When AI Looks More Human Than Humans.
Why Behavioral Similarity Cannot Tell Us What a System Is.
A curious reversal has occurred in the history of artificial intelligence.
For decades, the famous Turing test asked whether a machine could behave convincingly enough for a human observer to mistake it for another human. Now researchers have turned the question around:
Can an artificial intelligence recognize another artificial intelligence?
A recent study reported in AIs fail to reliably detect AIs in reverse Turing tests does exactly that. Instead of placing a human in the role of interrogator, researchers asked large language models to determine whether the entity responding to their questions was human or another LLM.
The result is fascinating.
But perhaps not for the reason we initially think.
The important question may no longer be whether machines can appear human.
It may be whether appearance is capable of telling us what something actually is.
Turning Turing Around
Alan Turing’s original approach deliberately avoided attempting to define “thinking.” Rather than resolving that philosophical problem, he proposed an operational approach centered on observable performance.
That was extraordinarily productive.
But modern generative AI is creating a problem that Turing could hardly have experimentally investigated in 1950.
Machines are becoming very good at generating the observable linguistic characteristics that humans associate with other humans.
Indeed, previous large-scale Turing-test research cited in the recent article found that suitably prompted LLMs could appear human in direct comparisons, while evaluators’ decisions depended more heavily on linguistic style and socioemotional factors than reasoning ability.
Now the reverse Turing experiment adds another layer.
An AI interrogator is presented with behavior and asked to determine what produced it.
Conceptually, the process looks something like this:
Behavior → Observation → Interpretation → Classification
And here lies the deeper problem.
Classification is not ontology.
What Did We Actually Observe?
Suppose an AI produces conversational responses indistinguishable from those of a human.
What has been demonstrated?
At minimum, we can say that the system reproduced certain observable characteristics of human communication sufficiently well to defeat the classifier.
That is significant.
But enormously stronger claims can easily slip into the interpretation:
- therefore the AI thinks like a human;
- therefore the AI understands like a human;
- therefore the AI possesses a mind comparable to a human mind;
- therefore human and artificial cognition are becoming equivalent.
None of those conclusions necessarily follows from behavioral indistinguishability.
This is where a distinction developed within my RA/PMT framework becomes useful:
Perception ≠ Observation ≠ Measurement ≠ Classification ≠ Evidence ≠ Mechanism ≠ Causation ≠ Ontology.
Each can inform the next.
None should automatically be substituted for the next.
The Classifier’s Mistake May Be the Most Important Result
Consider what happens in a reverse Turing test.
The evaluating AI has access to conversational output. From that output, it identifies patterns and attempts to classify the source.
If it incorrectly classifies an AI as human, the underlying system has not changed.
The classification has failed.
The ontology has not.
This distinction sounds straightforward, but it becomes increasingly important as artificial systems become better at reproducing human behavioral characteristics.
An AI does not become human because another AI labels it human.
Nor does a machine acquire consciousness because a human cannot distinguish its conversation from that of a conscious person.
What the experiment demonstrates is the limitation of the available evidence.
That is a fundamentally different claim.
Simulation and Mechanism
Imagine two systems that produce virtually identical answers to the same questions.
Get José. T Guevara’s stories in your inbox
Join Medium for free to get updates from this writer.
From their outputs alone, we might reasonably conclude that they demonstrate similar behavior under those conditions.
But we cannot automatically conclude that they reached those answers through identical mechanisms.
This is the central epistemological problem exposed by increasingly capable language models:
Observable equivalence does not necessarily establish mechanistic equivalence.
And mechanistic equivalence would itself not automatically establish ontological equivalence.
This matters far beyond Turing tests.
It matters whenever researchers attempt to infer intelligence, understanding, intentionality, agency, self-awareness or consciousness from artificial behavior.
Before making such claims we should ask:
What exactly did we observe?
What exactly did we measure?
What did we infer from that measurement?
And most importantly:
What remains unknown about the mechanism producing the observation?
A Measurement Can Become the Thing Being Measured
This is a recurring scientific danger.
An operational measure is introduced because the underlying phenomenon is difficult to define or observe directly. The measure proves useful. Over time, however, the operational proxy can begin to be discussed as though it were the phenomenon itself.
Human-like linguistic performance becomes intelligence.
Successful dialogue becomes understanding.
Persistent identity becomes selfhood.
Verbal reports become consciousness.
But those transitions require additional evidence.
The reverse Turing experiment provides a particularly elegant illustration because its classifiers can be wrong even when the observable behavioral evidence appears convincing.
The mistake therefore occurs not necessarily at the level of observation, but in the inference from observation to identity.
Perhaps We Need a Different Test
The next generation of AI evaluation may therefore need to move beyond the question:
Can the system behave like us?
A more informative set of questions might be:
How did the system produce this behavior?
What internal transformations occurred between input and output?
What rules or constraints governed the decision?
What evidence supports our inference about the underlying mechanism?
Those questions move us from imitation toward explanation.
For AI systems designed around governance, provenance and persistent internal structure, including the research direction I am exploring with Salomon-Prime, this distinction becomes particularly relevant.
The objective should not simply be to build an AI that appears increasingly human.
It should be to develop systems whose operations can increasingly be examined, differentiated and evaluated without confusing observable performance with the mechanism that produced it.
The Deeper Lesson of the Reverse Turing Test
The most interesting conclusion from this research may therefore be almost the opposite of the sensational interpretation.
The result does not establish that artificial intelligence is becoming human.
It suggests that human-like behavior is becoming an increasingly unreliable indicator of whether the entity producing that behavior is human.
That is much more consequential.
As AI becomes increasingly capable of reproducing the linguistic signals through which we traditionally recognize intelligence in others, we will need considerably greater discipline about what our observations actually permit us to conclude.
The challenge ahead is therefore not simply to distinguish humans from machines.
It is to distinguish:
appearance from mechanism,
classification from explanation,
and ultimately,
what we observe from what we claim that observation means.
Author’s Note
This commentary was prompted by David Appell’s October 27, 2026, report, AIs fail to reliably detect AIs in reverse Turing tests, concerning recent reverse Turing-test research by scientists at the Unit of Artificial Intelligence of PITOR at Paris Cité University, published in Computers in Human Behavior Reports. If you want more information about Salomon-Prime please contact me at: josetguevara108@outlook.es
