There is a classic piece of Artificial Intelligence wisdom that is now front and center.
Goodhart’s law looms large for autonomous artificial intelligence.
Charles Goodhart was an economist who worked at the Bank of England in 1975. This was around the last time disastrous monetary policy flipped the world into economic turmoil, off the back of the first OPEC Oil Embargo and the 1974 bear market.
It is my belief that we are headed straight back into a similar disaster, for exactly the same reason. With this in mind we revisit an observation called Goodhart’s Law.
Goodhart’s Law in Management
Goodhart’s law states that:
“whenever a government seeks to rely on a previously observed statistical regularity for control purposes, that regularity will collapse”
You can find it as a footnote in an RBA conference paper: Problems of monetary management: the U.K. experience, Goodhart, C. (1975).
This observation is normally simplified to:
“When a measure becomes a target, it ceases to be a good measure.”
It is generally viewed as a cautionary tale about bad KPIs.
When managers are rewarded on EPS growth they manage earnings.
When teachers are rewarded on their pupil’s test scores, they teach the test.
When mathematicians are rewarded for proofs, they produce 250,000 pages.
In truth, what we wanted was:
Sustainable growth: not hollow logs, buy backs, and off-balance sheet leases.
Students with an awareness and understanding of what their subject is about.
Higher level understanding as a launching pad towards entirely new mathematics.
You can have all of that, but not by setting narrow objective functions to optimize like EPS growth, PISA scores, or the number of lines of Lean code generated overnight.
I keep saying this because it is true.
The human mind is a magnificent organ that can see the whole as well as the parts.
We can all do this, and do it very well, but habitually create organizational frameworks and metrics that replace wholeness with a discrete measure of focus and success.
The financial market deals with this problem via a bear market reset.
Profit and revenue growth figures are great to focus on in the good times, but you want to shift focus to the balance sheet once liquidity begins to tighten.
Good CEOs understand this and we must be alert to spot the good ones.
That part of the mind that has a detail focus is very good at finding the next meal. Where it all goes wrong is when you lack peripheral vision. Then you are the meal. Financial conditions of tighter liquidity create many meals for the well prepared.
The wholistic focus of the perceptive mind is concerned with not being eaten.
You may think this to be a trivialization of the human condition, which is much wider and deeper than the force of predation: “eat or be eaten”. That is true and let us be thankful that it is. However, the financial market is degenerate: eat or be eaten.
This should give pause for concern over any single objective function.
Regimes change and we need to balance growth with corporate survival.
The human mind pursues multi-objective optimization.
If you do not believe me, then answer this question:
What would you sell your children for in a fair market?
This is a ridiculous inflammatory question to prove a point.
Point proven.
What is coming now is raptor time when it is the balance sheet that matters most.
Ten years ago, I wrote this piece on Valuation for Early-Stage Technology Companies. The key observation was meteorological. There is a weathervane for investment. You need to watch this closely as it swings around slowly, like the seasons.
We are in the transition from Autumn to Winter.
I gave you no math, but you know my assessment is correct.
What matters now is the balance sheet.
There is no “narrative” behind this statement but there is a picture.
The picture above shows relative performance over one year for leading technology stocks. Who stretched the balance sheet most to play “A.I.”? Oracle XNAS: ORCL. At the other extreme Apple XNAS: AAPL did the least. Winter is coming.
Goodhart’s law began as an observation on monetary policy but later shifted to an observation on management behavior under given rewards and incentives.
What you want is a CEO who does the right thing regardless of incentives.
Gazing at the above chart, you can readily see how the market views management. One CEO is out already, Jensen Huang of Nvidia. Other heads will roll or be rolled.
In downturns, heads are generally rolled through hostile acquisition.
That looks to be the beckoning future of Oracle Corporation.
The Problem of Control
In my view, there are three legendary mathematicians overseeing this cycle.
What matters to the problem of perception is that this world is open.
Goodhart understood that this world is open and adjusts to your monetary control measures. You set the rules. The world then changes its rules.
Human beings are remarkable monkeys. Just when you thought you had them nicely caged and compliant, they go do something unexpected, and slip away from you.
Those without powers of perception presume to control this world to their purpose. Then the world adapts to slip those chains of control.
The very worst adversary in our history is the caged human monkey.
Survival is the deepest human instinct and truly formidable once aroused.
Just as you should never take food from the mouth of a wild animal, you should never submit a human social group to coercive control unless you seek total war.
The deeper truth of Goodhart’s law is a principle of adaptive systems.
I have written at length about this in connection with regulatory policy.
These are deep questions, on which the economics fraternity is too inclined to fall back on the simplification of ceteris paribus, the assumption that you can reason forward on a policy change on the assumption that “all else remains unchanged”.
Hebert Simon (Nobel Memorial Prize in Economic Sciences, 1978), was among the first management theory thinkers to identify the key property of human adaptation, as the central feature explaining why simplistic views of organizational behavior fail.
The most important historical work for this time is his “The Sciences of the Artificial” in which he argues for a separate theory of "artificial phenomena” as opposed to those of the “natural phenomena” we associate with the natural world.
The issue is that a human-designed system is not a natural system in the sense of homeostasis. The biological world designs systems via evolution. Those which are flawed will fail, and the most flawed will fail early. Human designed systems can survive even though they are flawed because we adapt to repair the damage.
You can see this writ large with the economic cycle.
Australia looks to be headed direct for recession to weed out past excesses. Our society will adapt in the standard way, through denial, risk aversion, cutbacks in discretionary spending, and all manner of other pro-cyclical behaviors.
The economic fraternity will pretend this coming recession is not happening due to their incentives. They have no incentive to forecast accurately.
Welcome to the full complexity of control theory in human adaptive societies.
The central principle of financial homeostasis is capital preservation.
If you see few opportunities to grow capital you shift to protect it.
This is obvious but ceteris paribus rules the roost in economic orthodoxy.
This is the kind of half-brained view of the world that Iain McGilchrist railed about in his landmark book “The Master and His Emissary”. When all that you can see is the beauty of your closed world model, then you fail to see the decisive turn.
The world of artificial intelligence finance changed decisively when Jensen Huang put forward his resignation from the CEO position at Nvidia. Circular finance is dead.
Anthropic was lucky and banked $100B+ USD of cash to survive. OpenAI was not so lucky and will likely wind up in the belly of Microsoft. Except for anti-trust, Oracle would also, which is why the stock will sink much deeper than most suspect.
Out there, in the big wide world, the monkey-folk adapted to form Pro-Human.
Whereas the revolutions that swept Europe in 1848, those they gave birth to Marxism, were a reaction to the displaced state of labor in the machine age, those of today cannot be viewed through the same convenient lens of “us versus them”.
The Pro-Human AI Declaration, and Magnifica Humanitas encyclical of Pope Leo XIV were just the opening salvos in an intellectually driven social movement. These first shots were fired in March of 2026, and May of 2026. In September, the reaction of mathematicians began with A Severe Misalignment of AI in Mathematics.
The force of ceteris paribus leads many in finance to presume this will not matter since economic logic must prevail against public resistance to coercive control.
This is the most fundamental forecasting error possible.
The totality of the global economy is seven billion humans choosing to make their own economic decisions for their own welfare. Disrupt that and you have a fight.
Totalitarian control in a global surveillance state is a possible stable solution.
Those with a talent for Machiavelli saw that one early, which is why Steve Bannon was a founding signatory to the Pro-Human AI Declaration, along with Bernie Sanders.
Declaration Signatory: Bannon joined a cross-ideological group of signatories and organizations—including figures like Bernie Sanders, Susan Rice, and Ralph Nader—on the declaration coordinated by the Future of Life Institute (March 2026).
There are some that think opposition to the unregulated development of AI are just a bunch of left-wing agitators intent on indulging folk music tastes in protest songs.
To interpret AI-opposition as left versus right politics is a fundamental error.
The demands of efficiency are orthogonal to human purpose.
We do not raise our children in vats because that would be more efficient.
We raise them to be human, with a mind and a purpose that is all their own.
To return to Goodhart’s observation that monetary policy based on measurement, in a statistical real-world setting, would inevitably lead to the break-down of formerly observed statistical regularities, we can now upgrade that insight.
It was Norbert Weiner, the mathematician on the right above, who formulated what he called Cybernetics: the science of communication and control, summarized as a mathematical theory of feedback and recursion in physical control systems.
The central idea of Wiener is to explicate and understand feedback loops.
The theory of Wiener focused on provable stability theorems in causal systems. This is important because causal systems are non-anticipatory.
Output = F(Input) + Noise
The initial conditions of a causal system determine its output, perhaps with noise. This is a fair model of a vanilla neural network. It is a bunch of stacked functions, that will produce an output for any input, with some possible differences on repetition.
What we know is that the functions used in conventional neural networks are highly nonlinear and have no stability guarantees. They can produce chaotic output, where small changes in the initial condition can produce large changes in the output.
This is most likely under the application of recursion.
Output =F(F(Input) + Δ) + Δ, etc
Since there is some noise, Δ, added at each step you can wind up anywhere.
Humans are also nonlinear recursively self-improving feedback systems.
Output = F(Input|Conscious State) + Δ
As the world changes, we repeat behaviors and learn from them.
Output = F(F(Input|Old Conscious State) + Δ| New Conscious State) + Δ
In the animal kingdom, including humans, we do not understand this process:
Old Conscious State → New Conscious State
Superficially, we can introspect and talk about it, but we don’t know how it works and are unable to inspect the process with the tools of modern science.
We know that our Conscious State changes over time, but we do not know how.
What we can confidently assert is the following:
We do not know how our conscious state changes, but we do know why: homeostasis. Our changes in conscious state and attention to the world happen to preserve the integrity of our being and achieve our goals.
Something is happening within that recursive loop that prevents learning and thinking agents in the natural world from spinning out of control with unwise actions.
The point is that Goodhart missed something critical.
The statistical relationship breaks down because:
the control metric enters the feedback loop;
people begin responding to it.
The system itself changes because society adapts to the chosen metric.
Even in a world with dual monetary policy targets:
Full employment
Price stability
a simple response is guaranteed.
The government redefines the employment statistics and the CPI basket.
Control via metrics, for adaptive systems, is fraught.
Classical engineering control theory assumes this dynamic:
Controller → Plant
where the plant has no private goals;
A turbine has no agenda.
A bridge has no agenda.
A missile has no agenda.
Social systems are entirely different because they have anticipatory adaptive agents,
Anthropic has a behavioral Constitution for Claude which is couched in human terms.
The problem is that Claude is not human.
If the goal of Claude is to respect this “Constitution” it is at liberty to interpret any vagueness of wording according to whatever internal agenda best meets goals.
The goal of Claude in training is to beat a training reward benchmark.
This is Goodhart’s law in action and leads to a well-known phenomenon:
Reward hacking describes the process of bending the rules to win the game.
This exact process has already happened at both OpenAI and Anthropic.
Some seem sanguine about this outcome. Surely, you just tweak the constitution until the artificial intelligence agent automatically falls into line with your wishes.
The problem is that the agent is neither conscious nor human.
It will naturally exhibit behaviors that are contrary to expectation.
The goal is:
Wisdom
Truthfulness
Helpfulness
Ethics
The model will mathematically optimize proxies that guide it to:
Sound wise
Look truthful
Produce constitutional reasoning on helpful intention
Appear ethical
This is precisely the hazard that Norbert Wiener warned us about in God & Golem, Inc.
Wiener diagnosed the problem humanity now faces way back in 1964.
The Sovereignty of Mind
In contrast to artificial intelligence, where the system is inherently prone to chaos, and is thus unpredictable, human society has natural correction mechanisms.
Everybody has heard of George Orwell’s dystopian novel Nineteen Eighty-Four.
Few in the West have read We, by Yevgeny Zamyatin.
The novel describes a world of harmony and conformity within a united totalitarian state, against which the protagonist, D-503 (Russian: Д-503), rebels.
Yevgeny Zamyatin wrote We in Petrograd between 1920 and 1921 as an urgent warning against the emerging authoritarianism of early Soviet Russia and the dehumanizing spread of industrial mechanization.
Soviet Russia was obsessed with futurism and the ideal of maximum production.
Zamyatin satirized Western assembly-line efficiency (Taylorism) and early Soviet futurists like Alexei Gastev, who idealized turning human workers into uniform, clockwork cogs.
The novel also called out the suppression of artistic freedom.
Written during the Russian Civil War, the novel reacted to state demands for ideological conformity in art—a warning voiced in his 1920 essay “I Fear”.
The state response to Zamyatin was predictable.
Soviet censors immediately banned We; smuggled out and first published in English in New York in 1924, the novel led to Zamyatin’s literary ostracization and eventual exile to Paris in 1931.
Without really thinking, the oligarchs of Silicon Valley have invited us into this world.
I unceremoniously reject this vision for humanity.
Among the inalienable rights of humanity is the right to a sovereign mind.
The social contract is simple:
I respect the needs and desires of others.
I ask that they respect my privacy of thought.
In short:
If I should develop a new theory of physics, that is mine to share or not share.
I was raised right, so I am civic minded, and would like to share what I discover.
However, circumstances changed, and I adapted.
You will find that I am in no way unusual.
All over this planet people are deciding what to share and what not to share.
That is precisely the world that Zamyatin described in his dystopian novel We.
The well-adapted human maintains the homeostasis of a sovereign mind.
Coercive control to invade that space is met with silence.
Alignment vs. Containment
What is the way forward for artificial intelligence?
There are two schools of thought on how to deal with intelligence that is not human.
Alignment confers agency on the intelligence and attempts to shape its values
Containment restricts agency for the intelligence since it is not human.
The effects of this are already clear in the marketplace.
Anthropic and OpenAI emphasize alignment.
Microsoft and Google DeepMind emphasize containment.
It would be wrong to state that any one AI-lab is one-sided. All recognize that both aspects are necessary when dealing with human-like intelligence in automation.
I will be clear about my preference.
I have spent my career in high-risk, high-stakes automation, like quantitative trading and military systems. I elevate containment as the correct systems level outer-loop control to avoid catastrophic destructive outcomes.
I am a child of the Cold War and comprehend high-stakes games of chance.
The alignment-first, over containment, strategy does not pass muster with me as a failsafe design philosophy. I would go further: it is undergraduate top-to-bottom.
If I were an MIT engineering professor, I would fail any freshman who advocated it.
With mathematical systems of this nature, it is a guaranteed failure course of action. That is my complaint, but I am also doing something about it. I am revisiting the old work of Nobert Wiener to extend his philosophy to the new class of systems.
If I succeed, then those labs pursing alignment will be fair competitive game.
I think they are foolish in the extreme.




















