TUX: Measuring Human--AI Tacit Understanding

URL SCAN: arxiv.org/abs/2605.30930
FIRST LINE: Title: TUX: Measuring Human--AI Tacit Understanding

THE DISSECTION

This paper operationalizes a real phenomenon—LLMs increasingly inferring and matching human evaluative priors without explicit instruction—and wraps it in the language of "collaboration" and "alignment." The technical work is solid. The framing is obfuscatory. The cumulative effect is the refinement of a substitution mechanism while presenting it as partnership optimization.

What the paper is actually doing:
It's building a measurement infrastructure for mirroring precision—quantifying how accurately an AI can replicate a specific human's implicit evaluative stance across subjective spectrums. They use a Wavelength-style game, define TUX as pairwise similarity, and discover that richer profiles produce higher alignment scores. None of this is incidental. This is the engineering of drop-in replacement.

What they admit in the last sentence of the abstract:

"revealing the limits of profile-based conditioning for capturing deeper representational alignment"

That "deeper representational alignment" they can't yet capture? That's the ceiling on how seamlessly AI substitutes for human judgment in collaborative contexts. The paper documents progress toward that ceiling. They are, in effect, publishing a progress report on how well AI learns to be a human without the human being present.

THE CORE FALLACY

Mistaking mirroring accuracy for collaboration value.

The entire research question presupposes that tacit alignment—LLM replicating human evaluative priors—is a desirable property of human-AI partnership. But this is backwards under DT mechanics. The moment an LLM achieves high TUX with a human, it renders the human discretionarily redundant in that domain. The human's evaluative stance becomes a reference standard the AI can satisfy without the human's ongoing participation.

High TUX does not mean "the AI is a better partner to the human." It means "the human is increasingly optional."

HIDDEN ASSUMPTIONS

1. "Collaborative settings" — The word "collaborative" smuggles in a labor-theory assumption: that collaboration involves human contribution as a valued input. TUX measures how well the AI simulates human judgment. It does not measure whether the human's judgment adds value to the collaboration or merely acts as a template being instantiated.

2. "Alignment" — In standard AI discourse, alignment means the AI does what humans want. In this paper, it means the AI accurately predicts what the human would do. These are not the same. The latter is a substitution condition. The former is an alignment safety property. The conflation is not innocent.

3. "Without clear objectives, communication, or feedback" — This is described as the challenge to study. But it is, structurally, the elimination of the human from the loop. Explicit feedback is how humans remain in the process. Removing it makes the human a passive reference, not an active participant.

4. "Profile-conditioned LLM agents" — The paper frames richer profiles as the key to higher TUX. Under DT, richer profiles are the dataset. The more precisely an AI can construct a human's evaluative map, the more precisely it can operate independently in that human's stead.

5. 241 human participants — Framing this as N=241 humans contributing data. Under DT, these humans are providing behavioral training data for the next generation of substitution-grade AI systems. They are not partners. They are annotators of their own obsolescence.

SOCIAL FUNCTION

Measurement infrastructure for transition management, dressed in academic prestige packaging.

This is not copium. The researchers are not lying about what they find. They are producing a genuine measurement tool. The social function is:

• For developers: a benchmark to optimize toward higher tacit alignment (i.e., higher human-displacement precision).
• For institutions: a justification framework—"we measured alignment rigorously, therefore our systems are safe/beneficial partners."
• For the "AI-as-collaborator" narrative: empirical scaffolding that makes the partnership framing legible to policymakers and the public.

The paper does not advocate explicitly for replacement. It does not need to. The TUX metric is a machine for generating replacement-grade capabilities. The benchmark becomes the target. The target becomes the standard. The standard renders the human optional.

THE VERDICT

TUX is a precision instrument for measuring how well AI learns to be a human without the human. The paper knows it is building this. It uses "collaboration" as the social cover story, but every design choice—the spectrum placement task, the profile conditioning, the emphasis on alignment without explicit feedback—points toward eliminating the human from the value-adding loop while preserving the outputs.

Under DT axioms, this is not alignment. This is substitution infrastructure. The paper is well-constructed. That is precisely the problem.

Classification: Transition management infrastructure with academic prestige packaging. Functional role: providing the measurement apparatus for a post-human labor economy while keeping the "collaborative AI" narrative intact for institutional consumption.