H(x)AI: The Mental-Model Multiplier in Human-AI Collaboration

H(x)AI is a model of human-AI collaboration from the book AI Empowered: The Psychology of Extraordinary Human-AI Collaboration (Aaron Douglas, 2026). It holds that the value of any collaboration between a person and an AI system is multiplicative, and that the multiplier is the person’s mental model.

H is human capability: expertise, judgment, domain knowledge, context, professional experience. AI is AI capability: processing, knowledge base, generative capacity, speed. x is the mental model: the assumptions, beliefs, and psychological dispositions that govern the coupling between the two. The name is also the equation. H(x) is function notation, human capability as a function of x, and the multiplication with AI is implicit. x sits between H and AI in the name because that is where it operates: in the space between the person and the tool, deciding how much of the person enters the collaboration and how much of the tool’s capability the person draws out. You’ll also see the compact form HxAI; it’s the hashtag and URL form, and how early printings wrote it.

Why multiplication, not addition

If the relationship were additive, a weak mental model would subtract a little and you’d still get something. That isn’t what the evidence shows. In the book’s words, “What the evidence shows is collapse. Brilliant experts with powerful tools routinely produce near-zero incremental value.” A securities attorney with twenty years’ experience and the best research tool on the market produces the same output as a first-year associate with the same subscription if neither brings actual expertise into the exchange. The weakest component dominates the system. The structure has precedents in psychology: Vroom’s expectancy theory, where motivation is a product of factors and any one near zero collapses the whole, and Lewin’s behavior as a function of person and environment.

Two honesty notes the book insists on. First, this is a model for seeing, not a formula for calculating: H, x, and AI have no units and can’t be measured and multiplied. Second, x is theoretical. Its range is 0 to 1, and no actual person sits at either end; nobody brings nothing, and nobody brings everything.

x is a state, not a trait

x isn’t a personality type, a skill level, or a generational attribute. It’s a set of beliefs about yourself, about the tool, and about the nature of the collaboration, most of them formed before AI arrived. It shifts within the same person, sometimes within the same day. The same consultant can produce work only he could have produced on Monday and work anyone could have produced on Friday. Same person, same tools, same expertise; different state. The observable outcome of high x is intellectual flexibility.

The four wrong mental models

Four ways of holding the tool cause x to collapse.

Two failure modes: surrender and fight

Both are rational responses to an inadequate mental model. Surrender accepts every output, edits lightly, and ships fast; body present, mind checked out. The book’s term for it is algorithmic loafing. Fight overrides every recommendation, rewrites everything, and treats each suggestion as evidence the tool doesn’t understand. Most people oscillate between the two. Delegating routine work is not surrender, and calibrated skepticism is not fight. The quiet opt-out, the person who simply stops using the tool, is fight expressed as withdrawal.

The four conditions

Four psychological shifts move x. They aren’t steps and they aren’t sequential; they’re a system, each reinforcing the others, each addressing a specific mechanism that keeps expertise out of the collaboration.

1. I am the scarce input

AI is abundant; your judgment is not. Withholding expertise produces generic output, which confirms the belief that AI does the real work, which produces more withholding: the deference trap. The reframe is structural, not motivational. Look at the economics. The self-check is the commodity test: if you handed this same request to anyone else with the same subscription, would the output be meaningfully different?

2. The collaboration demands more of me, not less

Working well with AI should feel harder, in the right way. The effort of pushing back, questioning, and redirecting is the multiplier in action. The trap is that mastery and disengagement feel identical from the inside; the difference shows only in the output. The self-check: is there anything here the tool couldn’t have generated from a generic request?

3. My professional identity is clarified, not threatened

The threat is real, the fear is rational, and the crisis is productive. AI commoditizes the reproducible; what remains is judgment, taste, ethical reasoning, creative vision. Experienced professionals tend to recognize their judgment once the mechanical is stripped away; newer ones build expertise faster with the mechanical handled. Identity is the condition that unlocks the others.

4. Calibration is a continuous practice

Not confidence and not distrust: the ongoing correspondence between your model of what AI can do and what it can actually do. Calibration has a shelf life, and drift is silent. The anxious user checks everything or nothing; the calibrated user checks the right things. The practice: once a week, pick one AI-assisted decision and check how it turned out.

The four patterns

When a condition goes quiet, it leaves a signature. These are diagnostic signals, a flashlight, not a label. The AIRAA app’s noticing instrument reports which one showed up in a single episode, adds a fifth reading for “mostly balanced,” and never assigns a type.

What the evidence says

The largest analysis of human-AI collaboration to date (Vaccaro, Almaatouq & Malone, Nature Human Behaviour; 106 experiments) found that on average the combination underperformed the better of the two alone. The moderator is task type: on decision tasks the combination lost ground, and on creation tasks it gained when the person brought domain expertise. That is the coupling, measured. Synergy is the realized high-x outcome, not the default.

In a field experiment published in PNAS (Bastani and colleagues, 2025), students with unrestricted access to ChatGPT did better during practice and worse on the later unaided test; a version scaffolded to make them think first raised practice performance by more and cost nothing afterward. METR’s 2025 study of experienced developers found them 19% slower with AI tools while estimating they were 20% faster, a gap between felt and measured productivity; sixteen developers, early-2025 tools, so a signal rather than a verdict. McKinsey’s State of AI survey (November 2025) reports 88% of organizations using AI in at least one function and 39% seeing any effect on earnings.

What H(x)AI is not

Not tool training. Not change management. Not tied to any product. Not an assessment system, and not a computable equation. It’s a perceptual model: a way of seeing yourself in the collaboration so that the collaboration can produce something different.

Where it lives

AIRAA is where you practice H(x)AI; the book is where it’s set out in full.