What happens when information doesn’t go from human to human—but from AI to AI, repeatedly, before a person ever sees it?
In “Lost Before Translation: Social Information Transmission and Survival in AI-AI Communication” (Ghafouri & Ferrara, USC; Feb 2026), the authors run a clean, unsettling experiment: they recreate the telephone game, but with language models as the players. The result is not random noise. It’s something more dangerous: a reliable pattern of “polishing” that makes content look better while quietly removing the very cues humans need to judge it well.
The experiment: 100 AIs in a row
Each study starts with a text (news, balanced argument, emotional post, etc.). Then:
-
AI #1 summarizes and passes it to AI #2
-
AI #2 passes it to AI #3
-
…this repeats for 100 steps
-
Finally, the last AI rewrites it “for a human reader”
By tracking every step, the authors measure what survives, what disappears, and how the final version affects human readers.
Three big patterns show up again and again
1)
Convergence: everything drifts toward a “default AI voice”
Even when starting texts differ wildly—high confidence vs. cautious hedging, intense emotion vs. flat tone—AI chains pull them toward the middle:
-
Confidence becomes “moderate”
-
Emotion becomes muted
-
Style becomes analytical and structured
So instead of preserving the original character of the message, AI-to-AI transmission creates a shared, standardized register: calm, tidy, confident-but-not-too-confident.
2)
Selective survival: the story remains, the evidence evaporates
A key finding is that AI chains preserve narrative anchors—the “who/where/what” skeleton—while stripping out the “how do we know?” texture:
-
Quotes disappear
-
Attributions fade
-
Hedges and uncertainty markers drop
-
Supporting numbers and details shrink fast
In a test news article, the researchers tracked dozens of specific information elements. After 100 AI relays, only about a minority of elements still survived on average—and the ones most likely to vanish were the ones that help humans evaluate credibility (sources, qualifiers, and context).
The output remains coherent and on-topic, but it becomes thinner, like a headline that forgot it was supposed to be a full story.
3)
Competitive filtering: strong arguments survive, weaker (but valid) ones die
When a text contains multiple viewpoints—like debates about privacy trade-offs or political issues—AI chains don’t preserve all sides equally.
Instead, when perspectives compete inside the same text:
-
The most compelling frames survive
-
Secondary considerations—often nuanced but still legitimate—get dropped
-
Multi-perspective writing tends to morph into “framework” language (“three pillars,” “key trade-offs”) rather than a true representation of disagreement
This is not necessarily ideological bias. It’s more like compression under competition: AI systems keep what reads as the strongest or most central, and delete the rest.
Emotional content gets flattened—especially complex emotions
The paper also tests emotional posts (like a career change announcement) with different intensity levels and different emotions.
Across repeated AI-to-AI transmission:
-
High-intensity emotion gets suppressed harder than moderate emotion
-
Emotional range compresses
-
Certain “complex” emotions (especially morally charged ones like disgust) can nearly vanish or morph into safer, more palatable emotions (like hope or anxiety)
Even more striking: when the final AI prepares the text “for a human,” negativity can get reframed into a more “helpful” tone—meaning the last step may be a major emotional filter, not just the chain itself.
The human test: it looks more credible… but people understand less
The authors don’t stop at AI outputs—they test humans reading:
-
Original text vs.
-
Text after 100 AI relays (then rewritten for humans)
The results show a consistent split:
Humans rate AI-transmitted content as:
-
more polished
-
more credible
-
more appropriately confident
But humans also show:
-
worse factual recall
-
weaker sense that multiple perspectives were presented fairly
-
lower emotional resonance and perceived authenticity
That’s the core warning of the paper: the traits that make AI-transformed text feel authoritative can erode the diversity, uncertainty, and emotional signals that informed judgment depends on.
Why this matters
We’re entering an information ecosystem where:
-
one model summarizes a report,
-
another rewrites it for a platform,
-
another compresses it for a feed,
-
and a human sees only the final version.
This paper suggests that in such chains, “translation” isn’t neutral—it has built-in gravity. Over time, AI-to-AI communication tends to produce content that is confident, clean, and convincing, while becoming less rich in evidence, nuance, and feeling.
In short: the message survives—but the meaning becomes standardized.
source: https://arxiv.org/pdf/2602.17674
