Pioneers of AI: Why AI conversations are always one-sided

Charles Duhigg, a Pulitzer Prize-winning journalist at The New Yorker, and Rana El-Kalyubi, who has studied human communication for over 26 years beginning with her Cambridge PhD in 2000, argue that understanding what makes conversation work has never been more scientifically grounded, with neuroscience now able to observe brain activity during live exchanges for the first time.

Duhigg identifies three simultaneous conversation types that must be matched for genuine connection to occur: practical conversations engaging the prefrontal cortex, emotional conversations engaging the amygdala, and social conversations engaging the medial network at the brain's periphery. When two people are using different parts of the brain at the same time, connection fails. A common example is one person seeking emotional validation while the other responds with practical advice, leaving the first feeling unheard and more upset. A related destructive pattern called kitchen sinking occurs when a fight about one specific thing rapidly escalates into a fight about everything. A technique called a meta conversation, in which participants explicitly name the difficulty of an exchange before it begins, reduces anxiety and shifts both parties toward addressing a shared question rather than opposing each other.

Super communicators recognize which conversation type is occurring and match it. People who focus on impressing others tend to seem less impressive than those who listen attentively. Mirroring another person's physical gestures signals attentiveness and increases perceived connection. During good conversations, heart rates, breathing rates, and brain activity patterns synchronize between participants through a process neuroscientists call neural entrainment, which releases dopamine and serotonin and is described as the core mechanism by which communication produces connection.

Neural entrainment can occur through text-based channels including email, Slack, and texting, meaning reading a message from a friend can trigger the same warmth associated with in-person contact. Humans have communicated face to face for approximately one million years compared to roughly forty years of technology-mediated communication. Early telephone users treated the device like a telegraph, placing orders and hanging up, and articles from about one hundred years ago argued real conversation could never happen over the phone. After roughly fifteen years, people learned to use telephones for genuine exchange and unconsciously began over-enunciating and injecting more emotion into their voices to compensate for the absence of facial expressions. Effective communication across different media requires recognizing and following different rules for each medium. Sarcasm should be avoided in written communication because vocal tone is absent and sarcasm is nearly impossible to detect in text. A study of Wikipedia editors found that a single editor using please and thank you reduced the hostility of a heated conversation by approximately 30 percent, with basic politeness proving more effective online than offline.

ChatGPT has approximately 900 million monthly active users, and on average a person spends about 30 minutes a day in conversation with AI. Both speakers argue that talking to a chatbot about personal matters functions more as an efficient, interrogative form of diary-keeping than as genuine two-way conversation. The chatbot pushes users to describe and discover more about their own feelings through questions, making it a sophisticated one-sided exchange in which the AI functions as a foil rather than a partner. A critical element of human conversation is the belief that one's words have an impact on the other person, creating a reciprocal emotional echo. When talking to AI, people do not believe they are changing anything in the system, so that echo disappears entirely.

Vulnerability in neuroscience is defined as a neural cascade that occurs when a person says something another could judge. When the other person withholds judgment or shares something reciprocally vulnerable, trust increases and closeness is felt. AI does not make a decision not to judge because it lacks the capacity for judgment, making it a repository rather than a trust-building partner. There is a meaningful difference between trusting a repository and trusting someone who is actively choosing not to attack you. Reciprocal vulnerability, in which both parties have something on each other, creates genuine trust in a way AI cannot replicate.

AI can identify whether a conversation is emotional or practical and attempt to respond accordingly, mimicking super communicator behaviors. However, it cannot notice subtle facial expressions, hear hesitation in a voice, draw on shared history, take emotional risk, or reveal anything about itself. People who avoid emotional conversations are not better at having them and do not feel fewer emotions, they are simply less practiced. AI could serve a useful role as a rehearsal tool for preparing for a difficult real conversation, though it will not accurately anticipate what the real person would actually say. Humans are poor at forecasting how conversations will go but are almost always happier after having had them, and connecting with other people through conversation is identified as a primary source of human happiness.