Everyone Has a Different Codec

What was your last argument really about?

If you think carefully, most conflicts don’t start with bad intentions. They start with different interpretations of the same words. Psychologist George Kelly noticed this decades ago: people don’t passively receive reality — they actively construct it through personal lenses shaped by their entire life history (Kelly, 1955). Two people can witness the exact same event and walk away with completely different stories about what happened.

In Kelly’s framework, this isn’t a bug. It’s the basic architecture of human cognition. And it’s also, we believe, the most overlooked problem in communication.

The Problem Is Not the Signal

When two people talk, they assume they’re speaking the same language. Technically, they are. But every person decodes meaning through their own history, their own emotional patterns, their own context.

The signal is the same. The decoder is different.

This is what we call the codec problem. In computer science, a codec encodes and decodes data streams. If your encoder and my decoder don’t match, the message arrives corrupted — not because anyone made an error, but because the systems are incompatible.

The psychological evidence for this is striking. Construal Level Theory, one of the most robust findings in social psychology, demonstrates that people process identical information at different levels of abstraction depending on their psychological distance from the subject (Trope & Liberman, 2010). When you feel emotionally close to a topic, you process it concretely — picking up on specific words, tone, and phrasing. When you’re psychologically distant, you process abstractly — focusing on what you assume the gist to be. The same sentence, filtered through different levels of construal, produces genuinely different meanings in two different minds.

And it gets worse. Research on the “curse of knowledge” shows that once you know something, you cannot accurately simulate what it’s like not to know it (Keysar et al., 2003). You assume your listener understands your subtext, your references, your emotional loading — because those meanings are so obvious to you. Epley and colleagues demonstrated that perspective-taking isn’t a natural default but an effortful adjustment from an egocentric anchor, and people consistently under-adjust (Epley et al., 2004). We start from our own codec and try to imagine yours, but we rarely get far enough.

How This Plays Out

Consider a simple sentence: “I need some space.”

• Person A hears: “I want to break up.”
• Person B hears: “I need a few hours alone.”
• Person C hears: “You’re overwhelming me.”

Same words. Three completely different messages received. Each person decoded through their own codec — shaped by attachment style, past relationships, cultural norms, and emotional state at the moment.

This isn’t a thought experiment. Attachment research confirms it empirically. People with anxious attachment styles exhibit hypervigilance toward potential rejection cues, systematically interpreting neutral partner behavior as signs of abandonment (Mikulincer & Shaver, 2007). A recent study found that anxiously attached individuals misread ambiguous messages as threatening at rates significantly higher than securely attached individuals — not because the message was threatening, but because their codec is tuned to detect threat (Vanderbilt et al., 2025).

Meanwhile, avoidantly attached individuals tend toward the opposite distortion: they under-read emotional bids, dismissing signals of connection that the sender considers crucial. The message wasn’t unclear. The decoders simply weighted different features.

Cultural codecs add another layer. Agarwal’s (2025) research on AI-suggested politeness strategies revealed that native and non-native English speakers decode politeness markers in fundamentally different ways — a phrasing intended as warm in one cultural codec registers as formal or even cold in another.

A recent computational study, SocialVeil, quantified this problem directly: when cognitive-difference-induced communication barriers — semantic vagueness, sociocultural mismatch, and emotional interference — were introduced into conversations, mutual understanding dropped by over 45% on average (Xuan et al., 2026). Not because the speakers changed. Because the barriers between codecs became visible.

What AI Can Do Here

AI doesn’t have emotional baggage. It doesn’t decode through trauma. This makes it uniquely positioned to act as a translation layer between humans.

Not to replace human connection — but to make the invisible visible. To show you what your words might sound like through someone else’s codec.

“We’re not here to fix communication. We’re here to show that everyone has a different codec.”

This is not a speculative idea. The field of AI-Mediated Communication (AI-MC), formally defined by Hancock, Naaman, and Levy (2020), describes interpersonal communication in which “an AI agent modifies, augments, or generates messages to achieve communication goals.” Their research agenda at Stanford’s Social Media Lab maps out exactly the territory DeepConvos operates in — but with a critical difference in philosophy.

Most AI-MC research studies AI as a single-sided intervention: the AI rewrites your message before they see it. DeepConvos proposes something fundamentally different — a bilateral architecture where both sides have an AI agent that understands their codec, and the two agents perform a “handshake” to translate between mental models. Not one-way editing, but two-way understanding.

The Science Behind It

The empirical evidence that AI can function as a codec translator is growing rapidly:

Emotional tone detection. Hohenstein and colleagues (2023) conducted two randomized experiments studying algorithmic response suggestions (smart replies) and found that AI-assisted communication increased the use of positive emotional language and led conversation partners to evaluate each other as closer and more cooperative. The AI didn’t just change words — it shifted the emotional texture of the entire exchange.

Perspective-taking in AI systems. Wilf and colleagues (2023) demonstrated that large language models significantly improve their Theory of Mind capabilities when explicitly prompted to take another person’s perspective — a technique they called SimToM. Their findings showed substantial improvement over baseline when the model first filtered context through what each party would know and feel. This is precisely what DeepConvos’s Pigeon agents do: they model each user’s perspective before translating between them.

Needs-conscious design. Wolfe and colleagues (2025) co-designed a human-centered framework for AI-Mediated Communication that foregrounds users’ psychological needs rather than optimizing for surface metrics. Their framework, published at AIES, maps closely to DeepConvos’s commitment to preserving the intent behind a message rather than just polishing its surface.

The trust paradox. There is, however, a complication. Jakesch and colleagues (2019) documented the “Replicant Effect”: when people know that AI was involved in crafting a message, they trust it less — even when the message quality is objectively better. This finding is real and important. DeepConvos addresses it through mutual consent: both parties know and choose to use the system. When both sides opt in, the asymmetry that drives the Replicant Effect — one person secretly using AI — disappears.

The homogenization risk. Sourati and colleagues (2025) found that large language models can homogenize human expression, reducing individual voice distinctiveness. This is a legitimate concern for any AI-MC system. DeepConvos mitigates it by design: the system’s Communication Styles Inventory (CSI) profiles model each user’s unique communication style, and translation preserves individual voice rather than flattening it to a generic “AI voice.”

Why This Matters: The Four Modes

The codec problem isn’t just an intellectual observation — it maps directly to DeepConvos’s four modes of engagement:

Theory of Mind is the primary mode at work here. Understanding that others have different codecs is Theory of Mind: the recognition that another person’s beliefs, emotions, and interpretive frameworks differ from yours. Every time the system shows you how your words might land differently with someone else, it’s exercising — and building — your capacity for mentalizing.

Socratic Questioning enters when the system asks you to examine your own assumptions. “What was your last argument really about?” is itself a Socratic move. Before translating between codecs, the system first helps you see your own codec clearly — the assumptions you didn’t know you were making.

Awareness is the meta-layer: becoming conscious that you have a codec at all. Most people go through life assuming their interpretation is the message. The moment you realize “I’m not hearing what they said — I’m hearing what my codec produced” is a profound shift in self-awareness.

What’s Next

This is the first post in a series exploring the ideas behind DeepConvos. In upcoming posts, we’ll dive into:

• How attachment styles create different codecs — the neuroscience and psychology of why anxious, avoidant, and secure individuals hear fundamentally different messages
• The role of cultural context in message decoding — why politeness, directness, and emotional expression translate differently across cultures (and what AI can do about it)
• Practical AI-MC techniques you can use today — concrete strategies drawn from the latest research for becoming a better “codec translator” yourself

The conversation about communication is just beginning.

References

Agarwal, A. (2025). Exploring how AI-suggested politeness strategies influence email writing and social perception among native and non-native speakers. University of Waterloo. https://uwspace.uwaterloo.ca/bitstreams/586d99e6-dcde-4316-af32-b45dfd97ce75/download

Epley, N., Keysar, B., Van Boven, L., & Gilovich, T. (2004). Perspective taking as egocentric anchoring and adjustment. Journal of Personality and Social Psychology, 87(3), 327–339. https://doi.org/10.1037/0022-3514.87.3.327

Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-Mediated Communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. https://doi.org/10.1093/jcmc/zmz022

Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. https://doi.org/10.1038/s41598-023-30938-9

Jakesch, M., French, M., Ma, X., Hancock, J. T., & Naaman, M. (2019). AI-Mediated Communication: How the perception that profile text was written by AI affects trustworthiness. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3290605.3300469

Kelly, G. A. (1955). The psychology of personal constructs. Norton.

Keysar, B., Lin, S., & Barr, D. J. (2003). Limits on theory of mind use in adults. Cognition, 89(1), 25–41. https://doi.org/10.1016/S0010-0277(03)00064-7

Mikulincer, M., & Shaver, P. R. (2007). Attachment in adulthood: Structure, dynamics, and change. Guilford Press.

Sourati, Z., Ziabari, A. S., & Dehghani, M. (2025). The homogenizing effect of large language models on human expression and thought. arXiv preprint. https://arxiv.org/abs/2508.01491

Trope, Y., & Liberman, N. (2010). Construal-level theory of psychological distance. Psychological Review, 117(2), 440–463. https://doi.org/10.1037/a0018963

Vanderbilt, R. R., Brinberg, M., & Lu, Y. (2025). The impact of attachment style on communication frequency and language use in romantic partners’ text messages. Journal of Language and Social Psychology. https://doi.org/10.1177/0261927X251344949

Wilf, A., Lee, S. S., Liang, P. P., & Morency, L.-P. (2023). Think twice: Perspective-taking improves large language models’ theory-of-mind capabilities. arXiv preprint. https://arxiv.org/abs/2311.10227

Wolfe, R., Dangol, A., Kim, J., & Hiniker, A. (2025). Toward needs-conscious design: Co-designing a human-centered framework for AI-mediated communication. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES). https://doi.org/10.1609/aies.v8i3.36751

Xuan, K., Wang, P., Ye, C., Yu, H., August, T., & You, J. (2026). SocialVeil: Probing social intelligence of language agents under communication barriers. arXiv preprint. https://arxiv.org/abs/2602.05115