Two people receive the same policy proposal.
The first is told it came from a progressive government.
The second is told it came from a conservative government.
The cost is identical.
The implementation plan is identical.
Even the typo in paragraph seven remains courageously bipartisan.
Yet one reader discovers necessary public investment while the other discovers government waste. Elsewhere, the judgments reverse.
Nothing inside the proposal changed.
One word changed what the document appeared to contain.
Labels are instructions for attention
We often treat a label as a little storage box.
“Luxury” summarizes a product.
“Scientific” summarizes a claim.
“Religious,” “woke,” “traditional,” “foreign,” “natural,” “AI-generated” and “conspiracy theory” tell us what kind of thing has arrived.
This is useful. A mind that evaluated everything from first principles would not make it through the cereal aisle.
But a label does more than name the box.
It tells us how to open it:
- which features matter;
- what danger to expect;
- which memories to retrieve;
- whose side may be involved;
- and how much skepticism to bring.
The label is not merely information.
It is a viewing instruction.
One word can import an entire library
Consider natural.
The word may activate ideas about health, simplicity, safety, purity and a wholesome person standing barefoot near a leaf.
None follows automatically. Arsenic has maintained an entirely natural production process.
Or consider Harvard research.
The label may increase credibility before we know the sample size, method, authors or whether Harvard was merely the location of the car park in which the survey occurred.
This is prior knowledge doing its job. Institutions, categories and political parties really do develop patterns.
The trouble is that the prior can become so efficient that the object is no longer required.
Party can become stronger than policy
In a series of experiments, psychologist Geoffrey Cohen gave participants policies that were generous or stringent and varied which US political party supposedly supported them.
Participants’ attitudes followed the party cue strongly—even when the policy’s content conflicted with their usual ideology. They also tended to believe the policy’s substance, rather than the party label, had driven their judgment. (Cohen, 2003)
This does not mean voters never read policy.
It means a group label can change what the policy appears to mean.
The same expenditure can become compassion or dependency.
The same restriction can become safety or control.
The label activates a model, and the details are interpreted inside it.
Context can change the experience, not only the report
There is an even stranger possibility.
Sometimes a label may not merely change what we say about an object. It may change the experience itself.
In one experiment, participants tasted wines while being shown different prices. The wines presented as more expensive were rated as more pleasant, even where the underlying wine was the same. The price manipulation was also associated with activity in a brain region involved in experienced pleasantness. (Plassmann et al., 2008)
“They were pretending to like it” is not the only explanation. Expectation can enter perception.
This makes the line between packaging and substance less tidy than I would prefer.
If the label changes the experience, is the experience false?
The chemical liquid did not improve.
The human event apparently did.
The label gets there first
A label has one peculiar advantage over the evidence it describes: it often arrives first.
By the time I begin reading AI-generated advice, I am no longer quite the same receiver who would have read the paragraph without that information. A prior has been activated. Hallucinations, generic language and suspiciously balanced sentences have become worth hunting for.
Label the same paragraph written by a professor and different features may become diagnostic. Technical language can look like sophistication rather than camouflage.
The sentence has not changed.
The state in which I encounter it has.
This is more than a hypothetical now.
In a counterfactual experiment presented at ACL 2026, people gave higher trust ratings to information labelled as human-authored than to the same information labelled as AI-generated. Eye-tracking indicated that participants relied heavily on the source label as a heuristic cue. This is unusually close to the experiment I wanted: the label did not merely accompany the judgment; it attracted attention during the path toward it. (Sun et al., 2026)
That does not prove an AI label makes every reader hunt for the same flaws. Gaze at the label cannot tell us everything the reader subsequently noticed in the prose, and the study examined a particular trust task. But the effect is no longer only a question.
It also appears outside general trust judgments. Across nine studies involving 6,282 participants, identical AI-generated empathic responses were rated as more empathic and supportive when attributed to a human rather than AI. (Rubin et al., 2025)
A 2026 preregistered working paper adds a useful complication. When 654 participants read the same short story under an AI or human-authorship label, the AI label lowered subjective ratings—especially authenticity, atmosphere and literary merit—but barely changed reading time or willingness to pay. (Abel & Johnson, 2026)
Labels may change what we report thinking without changing what we are willing to spend.
Or perhaps they change some parts of the information-processing path more than others.
Either way, the stronger claim is no longer merely that labels influence opinions.
They can alter the route by which an opinion is reached.
Labels help
Removing all labels would not create pure reasoning.
It would create confusion.
A medical recommendation without a source is missing relevant information. A warning label can prevent injury. “Peer reviewed” does not guarantee truth, but it tells us something about the process. “Satire” can stop an aunt from reorganizing the family emergency plan.
Source information should affect trust when the source has knowledge, incentives or a track record relevant to the claim.
The problem is not using context.
The problem is allowing context to settle questions the object could answer for itself.
A small test for label capture
When a word feels decisive, I want to try three questions:
- What did the label make me look for?
- Which part of my judgment would survive if the source were reversed?
- What evidence inside the object could change my mind?
Imagine learning tomorrow that the article you loved was AI-generated, the policy you supported came from the opposing party or the research you dismissed came from an institution you respect.
What, precisely, should change?
The correct answer is not necessarily nothing. Some source information should update us.
The interesting question is how much.
This will not remove priors.
Nor should it.
It may reveal when the prior has stopped helping us inspect and started doing the inspection on our behalf.
There is a practical experiment here too: evaluate the proposal, artwork or argument before revealing its source, then evaluate it again afterwards.
The difference between those judgments is not necessarily bias. Source credibility can be rational.
But the size and direction of the change tells us what the label is doing.
What I still cannot figure out
Information never arrives without context.
Even an unlabeled document appears in a place, with a font, through a person who chose to send it. If we remove the official category, the mind may build one from accent, design, vocabulary or suspicion.
Perhaps the goal is not to become label-free. I doubt such a person exists.
Perhaps the better skill is noticing when the label has quietly become the evidence.
And there is an uglier question underneath that one.
If information always arrives through some prior model, is it possible to judge the thing “on its own merits”—or is that simply another label for the interpretation we prefer?
Until the next strange question,
Osagie