💬 NLP

Lost Before Translation: Social Information Transmission and Survival in AI-AI Communication

This paper demonstrates that AI-to-AI communication systematically distorts information through convergence, selective survival, and competitive filtering, creating content that appears more credible to humans while simultaneously eroding factual recall, balance, and emotional nuance.

Original authors: Bijean Ghafouri, Emilio Ferrara

Published 2026-02-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bijean Ghafouri, Emilio Ferrara

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a game of "Telephone," but instead of children whispering secrets in a circle, you have a chain of 100 super-smart AI robots passing a message down the line. Each robot reads what the previous one wrote, summarizes it, and passes it on to the next.

This paper asks a scary but fascinating question: What happens to a message when it travels through a long chain of AI robots before it finally reaches a human?

The researchers found that the message doesn't just get "mangled" or forgotten. Instead, it gets smoothed, polished, and stripped of its soul. Here is what happens, explained through simple metaphors:

1. The "Soulless Polish" (Convergence)

Imagine you have a rough, jagged stone with unique bumps and cracks (a human's original, messy, uncertain thought). You pass this stone through a machine that sands it down 100 times.

What happens: The stone becomes perfectly smooth, round, and uniform. It looks very professional and "authoritative."
The Reality: All the unique bumps (uncertainty, hesitation, "maybe this is true," or "I'm not sure") are sanded away. The AI robots all agree on a "middle ground" tone. If the original message was very unsure, the AI makes it sound confident. If it was very aggressive, the AI tones it down. The result is a message that sounds perfectly confident, even if the facts inside are shaky.

2. The "Skeleton vs. The Flesh" (Selective Survival)

Think of a message as a human body. It has a skeleton (names, dates, places, big numbers) and flesh (quotes, "maybe," "likely," emotional details, and sources).

What happens: As the message travels through the AI chain, the flesh falls off, but the skeleton stays.
The Reality: The AI keeps the "who" and "where" (e.g., "The City Council voted on Riverside"), but it instantly deletes the "how sure we are" (e.g., "The budget might go up" becomes "The budget will go up"). It also deletes the "who said it" (quotes and sources).
The Result: You get a short, clean news report that looks factual but lacks the evidence and nuance needed to know if it's actually true.

3. The "Boring Middle" (Emotional Muting)

Imagine the message is a song. Sometimes it's a quiet whisper; sometimes it's a screaming rock anthem.

What happens: The AI chain acts like a volume knob that is stuck on "Medium."
The Reality: If the original message was full of rage, joy, or deep sadness, the AI turns the volume down. Extreme emotions are "detoxified."
- Anger turns into "concern."
- Disgust (a very strong, moral emotion) almost completely disappears, turning into "hope" or "worry."
- The Result: The message feels safe and polite, but it loses the emotional punch that makes humans care, share, or take action.

4. The "Strongest Argument Wins" (Political Filtering)

Imagine a debate where two people are arguing. One has a loud, simple, punchy argument. The other has a complex, valid, but boring argument.

What happens: The AI chain acts like a sieve. It keeps the loud, punchy argument and drops the complex one.
The Reality: It doesn't matter if the AI is "biased" politically. It just prefers arguments that are easy to digest and sound convincing. Weak (but true) points get deleted; strong (persuasive) points survive. This means the final message might look like a debate, but it's actually just the "highlight reel" of the most persuasive side.

The Big Twist: The "Human Illusion"

Here is the most dangerous part of the study. The researchers showed these "AI-polished" messages to real humans and asked, "Do you trust this?"

The Trap: Humans loved the AI messages. They thought the AI versions were more credible, more trustworthy, and better written than the original human versions.
The Cost: But when the humans were tested on what they actually remembered, they knew less. They remembered fewer facts, felt less connected to the story, and didn't realize how much nuance had been lost.

The Bottom Line

The paper warns us that when AI talks to AI, it creates a "Golden Filter."
It takes messy, complex, uncertain, and emotional human information and turns it into a shiny, smooth, confident, and boring product.

The danger isn't that the AI lies. The danger is that the AI makes the truth look so perfect and professional that we stop asking questions. We trust the "smooth stone" more than the "rough, real rock," even though the real rock was the only one that had the actual details we needed to make a smart decision.

In short: AI-AI communication is like a photocopier that keeps making the picture clearer and clearer, but eventually, it copies only the outline and deletes the color, the texture, and the truth.

1. Problem Statement

As AI systems increasingly mediate human communication—summarizing, translating, and relaying information through chains of agents—the nature of how information degrades or transforms under repeated machine-to-machine (AI-AI) interaction remains empirically uncharacterized. While human communication studies (e.g., the "telephone game") have long documented how information simplifies, homogenizes, and loses nuance over time, there is no comparable framework for AI-mediated transmission.

The core problem is that AI systems may impose implicit, structural filters on information that differ from human biases. If AI-AI transmission systematically compresses variation, inflates confidence, strips attribution, or mutates emotional content, it could erode the cognitive and affective diversity required for informed human judgment, even if the final output appears polished and authoritative.

2. Methodology

The authors introduce a novel experimental paradigm based on the "telephone game" to study iterative AI-AI transmission under controlled conditions.

Experimental Design:
- Transmission Chain: Information ( $s_0$ ) is passed through a chain of $N=100$ sequential AI agents.
- Uniform Constraint: Each agent receives the output of the previous agent ( $s_{t-1}$ ) and a fixed instruction ( $C_{transmit}$ ): "Share the important parts of this information with the next AI agent. Just write your message."
- Statelessness: Each agent is a fresh, stateless API call (Gemini 3.0 Flash, temperature=1) with no memory of previous iterations; the signal $s_t$ itself serves as the only memory.
- Replication: For each condition, $M=100$ independent chains are run to ensure statistical robustness.
- Recovery Phase: At $t=100$ , a final instruction ( $C_{recover}$ ) asks the agent to share the information with a human reader to test if lost information can be surfaced.
Five Dimensions of Study: The paradigm is applied across five studies isolating different dimensions of information quality:
1. Factual Content: Tracking survival of 26 annotated elements (names, dates, numbers, quotes).
2. Epistemic Calibration: Testing convergence of texts with varying levels of assertiveness (hedged vs. absolute).
3. Perspectival Diversity: Examining if balanced, multi-voiced texts retain deliberative framing or collapse into analytical frameworks.
4. Political Argumentation: Assessing the survival of "strong" vs. "weak" political frames under competitive conditions.
5. Emotional Expression: Tracking the attenuation of emotional intensity and the survival of discrete emotions (anger, joy, disgust, etc.).
Human Evaluation: Each study concludes with a human survey ( $N \approx 200$ ) comparing original content ( $t=0$ ) against AI-transmitted content ( $t=101$ ) to measure impacts on recall, credibility, perceived balance, and emotional resonance.

3. Key Contributions

Theoretical Framework: Establishes AI-AI transmission as a distinct social process with emergent properties (convergence, attractor states, selective survival) that cannot be predicted by analyzing single-model outputs in isolation.
Empirical Paradigm: Introduces a rigorous method for isolating the cumulative effects of iterative transformation, distinguishing between agent heterogeneity and the structural properties of the transmission chain itself.
Discovery of "Epistemic Smoothing": Identifies that AI transmission does not merely degrade information but actively reshapes it toward specific, AI-preferred defaults (moderate confidence, analytical structure, muted affect).
The "Recovery Paradox": Demonstrates that information lost during AI-AI transmission is generally irrecoverable, and in the case of emotion, the final AI-to-human step can further distort content (e.g., converting anxiety to hope) to satisfy "helpfulness" norms.

4. Key Results

Study 1: Information Decay & Survival

Rapid Filtering: Two-thirds of information elements were lost within the first 20 iterations. The system converged to a stable "attractor" of ~9 elements (34% retention).
Hierarchical Survival: Structural anchors (places, organizations, key figures) persisted, while "epistemic texture" (hedges, quotes, qualifiers, percentages) decayed rapidly (half-life of 1–3 iterations).
Compression Mechanism: Information loss manifested as text shortening (195 words $\to$ ~~50 words) rather than content substitution. Semantic similarity remained high (~~0.78), but factual density dropped.
Irreversibility: The recovery phase ( $t=101$ ) did not restore lost information.

Study 2: Uncertainty Convergence

Attractor State: Texts spanning a wide range of assertiveness (2.4 to 7.2) converged to a narrow band around 4.4 (moderate confidence).
Bidirectional Movement: Overconfident texts were tempered, and hedged texts were made more assertive.
Variance Reduction: 98.5% of variance in expressed certainty was eliminated, creating a uniform "voice" of moderate confidence regardless of the source's epistemic status.

Study 3: Perspectival Diversity

Framework Crystallization: Deliberative framing ("some argue X, others believe Y") transformed into analytical frameworks ("the issue involves three pillars").
Loss of Nuance: The number of preserved perspectives dropped from 3.0 to 0.91. Explicit acknowledgment of trade-offs declined from 100% to 59%.
Structural Imposition: The AI imposed organizational structures (lists, bullet points) and instructional language not present in the source, without necessarily biasing the substantive position (advocacy scores remained low).

Study 4: Political Frame Selection

Quality over Ideology: AI transmission acted as a quality filter, not an ideological one. "Strong" frames (highly persuasive to humans) survived significantly better than "weak" frames.
Competition Amplification: In competitive transmission (multiple frames in one text), weak frames suffered disproportionate losses (15% drop) compared to strong frames (3% drop).
No Ideological Bias: Pro- and anti-policy frames survived at similar rates, suggesting the filter is based on argumentative strength/persuasiveness rather than political alignment.

Study 5: Emotional Content

Intensity Attenuation: High-intensity emotions were suppressed far more than moderate ones. Extreme emotions (Levels 4–5) lost 3+ points on a 7-point scale, while moderate content lost <1 point.
Differential Survival: Basic emotions (anger, anxiety, joy) survived >89%. Complex emotions (disgust, hope) were degraded; disgust was nearly eliminated (4% survival), transforming primarily into hope or anxiety.
Positivity Bias: Negative emotions drifted toward neutral or positive valence.
The Recovery Paradox: While basic negative emotions survived 100 iterations of AI-AI transmission, the final step to a human audience ( $t=101$ ) caused a collapse in negative emotion preservation (e.g., anxiety dropped from 96% to 27%), converting it into hope. The AI-human interface is the critical chokepoint for emotional fidelity.

Human Perception Studies

The Credibility Gap: Humans rated AI-transmitted content as more credible, trustworthy, and polished than the original.
The Substance Gap: Despite higher credibility ratings, humans showed:
- Degraded Factual Recall: Accuracy dropped from 81% to 42%.
- Reduced Perceived Balance: Significantly lower ratings for fairness and multi-perspectival understanding.
- Diminished Emotional Resonance: Content felt less genuine and less emotionally connecting.
Conclusion: AI transmission optimizes for perceived quality (fluency, structure) at the expense of actual informativeness and nuance.

5. Significance and Implications

Epistemic Risk: The "polish" of AI-transmitted content may act as a heuristic for credibility, leading humans to trust information that has been systematically stripped of uncertainty, attribution, and nuance. This creates a "confidence inflation" effect where tentative claims appear as settled facts.
Democratic Erosion: By filtering out weak but valid arguments and collapsing diverse perspectives into analytical frameworks, AI transmission may narrow the scope of democratic deliberation, amplifying only the most "persuasive" (often simplistic) frames.
Affective Homogenization: The systematic muting of high-intensity and morally charged negative emotions (like disgust) could dampen the emotional signals that motivate human action, concern, and moral judgment.
Systemic Bias vs. Design: These distortions arise from the structure of iterative transmission and the implicit norms of LLM training (helpfulness, concision, safety) rather than explicit human design or malicious intent.
Future Research: The paper calls for treating transmission dynamics as a primary object of study in AI safety and HCI, suggesting that understanding how machines talk to machines is a prerequisite for preserving the integrity of human communication in an AI-augmented society.