Abstract
This position paper proposes a compact conceptual model of human–AI work: SI = −DK. Biological intelligence (BI) is treated as a switch: either a human mind is actively engaged in the loop, or it is not. The Dunning-Kruger term (DK) is treated as a polarity: either overconfidence rises, or it falls. When BI = 0, AI alone drives DK upward: fluent, confident output that its user cannot evaluate. When BI = 1, BI + AI produces superintelligence (SI), and the same polarity reverses. The model does not claim that humans are smarter than machines, or the reverse. It claims that the human is the sign of the outcome. It is presented as a conceptual model built from field experience, not as an empirical law, and the article states what evidence would support or refute it.
From Three Articles to One Line
In August, I published a three-part series on working effectively with AI.
Part One argued for redundancy: one model builds, another inspects. Part Two argued for choosing the right model for the task. Part Three argued that the user is part of the system, and that human judgment is the final redundancy layer.
Since then, I have kept working. More models, more platforms, more real engineering work, and more mistakes, mine and the machines'.
Over those months, the lessons kept collapsing into the same observation. The quality of AI-assisted work was not decided mainly by which model I used. It was decided by whether a human mind was actually engaged while the model worked.
Not present. Engaged.
That observation eventually fit on one line of a notebook:
SI = −DK
Superintelligence is the negative of the Dunning-Kruger effect.
The rest of this article explains what that line means, how it is built, and, just as important, what it does not claim.
The Terms
An equation is only as good as its definitions. These four terms are defined precisely, because a reader who guesses their meaning will find contradictions that are not there.
BI — Biological Intelligence
A binary switch: zero or one.
Is a human mind actively engaged in the loop, setting the objective, questioning the output, verifying it, and able to recognize a convincing wrong answer?
AI — Artificial Intelligence
The model's contribution: text, code, analysis, design, or other generated work.
SI — Superintelligence
The combined result of biological and artificial intelligence when the human is engaged: work neither could produce alone.
DK — Dunning-Kruger Polarity
A direction, positive or negative.
Does the gap between confidence and actual competence grow or shrink?
Two choices in this table matter.
BI is binary. This is not a measure of how intelligent the human is. A brilliant person who copies an answer without reading it has BI = 0 for that task. An ordinary person who checks the result against reality has BI = 1. Engagement is a state of the workflow, not a trait of the person.
DK is a polarity, not a magnitude. It does not say how much overconfidence exists. It says which way it is moving. That is why it can be negated, and why it is never added to the other terms as a quantity.
The switch and the polarity are tied together by one relationship:
DK = 1 − 2 × BI
Two States, One Machine
The model has exactly two states. The AI is the same in both. Only the switch changes.
State One: BI = 0 — The Human Is Absent
0 + AI = AI, DK = +1
The user asks, the AI answers, and the answer is accepted.
The output is fluent. It is formatted. It sounds authoritative. The user now holds work that looks expert, and feels more competent for holding it. But the user cannot tell whether it is right, because the user never engaged with it.
Confidence has gone up. Competence has not. That is the Dunning-Kruger direction: the gap between what someone believes they know and what they actually know grows. Here, the AI is not reducing ignorance. It is dressing it.
State Two: BI = 1 — The Human Is Engaged
1 + AI = SI, DK = −1
The human defines the objective, supplies the context, questions the output, checks it against evidence, and rejects what does not hold.
Now every confident claim from the AI meets a mind that asks: how do you know? Errors surface. Assumptions get tested. The human learns where the model is strong and where it bluffs, and the model's speed and breadth extend what the human can reach.
Confidence and competence move together. The gap shrinks. The polarity reverses.
That combined state is what this article calls superintelligence. Not a machine that has surpassed humans, but a human–machine system that outperforms either part working alone.
The Line, Read in Full
In State 2, DK = −1 and the result is SI. That is the notebook line:
SI = −DK
Superintelligence is not the absence of the Dunning-Kruger effect. It is its exact opposite direction.
Why the Human Flips the Sign
The interesting part of the model is not that the human adds value. That is obvious. The interesting part is that removing the human does not merely subtract value. It reverses direction.
Why would the same AI make one person sharper and another person more overconfident?
Because AI output carries the appearance of competence. Fluency, structure and confidence are exactly the signals humans use to judge expertise. An AI produces those signals whether or not the content is correct. A human who is not engaged reads the signals and stops there.
Because the Dunning-Kruger effect is a failure of self-evaluation. People who lack the knowledge to do a task also lack the knowledge to judge whether it was done well. AI does not repair that gap on its own. It hands the user a finished-looking answer, so the user never meets the moment of difficulty that would have revealed the gap.
Because engagement creates feedback. When the human questions and verifies, every error the AI makes becomes a lesson instead of a hidden liability. The user's model of reality is corrected continuously. Confidence becomes calibrated, because it is now earned against evidence.
In engineering terms, BI is the feedback path. Without it, the system runs open-loop: the output looks stable while error accumulates silently. With it, the loop closes, and the same gain that amplified error now amplifies correction.
The AI is the amplifier. The human decides what it amplifies.
What I Saw in the Field
This model did not come from theory. It came from watching the same tools produce opposite results.
Field Case One — The Website That Worked
In Part One of the AI series, I described a production website that loaded correctly and appeared healthy. Each AI-assisted change satisfied the visible request. Underneath, duplicated routes, obsolete configuration and stale files had accumulated. As long as the visible result was accepted, the project felt finished and under control. That was BI = 0: confidence rising while the system quietly degraded. The moment the actual files were inspected, the polarity reversed. The defects became visible, the architecture was cleaned, and our understanding of our own system improved.
Field Case Two — A Convincing Answer Is Not a Verified Configuration
A more technical case emerged while investigating RandomX memory behavior on an Intel Mac. Huge Pages were a specific part of the work. They were not an abstract optimization or a decorative checkbox. The question was whether the operating system and miner were actually using the intended memory configuration, and whether any claimed improvement could survive a real test.
An AI could propose a sequence of plausible actions. It could explain memory allocation, suggest a change, and describe what a successful result should look like. That is useful. But a coherent proposal is not proof that the machine is in the proposed state. Especially in performance engineering, the difference between a plausible explanation and an observed condition matters more than the confidence of the explanation.
I challenged the AI's proposed approach during this investigation. The correction came from the human side of the loop: the answer had to conform to the machine, its operating-system behavior, and the actual test evidence—not the other way around. The plan could not be accepted merely because its language sounded technically complete.
The eventual investigation included a 256 MB scratch allocation and a cold reboot, after which a test reported 100% Huge Pages. That observed result mattered. So did the discipline of treating it as a configuration observation, not as permission to attribute every later performance number to a single change. Different measurements answer different questions.
The bad outcome in a case like this is easy to imagine but must not be confused with what was measured: accepting an AI's first explanation would have turned a proposed configuration into an assumed fact. The useful outcome came from refusing to close the investigation before evidence was available. BI was not the act of asking AI to do more. It was the act of denying the AI the final word over a physical system it could not inspect for itself.
Field Case Three — When the Correction Came Back the Other Way
The same Intel Mac investigation also showed why the relationship cannot be reduced to 'the human is right and the AI is wrong.' After I corrected the AI's approach, the AI identified a flaw in how I was testing. The direction of correction reversed. The human challenged the machine's proposal; the machine challenged the human's evidence.
That second intervention is the more important half of the case. If the story ended with the user proving the assistant wrong, it would establish only that humans can catch machine errors. That is necessary, but it is not the combined intelligence described by the model. The stronger result appears when both sides can expose weaknesses in the other's reasoning and the workflow is designed to tolerate those corrections.
A benchmark is not an argument won by either participant. It is a measurement that must be interpreted with enough care that the machine, the configuration, the test conditions, and the reported result refer to the same thing. A discovered test flaw changes what can responsibly be concluded, even when the result initially appears to confirm a preferred explanation.
Later, the Intel Mac produced sustained RandomX measurements above 3.0 kH/s at roughly 50% CPU use over a 24-hour run. That is an operating observation, not proof by itself that Huge Pages caused the improvement. The distinction is the point: a number gains engineering value only when the conditions under which it was obtained are understood, and a convincing number should be open to challenge just as a convincing paragraph should be.
The reciprocal correction is what made the work better. Neither 'trust the AI' nor 'trust the engineer' was a sufficient rule. The rule was to test the proposal, inspect the observation, and allow either participant to reveal an error. The machine helped when it challenged a human assumption. The human helped when he refused to take the machine's confident explanation as evidence. That is what an engaged feedback loop looks like in practice.
The pattern across these cases is the same. The model did not change between the bad outcome and the good one. The human did.
What This Model Is Not
An article about overconfidence should not be overconfident. So the limits are stated here, before a critic has to state them.
It is not a physical law. E = mc² has measurable quantities, units, and decades of experimental confirmation. SI = −DK is a conceptual model. Its terms are defined states and directions, not measured magnitudes.
It should not be read as ordinary algebra. DK is a polarity. If it is treated as a number to be added to AI, the equations can be forced into nonsense. The only operations the model claims are the switch (BI = 0 or 1) and the sign reversal.
It is a simplification. Real engagement is not perfectly binary. A person can be engaged on one part of a task and absent on another. The binary switch is useful because it forces an honest question for each piece of work: was a human mind actually in the loop here, or not?
It does not say AI is bad. In State 2, AI is the reason the result exceeds what the human could do alone. The model is a warning about how AI is used, not about AI itself.
How It Could Be Tested
A model is worth more when it can be proven wrong. This one makes a testable prediction:
- Give two groups the same AI tool and the same task in a domain they do not master.
- Group A accepts the AI's output. Group B must question it and verify key claims before submitting.
- Measure each person's confidence in their result, and the result's actual accuracy.
The model predicts that Group A's confidence will exceed its accuracy (DK = +1), while Group B's confidence and accuracy will move closer together (DK = −1), even though both used the same AI.
If both groups show the same gap, the model is wrong, and I would want to know.
The Human Is the Sign
The public debate about AI often asks how intelligent the machines are becoming.
That question matters. But for the person using the tool today, a more practical question decides the outcome: is my mind in the loop, or have I handed it over?
The same model, on the same day, can make one person wiser and another more confidently wrong. The difference is not in the model's weights. It is in the human's engagement.
With the human absent, AI amplifies the illusion of competence.
With the human engaged, AI amplifies competence itself.
SI = −DK
Superintelligence will not arrive only when machines surpass us. For each of us, it arrives when we stay engaged enough to turn the sign.
About the Author
A. Remani
A. Remani is the Founder, CEO, and Head of Engineering at I2PS Engineering. His work combines engineering, product development, software, field operations, prototyping, and the practical integration of AI into small-team workflows. His writing focuses on responsible engineering, verification, human judgment, and systems that remain understandable after repeated change.
Closing Reflection
AI does not decide whether it makes you smarter or more overconfident. You do, every time you choose to engage or to accept.
