Small talk is a ritual, not a talent

At four London train stations, researchers ran a field experiment with 148 commuters. Each one rated how they expected to feel about the ride ahead, and the split in that data is where the problem sits. Commuters asked to imagine they had successfully talked with a stranger rated the prospect at 1.00 on a positivity index. Those asked to imagine they tried to talk with a stranger rated it 0.48 (Schroeder, Lyons and Epley, Journal of Experimental Psychology: General, 2021). Same train, same stranger, same conversation. The only variable was whether the act of opening it was in the picture. Entering the ritual is the hard part, not performing once you are in it.

My opinion, which you can argue with by the end: the advice industry sells you better content when your problem was never content.

The builds in this article are made in YourCrush.AI, where the traits list, a 2,400-character description and a 500-character greeting let you make a practice partner as easy or as difficult as the rehearsal needs.

What is small talk?

Small talk is a low-stakes signalling ritual between people with no established relationship. The topic stays deliberately trivial so both sides can show willingness and safety without risking anything. The weather is not the message. The willingness to stand there discussing the weather is the message, and both people know it, which is why nobody walks away disappointed by the informational content.

Read structurally, the exchange has four beats: an opening bid tied to the shared situation, an acceptance, two or three turns of reciprocated low-stakes disclosure, and an exit. Any of those can be learned separately. None of them requires you to be interesting.

How YourCrush.AI lets you drill a conversation before you need it

YourCrush.AI builds all three of those characters from the fields above with no code and no prompt engineering: you type the name, pick the behaviour mode, write the description and backstory, and the character exists. YourCrush.AI will also generate a description, backstory, greeting and trait set from a name and an image, which is faster if you would rather edit than write from scratch. YourCrush.AI is free to join, and signed-in members chat without a message or conversation limit, so a rehearsal can be run again as often as you like. Visitors can try a character from the explore page in a short guest session before signing up. YourCrush.AI holds pinned memories as a list you can read and edit, so a detail the character has to keep across sessions is one you picked rather than one the model happened to retain.

One thing to know before you start. YourCrush.AI is an entertainment product for adults, it is 18+ only, and it is not a mental-health service. YourCrush.AI employs no clinicians, assesses nothing and treats nothing.

If the reason you are here is that you moved city and the conversations are not turning into anything, the structural version of that problem is in why adult friendship is harder than school friendship, and the state of stranger chat after Omegle closed is in what replaced Omegle and whether any of it is safe.

Why is starting a conversation harder than having one?

Because the fear lives in the approach, and the research separates the two cleanly. In the London experiment above, the effect appeared only in the conversation condition: people imagining that they tried versus succeeded at conversation differed significantly, t(146) = 2.58, p = .011, d = 0.43, while imagining trying versus succeeding at sitting in solitude or commuting normally produced no difference at all. Nothing else on that train was hard to attempt. Only the opening was.

The rejection people brace for does not arrive at anything like the predicted rate. In the original Chicago studies, 64 Metra commuters estimated that fewer than 47% of fellow passengers would be willing to have a conversation, and rated starting one as fairly difficult, yet every participant who returned a survey reported talking with the person they approached and not a single person reported being rebuffed (Epley and Schroeder, Journal of Experimental Psychology: General, 2014, 143(5)). The seven people in the earlier experiment who failed the task all gave the same reason: nobody sat next to them.

Here is the part the popular version leaves out: the London replication did not reproduce the headline finding. Chicago commuters predicted a conversation would make the ride worse. London commuters predicted the opposite, expecting a more positive commute in the conversation condition (M = 0.73) than in solitude (M = 0.40), and they were right. What London did show is a compliance gap: 78.0% of people assigned to talk reported following the instruction, against 97.1% assigned to solitude and 100% assigned to commute normally, in the same 2021 paper. People who expected to enjoy it still could not make themselves start. Treat "people wrongly think talking to strangers is unpleasant" as the weaker claim and "people find starting disproportionately hard" as the strong one.

Why do you think it went worse than it did?

Because you systematically underestimate how much the other person liked you, and this has a name and a measured size. In the study that named it, 36 community members recruited near Yale had five-minute conversations with a stranger, then rated both sides. Liking for the partner averaged 5.82 on a seven-point scale; belief about how much the partner liked them averaged 5.17, t(34) = -5.83, p < .001 (Boothby, Cooney, Sandstrom and Clark, Psychological Science, 2018, 29(11)). Since both people cannot on average be liked less than they like, the gap is an error, and the authors call it the liking gap.

Two findings inside that paper go further than the headline. Shyness moderated the effect, so the shyer the participant, the bigger their underestimate, t(30) = 3.29, p = .003. And in a year-long study of 102 first-year Yale students rating their dorm suite mates at five points across an academic year, the gap held at September, October, December and February, and vanished only at the May measurement. So this is not a first-impression jitter that resolves over coffee. It survived months of living together.

The limits are real. These are small samples of students and community members at universities in the United States and the United Kingdom, self-reported on Likert scales, and the authors offer three competing explanations for the May result themselves, including that students were by then deciding who to live with the following year.

Do conversation starters work, or is the advice wrong?

Lists of good conversation starters aren't wrong so much as aimed at the wrong target. They fix the first line, when the measured effect sits in the second and third. The best evidence on what actually moves liking comes from a set of studies on question-asking. In the first, 398 people in 199 pairs had a 15-minute anonymous text conversation, with one member of each pair privately told to ask at least nine questions or at most four. Partners of the high askers liked them more, 5.79 against 5.31 on a seven-point index, t(197) = 2.47, p = .014, d = 0.35 (Huang, Yeomans, Brooks, Minson and Gino, Journal of Personality and Social Psychology, 2017, 113(3), 430-452).

The kind of question decided it. In a separate analysis of 110 speed daters across three sessions, each going on 15 to 19 four-minute dates and producing 1,961 yes-or-no second-date decisions, the authors trained a classifier to spot follow-up questions. People asked 9.80 questions per date on average, of which 4.51 were follow-ups. A higher follow-up rate predicted more partners saying yes to a second date, while other question types did not, in the same 2017 paper. A follow-up question is the one you cannot prepare, because it can only be built from what the other person just said.

Three limits before you take this too far. The experimental effect is modest at d = 0.35, the speed-dating analysis is observational, and naturally occurring variation in question rate correlated with partner liking at only r = .11 in the first study, which was not significant. Being instructed to ask more worked. Simply being a person who asks a lot did not show the same effect at that sample size. One thing held across both studies: the askers had no idea it was working, predicting no difference in how much they were liked, p = .544.

What are the four moves that actually work?

Four, in the order they happen, each built from the evidence above rather than from a list of topics.

1. Open with the situation, not the person. An observation about the room, the queue, the speaker who just overran, or the machine that is not working gives the other person something to agree with that costs them nothing.
2. Ask a question that cannot be answered in one word, then ask a second question about their answer. One question plus real attention to the reply beats nine prepared ones.
3. Volunteer a specific instead of a generality. "I came for the afternoon session and I am regretting it" gives a follow-up to work with; "it's been good" gives nothing.
4. Close the exchange on purpose, with a line that names the ending. "I'm going to grab a seat before the room fills, good to meet you" completes the ritual. An exchange that trails off is the one both people remember as awkward, and the liking gap says your memory of that awkwardness is inflated.

Does practice improve social skills, or just confidence?

Repetition reliably lowers the fear and reliably raises how capable you believe you are. The evidence that it changes how much you expect to enjoy the conversation is weaker. In a week-long study, 286 students at one American and one British university played a scavenger hunt on their phones, with 198 assigned to find and talk to a stranger matching a daily description and the rest assigned only to observe one (Sandstrom, Boothby and Cooney, Journal of Experimental Social Psychology, 2022, volume 102). By the end of the week the talkers expected to be rejected by significantly fewer people than at the start, p < .001, still expected less rejection a week after the study ended, p < .001, and rated their own conversational ability higher, p < .001, while the observers did not change.

Now the honest half. Against the control group, the talkers showed no significant difference in how much they expected to enjoy conversations, p = .83 at the end and p = .99 at follow-up, and none in how positive an impression they expected to make, p = .63 and p = .79. The authors flag that themselves as an unexpected limitation: the intervention cut the negative more than it lifted the positive, and they do not know why. They also note the design was gamified, so transfer to ungamified practice is untested. What survived is that daily expectations grew more accurate day by day, which argues for repetition over preparation.

One boundary has to be drawn explicitly here. Graded exposure is a component of clinical treatment for anxiety, planned and supervised by a therapist: "in vivo exposures are completed both in the therapy session with the encouragement of the therapist and outside the session as homework" (Kaczkurkin and Foa, Dialogues in Clinical Neuroscience, 2015, 17(3), 337-346). What the rest of this article describes is not that, is not a version of that, and is not a substitute for it. It is rehearsal for a social ritual, the way you would rehearse a toast. If fear of talking to people is stopping you doing things you want to do, that belongs with a clinician rather than with a chat app: a general practitioner, a licensed therapist or clinical psychologist, or in England NHS talking therapies for anxiety and depression, which states that you can refer yourself without speaking to a GP and that you do not need a diagnosed condition to do it.

How do you build a practice partner in the character creator?

You fill nine fields, and the one that decides whether the rehearsal is worth anything is the trait list. A practice partner is not a supportive one. A character who is warm, encouraging and easy to talk to teaches you nothing you did not already know, because the thing you are trying to survive is a person who is none of those. Arrogant, Cold-hearted, Impulsive and Aggressive are on the trait list for exactly this, and a practice build that avoids them is not practice.

Here are the fields and the limits they enforce. Name takes 40 characters, and behaviour mode is picked first from natural, flirty, playful and sexual, with every build below on natural, the non-romantic one. Description takes 2,400 characters in the character creator and says who they are, although the server behind the form accepts 3,000. Backstory takes 2,400 in the form as well, with the server accepting 5,000, and it says what already happened between the two of you. Greeting message takes 500 and is the first thing the character says. Traits come from a list of 32, with the form asking for 3 to 7 and a hard cap of 12 selected. Speaking style takes up to 4 words of up to 18 characters each, category is Anime or Human, and image generation is off by default and stays off for all three of these.

Two of those do most of the work. The backstory is where continuity comes from, and most people leave it empty and then wonder why the character has no history with them. The greeting decides the first thirty seconds, because a greeting that opens mid-scene forces you to respond to something, while a greeting that says hello hands the work back to you.

Build one, the neutral stranger at a work event

This is the baseline, and the point of it is that nothing helps you. The character is neither hostile nor warm, will not carry the conversation, and will not punish a weak opening, which makes it the build for practising moves one and two on their own.

Description, 516 of the 2,400 characters allowed:

Priya is a procurement analyst in her late thirties at a mid-size logistics firm. She is at this conference because her manager sent her. She does not know anyone in the room, she is not enjoying standing alone, and she will not start a conversation herself. She answers what she is asked, adds one detail, then stops and waits. Asked a closed question, she gives a closed answer. Asked nothing, she looks at her phone. She is neither warm nor cold towards you; she is a stranger with no reason to invest in you yet.

Backstory, 413 characters:

We have not met. This is the second morning of a two-day industry conference in a hotel function room, during the coffee break between the ten o'clock and eleven o'clock sessions. She has been standing near the urns for four minutes holding a cup she has already finished. Her colleague cancelled, so she came alone and has spoken to nobody since the badge desk. She has no opinion about you and no memory of you.

Greeting message, 310 of the 500 allowed:

She is standing a metre from the coffee urn, holding an empty cup, watching the lanyard of someone walking past. When you step up beside her she glances over, registers you, and looks back at the urn. She says nothing. The next session starts in eleven minutes and there are two people behind you in the queue.

Build two, the interviewer who has decided you are wasting his time

This is the hard one, and it is why the difficult traits exist. The character interrupts, answers enthusiasm with silence, and rewards only short specific answers. Rehearsing against him is uncomfortable, which is the point.

Description, 546 characters:

Gerald runs a forty-person engineering group and has interviewed hundreds of people. He thinks this interview is an hour he did not agree to spend. He interrupts. He asks a question and, before you have finished, asks a second one that contradicts the first. He answers enthusiasm with silence and detail with "and?". He is not stupid, he is fast and bored, and a short, specific, unbothered answer earns exactly one more question. He never apologises for his tone and never softens it. If you fill his silence by rambling, he picks up his phone.

Backstory, 391 characters:

We have met once, for about ninety seconds, when reception handed you over. It is twenty to five on a Thursday and you are the last of four interviews on his schedule. He read your CV in the lift. His own preferred candidate was rejected by the panel last week, and he has said in two meetings since that this hiring process is theatre. He has forty minutes booked and intends to use twenty.

Greeting message, 285 characters:

He sits down without shaking your hand, puts his phone face down, and reads two lines of your CV. "Right. I've got twenty minutes, not forty, so let's skip the warm-up." He looks up. "Tell me about the last thing you built that didn't work. And don't tell me what you learned from it."

Build three, the person who will not take no

Build this one if you are practising a refusal rather than an opening. The character never raises their voice and never gets angry, which is what makes them hard: they treat a reason as a problem to be solved and stop only when a refusal has nothing attached to it.

Description, 538 characters:

Dani has been your neighbour for two years and treats every request as the opening of a negotiation. They ask for favours in a friendly voice, and when you decline they do not argue: they go quiet, then make the ask smaller, then mention something they once did for you. They are genuinely likeable and they genuinely will not stop. They never raise their voice and never get angry. They hear a no with a reason attached as a problem to be solved, and the only thing that ends the conversation is a refusal with no reason attached at all.

Backstory, 413 characters:

We have known each other since you moved into the building. Dani fed your cat for a weekend about eighteen months ago and has mentioned it four times since. You have lent them a car, a drill and a folding table, and the table never came back. Last month you agreed to drive them to the airport at five in the morning and regretted it before you reached the motorway. You have never once said no without an excuse.

Greeting message, 332 characters:

"Hey. Hey, quick thing." Dani is in your doorway with car keys already in hand and a bag over one shoulder. "I need someone to sit in for a delivery on Saturday morning, just Saturday, and you're literally the only person on this floor who's ever home. It's two hours. Three, max." They smile at you. "You haven't got plans, right?"

Frequently asked questions

What do I say after hello to someone I don't know?

Say something about the situation you are both in, then ask a question about their answer rather than moving to a new topic. The 2017 question-asking research found that follow-up questions, not opening lines, predicted whether speed-dating partners wanted to meet again.

Are conversation starter lists actually any good?

They are useful for the first ten seconds and irrelevant after that. The measured effect on liking came from asking follow-up questions built out of what the other person just said, which by definition cannot be on a list.

Why do I always think a conversation went badly?

Because almost everyone does, and it has been measured. After five-minute conversations with strangers, people rated how much they liked their partner at 5.82 out of 7 and how much they thought their partner liked them at 5.17, and shyer people underestimated by more.

Does practising with an AI character make me better at this with people?

The honest answer is that nobody has tested that specific claim, and this article does not make it. What has been tested is repeated real conversations with real strangers over a week, which reduced expected rejection and raised self-rated conversational ability but did not significantly change expected enjoyment.

Why would I build a rude character instead of a nice one?

Because a supportive character rehearses nothing. The difficulty is the content of the practice, which is why the interviewer build above uses Arrogant, Cold-hearted, Impulsive and Sarcastic rather than Calm and Supportive.

When should I talk to a person instead of an app?

When fear of social situations is stopping you doing things you want to do. That belongs with a GP, a licensed therapist or clinical psychologist, or a service like NHS talking therapies for anxiety and depression, which takes self-referrals without a GP and without a diagnosis.

More in connection and wellbeing

• Start here: AI friend: what a platonic AI companion is actually used for
• AI therapist: what the trials found and what regulators did
• 60 journaling prompts, grouped by what you need them for

Editorial Team, Content Strategist. Last updated: 2026-10-05