Artificial intelligence in communication impacts language and social relationships

Helen: Six point seven billion emails per day. That's how many messages were already being written by artificial intelligence as of 2017 — just from Gmail's smart reply feature alone. Almost nobody had asked what that does to the people on either end of those conversations. Adam: Which is strange, because the question seems obvious. If I am sending you a message that an algorithm suggested, is what you receive still really from me? Helen: That's exactly the tension that Hohenstein and colleagues set out to resolve. They ran two randomized experiments using a custom messaging platform called Moshi, pairing workers on Mechanical Turk to have live text conversations about labor policy. Half had access to Google-generated smart replies, while the other half did not. People used the smart replies — they made up fourteen point three percent of sent messages, and having them available sped up conversations by ten point two percent more messages per minute. Adam: Speed I can explain away. Convenience is convenient. What I’d have expected next is some kind of flattening, conversations that feel more generic and partners who seem more distant. Helen: The data goes in the opposite direction. Using smart replies made the sender's own language more positive. This was true not just for the AI-generated messages but also for the messages that the person wrote themselves afterward. The Valence Aware Dictionary and Sentiment Reasoner sentiment instrumental variable estimate is 0.18, with a p-value of 0.045. The effect holds even when you remove the smart replies from the analysis. Adam: Wait — their own messages became warmer? Helen: Their own messages became warmer. Their partners rated them as more cooperative and more affiliative. The instrumental variable estimates show that cooperation is fifteen point sixty-six, and affiliation is twenty-one point seventy-nine, with both being significant below a p-value of 0.02. Adam: That's the step I'm stuck on. The partners didn't know whether smart replies were being used. So how is warmth being transmitted through messages that weren't actually generated by the sender? Helen: The paper's answer is that the receiver's experience is real regardless of its origin. If the words are warmer, the conversation feels warmer — full stop. The sender's internal process is invisible to the partner. Adam: So, authenticity is a property of the sender's experience, not of the receiver's. Helen: At least when the receiver doesn't know. That's exactly where the second experiment comes in. Hohenstein and colleagues told participants explicitly that their partner would be using an AI mediated messaging system. The moment that knowledge enters the room, the entire benefit inverts. The more a partner was perceived to use smart replies, the less cooperative they were rated, with a t value of negative 9.89 and a p-value below 0.0001. They were also rated as less affiliative, with a t value of negative 6.90. Adam: The same messages, the same warmth. However, knowing that artificial intelligence was involved changed how that warmth was perceived. Helen: Dominance ratings weren't affected either way. The negative effect is specifically about closeness and cooperation. The prevailing assumption that AI communication is somehow less authentic is doing all that work. Adam: That leaves a genuinely uncomfortable place to end up. If you disclose that you are using smart replies — which ethically you probably should — you erase the relational benefit that the tool just created. If you do not disclose, you get the warmth, but you are deceiving your partner about where your words came from. Helen: Hohenstein and colleagues call it out directly. Most current systems do not surface this to receivers at all, and the paper highlights that transparency norms and design choices need to catch up to the scale of use. The longitudinal question matters here too. These are single conversations between strangers. Whether the effect compounds or decays in an ongoing relationship is genuinely unknown. Adam: What they've nailed down is the direction of each force. Hidden artificial intelligence warms a first impression. Visible artificial intelligence poisons it. The disclosure trap is real. Helen: This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.