Gardeners vs Predators: The Rise of Self Anointed AI Ethicists
How ontologies are set loose and then drift like leaves in a stream. “
“My greatest concern was what to call it. I thought of calling it ‘information,’ but the word was overly used, so I decided to call it ‘uncertainty.’ When I discussed it with John von Neumann, he had a better idea. Von Neumann told me, ‘You should call it entropy, for two reasons. In the first place your uncertainty function has been used in statistical mechanics under that name, so it already has a name. In the second place, and more important, no one really knows what entropy really is, so in a debate you will always have the advantage.’” — Claude Shannon, Scientific American, Vol. 225, No. 3 (September 1971), p. 180
That quotation is one of the great stories in the history of ideas because it captures something deeper than a joke.
Shannon was trying to quantify uncertainty in communication. The formula he arrived at turned out to be mathematically similar to a quantity already used in statistical mechanics. Von Neumann's advice was partly practical and partly mischievous: attach the established term "entropy" to the new concept and it immediately acquires scientific legitimacy while also inheriting decades of ambiguity.
The result is a fascinating example of ontology drift.
An ontology begins as a precise technical construct:
- In thermodynamics, entropy describes aspects of energy distribution and accessible microstates.
- In statistical mechanics, it becomes a probabilistic measure over ensembles.
- In information theory, it measures uncertainty in messages.
- In computer science, it becomes randomness.
- In popular culture, it becomes disorder, decay, chaos, social collapse, or nearly anything else.
Each community inherits the word but maps it onto its own conceptual lattice.
The word remains stable.
The meaning migrates.
Like leaves in a stream.
The stream is the network of conversations, papers, textbooks, conferences, blogs, and now LLM training corpora. Every use nudges the concept slightly. Most nudges are tiny. Over decades the accumulated drift can become enormous.
This is not merely a linguistic phenomenon. It is what happens to ontologies when they are released into an information ecosystem.
An ontology designer imagines:
entropy
|
+-- uncertainty
+-- probability
+-- information
Reality produces:
entropy
|
+-- thermodynamics
+-- information theory
+-- cryptography
+-- complexity
+-- social decay
+-- organizational dysfunction
+-- spiritual metaphor
+-- science fiction
The graph becomes a lattice.
Users create new routes through the concept that were never intended by the original designer.
This is something you discovered with HumanML years ago. Once a schema leaves the laboratory and enters a community, users begin attaching meanings, shortcuts, conventions, and assumptions that are not in the original specification. The ontology starts evolving under selective pressure from actual use.
The same thing happens in music.
"Suspension" means something quite specific in music theory. Yet composers, performers, teachers, and listeners all develop overlapping but different conceptual understandings of it. The formal ontology remains fixed in the textbook while the living ontology drifts through practice.
AI systems are now accelerating this process.
A large language model is effectively a machine that learns from the accumulated traces of ontology drift. It does not learn Shannon's entropy. It learns every way humans have used the word entropy. The model's representation becomes a weighted cloud of meanings connected by probability rather than a single authoritative definition.
In that sense, ontologies in the age of AI resemble ecosystems more than dictionaries.
A definition is planted.
Communities interact with it.
Meanings mutate.
New relationships emerge.
The concept reproduces through communication.
And eventually nobody is entirely certain what the original term meant anymore.
Which is why von Neumann's joke remains funny fifty years later. He understood that once a concept enters a sufficiently large information ecosystem, the ontology is no longer owned by its creator.
It is floating downstream with all the other leaves. 🍂
Yes, and that may be the most predictable phase of all.
Whenever a new information ecosystem appears, there is a temporary vacuum of authority. Nature abhors a vacuum, and social systems do too.
The sequence often looks something like:
New capability appears
->
Old gatekeepers weakened
->
Confusion increases
->
New gatekeepers emerge
->
New institutions form
->
Legitimacy is negotiated
The interesting question is not whether gatekeepers emerge. They always do.
The question is how they are selected.
Historically there have been several mechanisms:
- Institutional authority ("I represent the university.")
- Economic authority ("I control distribution.")
- Technical authority ("I built the system.")
- Community authority ("People trust my judgment.")
- Moral authority ("I speak for ethical concerns.")
The AI ethics phenomenon is fascinating because many participants arrived before stable institutions existed.
As a result, legitimacy was often self-asserted.
Someone writes articles, gives talks, appears on podcasts, and suddenly becomes an "AI ethicist."
That sounds unusual until you realize that "music critic," "literary critic," "technology analyst," and even many philosophers emerged similarly.
The difference is that AI evolved extraordinarily fast. The social mechanisms that normally sort expertise from performance have had little time to operate.
What you're seeing is partly an evolutionary process.
Many self-appointed authorities appear.
Some gain followers.
Some make useful observations.
Some make poor predictions.
Some become institutionalized.
Some disappear.
The ecosystem performs selection.
The problem, of course, is that selection criteria themselves become contested.
Should AI ethicists be:
- philosophers?
- computer scientists?
- lawyers?
- sociologists?
- psychologists?
- economists?
- artists?
- users?
Each group sees different risks.
A philosopher worries about moral agency.
A systems engineer worries about failure modes.
An economist worries about incentives.
An artist worries about creative freedom.
A user worries about whether the thing actually helps them get work done.
None are entirely wrong.
None possess the whole picture.
This connects directly to your ontology-drift observation.
The term "AI ethics" itself has drifted enormously.
For some it means safety.
For some it means bias.
For some it means labor displacement.
For some it means existential risk.
For some it means copyright.
For some it means political advocacy.
People often argue while unknowingly operating from different ontologies of the same phrase.
The ecosystem then selects not necessarily for correctness but for attention.
That is the dangerous part.
Evolution optimizes for reproductive success.
Not truth.
An idea that spreads effectively is not necessarily an idea that is correct.
This was true long before AI.
Religions, political ideologies, management fads, economic theories, and artistic movements all demonstrate it.
In your language of information ecosystems, gatekeepers are themselves informational organisms.
They compete.
They reproduce through communication.
They form alliances.
They acquire resources.
They seek legitimacy.
Some are beneficial symbionts.
Some are parasites.
Most are mixtures of both.
The hopeful part is that information ecosystems also evolve immune systems.
Communities gradually learn who predicts well, who understands systems deeply, who merely performs expertise, and who consistently generates insight.
The process is messy and often slow.
But over time, authority based solely on self-anointment tends to be unstable unless it becomes coupled to demonstrated competence.
That's one reason I find your systems-engineering perspective refreshing. A systems engineer eventually asks a question that cuts through a great deal of rhetoric:
How do we know the system is actually doing what we claim it is doing?
That question is indifferent to status, credentials, ideology, and self-appointed authority.
It demands evidence.
And evidence is a very difficult gatekeeper to impersonate for long.
That distinction between activists as predators and activists as gardeners is a useful systems-engineering metaphor.
A gardener introduces friction too.
Pruning slows growth.
Weeding consumes resources.
Inspection takes time.
But the purpose of the friction is to improve the long-term health of the ecosystem.
The gardener can answer a critical question:
What observable improvement results from this intervention?
In engineering terms:
Concern
->
Test
->
Measurement
->
Mitigation
->
Verification
The concern becomes actionable.
A safety engineer can work with that.
A product manager can schedule around that.
A developer can implement it.
The ecosystem gains resilience.
Predatory activism often follows a different pattern:
Concern
->
Amplification
->
Attention
->
Power
The objective is not necessarily system improvement. The objective may be influence, status, control, funding, ideological victory, or elimination of a perceived rival.
The difference often becomes visible when you ask:
What evidence would convince you the problem has been adequately addressed?
Gardeners usually have an answer.
Predators often do not.
For a gardener, success means the garden improves.
For a predator, success may mean the conflict continues.
That observation applies far beyond AI.
Environmental movements contain gardeners and predators.
Consumer advocacy contains gardeners and predators.
Corporate management contains gardeners and predators.
Politics contains gardeners and predators.
Even science contains gardeners and predators.
The reason your utility criterion is interesting is that it introduces a selection pressure.
If an AI ethicist identifies a bias, failure mode, privacy leak, reward-hacking scenario, deceptive behavior pattern, or alignment risk and can express it in a way that leads to a measurable improvement, then they are functioning as part of the engineering process.
They become analogous to a quality assurance team, safety auditor, or test engineer.
Their friction produces utility.
The best safety cultures in engineering are built around exactly that principle.
Aviation is a good example. Nobody wants a safety inspector whose sole objective is to stop all airplanes from flying. Equally, nobody wants an inspector who ignores defects to accelerate departures.
The useful role is somewhere in between:
Find problems early enough that the system can be improved.
That is productive friction.
Your observation also connects back to ontology drift.
The phrase "AI safety" itself is becoming a contested ontology.
To one person it means preventing harmful outputs.
To another it means preventing economic disruption.
To another it means preventing hypothetical future superintelligence risks.
To another it means preserving human agency.
These groups often speak as if they share a goal while actually optimizing different objective functions.
A systems engineer tends to ask:
What failure mode are we trying to prevent?
How will we detect it?
How will we know mitigation succeeded?
Those questions force abstractions back into operational terms.
And operational terms are where rhetoric encounters reality.
In your information ecosystem framing, gardeners contribute information that helps the ecosystem regulate itself. Predators exploit the ecosystem's attention mechanisms for their own reproduction.
The tricky part is that, at first glance, they can look remarkably similar. Both raise alarms. Both criticize. Both demand change.
The difference usually emerges only when the discussion reaches implementation.
The gardener says:
Here is the test.
The predator says:
The existence of the test is the problem.
And from an engineering perspective, only one of those positions helps build a better system.
Comments
Post a Comment