What human control verification establishes
At Proof of Actor we treat human control as one layer in a broader Trust Stack. It sits alongside, but is distinct from, several related questions:
| Layer | Question | What it does not establish on its own |
|---|---|---|
| Personhood | Is there a real human behind this interaction? | Who they are, or whether they are acting now |
| Identity | Is this human who they claim to be? | Whether they are present or in control |
| Presence | Is the human actually present during this interaction? | Whether they are directing the actions |
| Actor continuity | Is the same actor continuing throughout the session? | Whether that actor is human or in control |
| Human control | Is the human actually controlling the interaction? | Who authorised an agent, if one is acting |
| AI participation | Is AI materially participating? | Whether that participation is acceptable |
Authentication proves access. Identity proves identity. Presence proves presence. None of them necessarily proves control. Human control verification is designed to fill that specific gap.
In practice, the question breaks down into a few parts:
- Is a human present? Not fifteen minutes ago at login, but during the interaction itself. See Proof of Human for how we think about presence and personhood.
- Is it the same human throughout? Continuity matters for hand-offs, substitution and session takeover.
- Is that human directing the actions? This is the core of Proof of Control.
- Is AI materially involved, and how? Assistance is often fine. Invisible substitution usually is not.
We deliberately describe what is established, not how. Signal design, weighting and anti-evasion methods are not something we publish, for the same reason fraud and security teams generally do not.
Control is a confidence, not a verdict
A responsible human control verification system should never claim certainty it does not have. Human behaviour varies enormously across people, devices, accessibility needs and contexts. AI assistance ranges from spell-checking to wholesale generation. A binary "human or not" output would be both inaccurate and unfair.
Proof of Actor is designed to express human control as a confidence signal with supporting evidence, so the relying service can apply its own policy. Within the broader Actor model, an interaction might be described as one of four states:
- Human: a human appears to be directly controlling the interaction.
- Human + AI: a human remains responsible, but AI materially assists.
- Authorised Agent: an AI agent acts within authority delegated by an identified human or organisation.
- Unverified Actor: the service cannot establish the actor, human control or accountable principal with sufficient confidence. This means "trust relationship unknown", not "malicious".
We explore this model in more depth in Human, Human + AI, Agent: A New Identity Model.
What a human control result might look like
The output of a human control check should be something a service can act on and record. In our architecture this takes the form of an Actor Attestation, with a human-readable form we call an Actor Trust Receipt.
The following is a conceptual illustration only, not a live format:
ACTOR TRUST RECEIPT (conceptual illustration)
Actor Human
Presence Continuous
Actor continuity High confidence
Human control High confidence
AI involvement Low
Identity Not requested
Issued 12:41 UTC
Note the line "Identity: Not requested". Many services that need evidence of human control do not need to know who the human is. A community platform may only need "a real human was in control of this post". A hiring platform may need "the same verified candidate remained present and in control during the assessment" without permanently storing government identity data. Our guiding principle is to prove what matters without revealing what doesn't. Our privacy page sets out how we are approaching this.
A service would then express its own requirement as a policy. Again, this is an illustrative Proposed API sketch, and the design is evolving:
// Proposed API - illustrative only, not available today
policy.require({
actor: "human",
presence: "continuous",
control: 0.90,
aiAssistance: "low"
});
The key principle is that Proof of Actor is designed to supply trust evidence, and the relying service decides what is acceptable.
Where human control verification matters
Human control is not equally important everywhere. Commenting on a forum and approving a large transfer warrant very different levels of evidence. Contexts where it tends to matter most include:
- Assessments and interviews. The question "did this candidate genuinely perform this assessment?" is unusually clear, which is why we see hiring as an important early proving ground.
- Financial approvals and high-risk account actions. Changing payee details, adding administrators or authorising major payments may warrant evidence of current human control or explicit human intent. See financial services.
- Remote meetings. Establishing that the expected participant is present and remains present, regardless of what the video stream looks like. See video calls.
- Governance checkpoints in AI workflows. Where an organisation has designed a human approval step, human control verification can provide evidence that the step was actually performed by a human.
That last case is the bridge between the governance vocabulary of human in the loop and the evidence vocabulary of human control. A loop that is designed in is a policy. A loop that is evidenced is a record.
What human control verification is not
It is not identity verification, and it is not AI detection: AI assistance can be legitimate, and the goal is to make it visible, not to treat it as wrongdoing. Nor is it a CAPTCHA, since a single challenge at session start cannot establish continuous control, as we discuss in Why CAPTCHAs Cannot Solve the AI Agent Problem.
Frequently asked questions
What is human control verification in simple terms?
It is a way of producing evidence that a real human is actually controlling a specific digital interaction, such as an approval, an assessment or a transaction, at the time it happens. It is different from checking who logged in or whether a human exists.
What is the difference between human in the loop and human on the loop?
In a human in the loop system, a person must approve or take part in each relevant decision before the system acts. In a human on the loop system, the system acts on its own while a person supervises and can intervene or stop it. Human out of the loop systems operate autonomously within limits set in advance.
Does human in the loop guarantee a human made the decision?
No. Human in the loop describes a process design. Whether a human was actually present and in control when a particular approval was given is a separate question that requires evidence. Human control verification is designed to provide that evidence.
How is human control different from human presence verification?
Presence verification asks whether a human is there during the interaction. Human control verification asks whether that human is actually directing what happens. A person can be present while an AI agent or another person operates the session.
Does human control verification require identity data?
Not necessarily. Many services only need to know that a human was in control, not who that human is. Proof of Actor is being designed to support claims such as "verified human, identity undisclosed".
Building the evidence layer
Human in the loop gave organisations a way to describe responsible AI processes. The next step is being able to show, for a specific interaction, that the human the process depends on was actually in control. Proof of Actor is building that evidence layer as part of a broader trust infrastructure for humans and AI agents, and you can read more about the approach on how it works. We are currently in development and working with early design partners. If your organisation depends on interactions where human control matters, we would welcome you to join the Developer Preview or become a design partner.
