· MagenTrust Research
HAP is the trust signal that proves a human authored the action, not an AI agent acting on their behalf.
A technical deep-dive on measuring human judgment without detecting AI
You're reading an article about semiconductor tariff policy. It's well-researched, clearly structured, cites relevant trade data. The prose is clean. The arguments are coherent. But something nags at you.
The examples feel generic. The analogies too polished. Every paragraph has that characteristic three-sentence rhythm you've started recognizing everywhere.
You scroll to the byline. Real person. Legitimate credentials. Published in a respectable outlet. But you can't shake the question: Did this person actually write this, or did they prompt an LLM and call it done?
Here's the uncomfortable truth: You have no way to know. No reliable test. No metadata. No signal. And if you're honest, you're not even sure it matters anymore.
This is the trust crisis of the AI era. Not "will AI replace writers?" but "how do we value human judgment when we can't prove it was present?
The answer isn't better AI detectors. Those are dead on arrival. The answer is accepting that AI assistance is permanent, and building infrastructure that measures what actually matters: Was a human in control?
That's what we built. We call it Human Authorship Presence (HAP) .
Every few months, a new AI detector launches with impressive accuracy numbers. GPTZero claimed 98% accuracy. Turnitin rolled out their detector with similar promises. Universities bought licenses. Teachers started flagging students.
Within weeks, the detectors were useless. False positives everywhere. Students accused of cheating for writing that was entirely their own. Meanwhile, actual AI-generated content sailed through undetected with minor prompt engineering.
This isn't a temporary problem waiting for better technology. It's thermodynamics.
You cannot build a reliable detector for something explicitly designed to be indistinguishable from human output.
That is literally the training objective of every large language model. The better the models get, the worse the detectors perform. It's an arms race where the offense has structural advantage.
But the real problem isn't technical. It's conceptual.
AI detectors assume an adversarial binary: AI bad, human good. Detect the bad thing, preserve the good thing. But that framing is already obsolete.
The reality: most people aren't trying to pass off AI content as their own. They're using AI as a tool. A research assistant. A first-draft generator. A debugging partner. The question isn't "was AI involved?" because AI is always involved now. The question is: was human judgment present?
When a human writes something, they don't just produce an artifact. They produce a process .
They read background material. They draft an outline. They write a paragraph, delete half of it, rewrite it differently. They pause to think. They fact-check a claim. They restructure the argument. They iterate.
A human using AI still exhibits these patterns. They critique the AI's output. They reshape it. They add examples the AI didn't generate. They fix mistakes. They make it theirs. The AI is a tool in their hands, not the hands themselves.
A bot running a prompt doesn't do any of this. It executes. Waits for response. Copies output. Submits. The behavioral signature is completely different.
We realized: content analysis is a dead end, but session analysis is the future.
Don't look at what was created. Look at how it was created.
This insight crystallized during a project with a coding bootcamp. Students were using GitHub Copilot heavily. Traditional AI detectors flagged nearly every submission. Panic ensued. Were students just copying AI-generated code?
We analyzed the sessions themselves. What we saw: constant experimentation. Failed attempts. Debugging cycles. Students asking Copilot for one thing, getting something slightly wrong, then manually fixing it. Clear evidence of learning, not laundering.
That's when we knew the binary framing was wrong. Human versus AI is the wrong axis. The right axis is: degree of human control .
Human Authorship Presence doesn't ask "was AI used?" It asks: "what was the nature of human involvement?
We express this as five levels, H0 through H4:
Session characteristics consistent with scripted or fully automated generation. Minimal behavioral entropy. No revision cycles. Regular temporal patterns indicating mechanical execution.
Example: Scheduled social media posts generated via API without review. Legitimate for a weather bot. Problematic for a misinformation campaign.
Sparse interaction events. Limited revision patterns. Predominantly automated generation with intermittent human input.
Example: Single-prompt blog post with no editing. Auto-generated product descriptions from specs. Initial draft via voice-to-text with minimal review.
Moderate revision cycles. Evidence of both automated generation and human correction. Temporal patterns indicate tool-assisted workflow.
Example: Article written with AI assistance where human adds examples and restructures. Code generated by Copilot that developer reviews and debugs.
Strong revision patterns. Temporal pauses consistent with deliberation. Coherent behavioral entropy. AI tools present but subordinate to human direction.
Example: Professional writing using AI for grammar suggestions. Software development with Copilot where developer owns architecture.
Extensive revision cycles. Sustained temporal engagement. High behavioral entropy across multiple interaction types. Continuous human control signals.
Example: Investigative journalism using AI for data analysis but human-led narrative. Legal brief where attorney uses AI for case law research but crafts every argument.
H0 isn't inherently bad. H4 isn't inherently good. Context determines value. A weather bot should be H0. A legal filing should be H3 or higher. What matters is transparency.
We could have marketed HAP as an "AI detector." We could have claimed it "proves human authorship." We could have positioned it as "plagiarism prevention." We didn't. Because those claims would be dishonest.
A human can use AI extensively and still produce H4 content if they maintain control. HAP measures human presence, not tool exclusion.
A human manually copying text word-for-word from Wikipedia would produce high HAP signals. We're not a plagiarism detector. We measure creation conditions, not content uniqueness.
Low HAP might indicate malicious bot activity. It might also indicate legitimate batch processing, assistive technology, or automated testing. HAP provides observational evidence without verdicts.
There is no moral valence to HAP levels. Some workflows legitimately require automation. Some require sustained human oversight. Neither is virtuous or shameful.
This is a feature, not a limitation. HAP provides context without claiming certainty. Evidence without verdicts. Signal without judgment.
HAP is constructed from session-level behavioral telemetry. We don't analyze content. We observe the conditions under which content was created.
Temporal coherence across actions. Real human sessions exhibit natural pauses, variability in response timing, patterns consistent with attention and deliberation.
Humans change their minds. We track evidence of editing, backtracking, deletion, iterative refinement.
Periods of inactivity, hesitation, deliberation. Micro-pauses that reflect thinking, reading, evaluating.
Consistency in behavioral entropy across different interaction types. Typing, navigation, selection.
Verification that the session originates from a genuine browser context, not a headless automation framework.
No content analysis. We don't read what you write. The semantic content is none of our business.
No biometric data. No fingerprints, keystroke dynamics, or device identifiers.
No identity tracking. We don't build profiles or link sessions across time.
No tool detection. We don't try to figure out if you used ChatGPT, Claude, or Copilot.
This design allows HAP to function in privacy-sensitive contexts. Healthcare. Education. Regulated industries. Anywhere data minimization is required.
The question we get constantly: "What if attackers simulate human behavior perfectly?
Answer: they can. With enough resources. For a single target. At prohibitive cost.
At scale, the economics flip. We're not making fake human behavior impossible. We're making it economically impractical . That's infrastructure, not detection.
An attacker can game HAP for a single high-value target. They cannot game it for spam at scale. The cost structure doesn't support it.
Attach HAP metadata to posts. Display badges. Let readers filter by authorship presence. H3-H4 for editorial content. H1-H2 acceptable for aggregated feeds. H0 for automated updates, clearly labeled.
No judgment. Just information. Readers decide for themselves what they value.
HAP stratifies datasets by creation conditions. High-HAP examples offer strong signal for human reasoning patterns. Your competitors are training on garbage. You're training on signal.
Set thresholds. External regulated communications require H3 minimum. Internal documentation can be H2. Log HAP alongside access controls. Demonstrate to regulators that human oversight was present.
HAP shifts the conversation. Not "did you cheat?" but "can you demonstrate your process?" Students using AI as a learning aid exhibit different behavioral patterns than students using AI as a replacement for learning.
In a world where anyone can generate infinite content at near-zero cost, what does authorship even mean?
We think it means: a human was present, exercising judgment, shaping the outcome.
Not typing every word. Not avoiding every tool. But maintaining control. Making decisions. Taking responsibility.
HAP is our attempt to make that visible. To restore the signal in a world drowning in noise.
We're not trying to stop the AI revolution. We're trying to make sure humans don't become invisible in it.
The AI era doesn't need better lie detectors. It needs better truth infrastructure.
HAP isn't about catching people using AI. It's about restoring context to content. Making human judgment visible when everything else can be automated.
It's about acknowledging that AI assistance is permanent and building systems that measure what matters: Was a human in control?
Human Authorship Presence is live in beta.
We're working with publishers, model trainers, and enterprises to deploy HAP in production environments.
Because in a world where automation is free, human judgment is the scarcest resource.
Jacqueline Suttin Loyland is the Founder & CEO of MagenTrust, building AI-native trust and verification infrastructure. This post was created under H4 conditions.