AI-Mediated Communication: What It Is and Why It Matters

You probably used AI today without noticing it.

Your email app suggested a reply. Your messaging app autocorrected a misspelling. A chatbot answered your customer service question. These interactions are so seamless that they’ve become invisible — like plumbing. You don’t think about the pipes until the water stops flowing.

But what happens when AI moves beyond correcting your typos and starts shaping what you say to other people? That’s the territory we’re entering now, and it has a name: AI-Mediated Communication.

Defining AI-MC

AI-Mediated Communication (AI-MC) was formally defined by Jeffrey Hancock, Mor Naaman, and Karen Levy in a landmark 2020 paper published in the Journal of Computer-Mediated Communication. Their definition is precise: AI-MC is “interpersonal communication in which an intelligent agent operates on behalf of a communicator by modifying, augmenting, or generating messages to accomplish communication goals” (Hancock et al., 2020, p. 89).

The key distinction — and it’s worth pausing on — is that AI-MC concerns human-to-human communication facilitated by AI. Not chatting with an AI. Not asking ChatGPT a question. But rather: AI sitting between two humans, altering what passes from one to the other. The AI is the medium, not the interlocutor.

This distinction matters because it places AI-MC squarely within the tradition of Computer-Mediated Communication (CMC) research — the field that has studied how technology shapes interpersonal dynamics since the early days of email and instant messaging. As Hancock and colleagues argue, the advent of AI-MC “requires re-evaluation — and potentially expansion — of many of CMC’s key theories, frameworks, and findings” (Hancock et al., 2020, p. 90). The AI isn’t just transmitting your message anymore. It’s transforming it.

More recently, Boyd and Markowitz (2026) introduced the Machine-Integrated Relational Adaptation (MIRA) model, which provides the field with a crucial conceptual expansion. MIRA distinguishes two roles that AI plays in human social ecosystems: relational partner (a direct interaction companion, like a chatbot) and relational mediator (shaping human-to-human communication, like smart reply). This dual framework helps us understand why the same technology — a large language model — can serve fundamentally different social functions depending on whether it’s talking to you or talking for you.

Three Pillars of AI-MC Research

Hancock et al. (2020) mapped out a research agenda that progresses from individual-level effects to relational dynamics to societal impact. Three areas of active research illustrate the breadth of this field:

1. Smart Reply and Predictive Text

The most familiar example of AI-MC is Gmail’s Smart Reply, introduced by Kannan and colleagues (2016) at Google. The system uses deep learning to generate semantically diverse response suggestions that users can send with a single tap. At launch, Smart Reply was responsible for assisting with roughly 10% of all mobile email responses — billions of messages flowing through an AI intermediary every day (Kannan et al., 2016).

But Smart Reply isn’t neutral. Hohenstein and colleagues (2023), in two randomized experiments published in Scientific Reports, found that using algorithmic response suggestions measurably changes both language and relationships. Participants who used smart replies increased their use of positive emotional language, communicated faster, and — crucially — their conversation partners evaluated them as closer and more cooperative. The AI didn’t just speed things up. It shifted the emotional texture of the exchange.

However, the same study revealed a critical tension: when people suspected that their partner was using AI-generated responses, they evaluated that partner more negatively — even when the message quality was objectively better. This echoes the “Replicant Effect” first documented by Jakesch and colleagues (2019): knowing that AI authored a message reduces perceived trustworthiness, even when the message content is superior.

Mieczkowski and colleagues (2021) found converging evidence in a controlled referential communication task: smart replies introduced a “positivity bias” into conversations — making language systematically more positive than the speaker would naturally produce. This finding raises a question that sits at the heart of AI-MC ethics: if the AI makes you sound nicer than you are, is the resulting communication authentic?

2. Sentiment-Aware Reframing

Beyond suggesting pre-written replies, more advanced AI-MC systems can detect the emotional tone of a message and propose alternative phrasings. The goal is to preserve the sender’s intent while reducing the receiver’s potential for negative interpretation.

This isn’t a hypothetical capability. Ziems and colleagues (2022), in work presented at ACL, introduced the task of “positive reframing” — neutralizing a negative perspective and generating a more positive one without contradicting the original meaning. Their benchmark dataset of over 8,000 sentence pairs demonstrated that neural models can learn to shift distorted text toward more constructive perspectives using strategies drawn from positive psychology: growth mindset framing, impermanence, neutralizing, and optimism. The insistence on meaning preservation makes this fundamentally different from censorship — the goal is translation, not suppression.

This approach maps onto what Wolfe and colleagues (2025) call Needs-Conscious Design — an AI-MC framework built on the principles of Nonviolent Communication (NVC). Through interviews with 14 certified NVC trainers and co-design sessions with 13 lay users, they identified three pillars of AI communication design that centers human relationships: Intentionality (the AI should facilitate purposeful expression), Presence (it should support rather than replace genuine attention), and Receptiveness to Needs (it should help both parties understand the underlying needs behind messages).

Wolfe’s team also identified a critical risk they named Empathy Fog: uncertainty over how much empathy, attention, and effort a user has actually invested in an AI-facilitated interaction. If the AI makes everyone sound empathetic, how do you know if the person is actually empathetic? This is a genuine design challenge that any responsible AI-MC system must address.

3. Codec Translation: The DeepConvos Approach

Most current AI-MC systems operate as single-sided interventions: the AI rewrites your message before the other person sees it. This creates an inherent asymmetry — one person has AI assistance, the other doesn’t even know it’s happening.

DeepConvos proposes something fundamentally different: a bilateral architecture where both participants have an AI agent — a Pigeon — that understands their unique communication style. The two Pigeons perform a “handshake,” translating between mental models. Not one-way editing, but two-way understanding.

This approach directly addresses three problems that pervade current AI-MC:

The authenticity problem. When AI rewrites only one side, the resulting message may not represent the sender at all. DeepConvos’s Communication Styles Inventory (CSI) profiles each user’s natural communication patterns and preserves individual voice during translation — addressing the homogenization risk documented by Sourati and colleagues (2025), who found that large language models systematically reduce individual voice distinctiveness when mediating human expression.

The transparency problem. Boyd and Markowitz’s (2026) MIRA model identifies interpersonal trust as one of four core principles governing AI’s relational impact. Trust requires transparency. In single-sided AI-MC, the receiver doesn’t know AI was involved — creating exactly the conditions under which the Replicant Effect operates (Jakesch et al., 2019). DeepConvos’s bilateral design means both parties opt in, eliminating the information asymmetry that erodes trust.

The empathy fog problem. Because both Pigeons model their respective users, the system can distinguish between “this person expressed concern” and “the AI made this person sound concerned.” The architecture preserves the signal of genuine human effort that Wolfe et al. (2025) correctly identify as essential to meaningful connection.

Ethical Considerations

AI-MC raises at least three fundamental ethical questions that the field is actively grappling with:

Message authenticity. If AI substantially rewrites your message, is it still yours? Hancock et al. (2020) frame this as a question about “self-presentation” — AI-MC gives communicators unprecedented ability to manage how others perceive them, but at the cost of authenticity. Nakano and colleagues (2025) found that when readers learn a text was AI-authored, their perception of both the text and its supposed author shifts significantly — suggesting that AI involvement fundamentally changes the social contract of communication.

Disclosure and consent. Should receivers know when AI has helped craft a message? The empirical evidence is clear: awareness of AI involvement changes evaluation (Hohenstein et al., 2023; Jakesch et al., 2019). But blanket disclosure creates its own problems — as Agarwal (2025) found in cross-cultural contexts, the social meaning of “AI-assisted” varies significantly across cultures and communication contexts. What reads as helpful technology in one cultural codec may register as dishonesty in another.

Default communication styles. AI-MC systems encode particular assumptions about what “good” communication looks like. Smart Reply’s positivity bias (Mieczkowski et al., 2021) privileges a particular Western, professional, conflict-avoidant communication style. Baumler and Daumé (2024) demonstrated that predictive text suggestions can reinforce gender stereotypes — the AI doesn’t just reflect existing biases, it amplifies them into new messages. Any AI-MC system must be designed with awareness that its defaults carry cultural and political weight.

DeepConvos addresses these concerns through three design principles: mutual consent (both parties know and choose to use the system), voice preservation (the CSI ensures translation preserves individual style rather than flattening it), and needs-centered design (the system optimizes for mutual understanding, not surface politeness).

The Future Trajectory

The trajectory of AI-MC is clear: it will become standard infrastructure for digital communication.

Consider the analogy. Spell-check was once a novelty. Then it became a setting you could toggle. Now it’s invisible — you’d notice its absence, not its presence. Grammar suggestions followed the same path. AI-MC is next.

Within the next decade, every messaging platform will likely incorporate some form of intelligent mediation — tone detection, cultural context adaptation, real-time translation between communication styles. A Pew Research Center survey of technology experts predicted that by 2035, AI will be “as invisible and integral to the fabric of everyday life as Wi-Fi” (Nadeem et al., 2023) — and interpersonal communication is one of the most intimate fabrics there is.

But the question isn’t whether AI-MC will become pervasive. It’s whose values it will encode. Will AI-MC systems default to corporate-friendly positivity, smoothing away legitimate grievances? Will they homogenize communication across cultures, creating a global “AI voice” that erases local expression? Or will they be designed — as DeepConvos envisions — to enhance understanding while preserving the irreducible individuality of each person’s communication style?

The research is clear on one point: the design choices made today will shape the communication norms of tomorrow. As Hancock and colleagues wrote: AI-MC “raises new questions about how technology may shape human communication” (Hancock et al., 2020, p. 89). Those questions aren’t theoretical anymore. They’re being answered — in code, in product decisions, in the defaults that billions of people encounter every day.

The conversation about how we communicate is itself being mediated by AI. The least we can do is have that conversation deliberately.

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

Baumler, C., & Daumé III, H. (2024). When stereotypes GTG: The impact of predictive text suggestions on gender bias in human-AI co-writing. arXiv preprint. https://arxiv.org/abs/2409.20390

Boyd, R. L., & Markowitz, D. M. (2026). Artificial intelligence and the psychology of human connection. Perspectives on Psychological Science. https://doi.org/10.1177/17456916251404394

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

Kannan, A., Kurach, K., Ravi, S., Kaufmann, T., Tomkins, A., Miklos, B., Corrado, G., Lukacs, L., Ganber, M., Bober, P., & Toutanova, K. (2016). Smart Reply: Automated response suggestion for email. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 955–964. https://doi.org/10.1145/2939672.2939801

Mieczkowski, H., Hancock, J. T., Naaman, M., Jung, M., & Hohenstein, J. (2021). AI-Mediated Communication: Language use and interpersonal effects in a referential communication task. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 1–14. https://doi.org/10.1145/3449091

Nadeem, R., Anderson, J., Auxier, B., Rainie, L., & Vogels, E. A. (2023). As AI spreads, experts predict the best and worst changes in digital life by 2035. Pew Research Center. https://www.pewresearch.org/internet/2023/06/21/as-ai-spreads-experts-predict-the-best-and-worst-changes-in-digital-life-by-2035/

Nakano, H., Takezawa, J., Matulic, F., Yang, C.-L., & Yatani, K. (2025). Understanding reader perception shifts upon disclosure of AI authorship. arXiv preprint. https://arxiv.org/abs/2510.24011

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

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

Ziems, C., Li, M., Zhang, A., & Yang, D. (2022). Inducing positive perspectives with text reframing. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 3120–3132. https://doi.org/10.18653/v1/2022.acl-long.257