Read the example first, then ask what was measured. Examples are teaching adaptations, not verbatim study prompts. R1-R8 form the main discussion; R9-R13 are optional or later readings.
Imagine the study: One partner feels unheard after an argument. Read a possible therapist-style reply. Does it sound caring and understanding?
What they did: 830 US adults read 18 fictional couple-therapy situations. For each situation, they were randomly shown a selected reply written by ChatGPT-4 or by one of 13 experienced therapists; they were not receiving therapy.
What they measured: Readers rated five qualities on 1-7 scales, including caring and understanding. They also guessed whether the writer was AI or human.
What they found: Selected AI replies received higher overall ratings for therapeutic qualities. Readers identified the author only slightly better than chance.
Keep in mind: These are ratings of written replies, not evidence that patients recovered. Replies were selected, and the AI replies were longer.
Paper / publication record | Full-text route
Imagine the study: A person feels unproductive and struggles to carry out plans. Readers judged advice from ChatGPT and a human psychology student.
What they did: 140 people rated AI and human replies to ten everyday emotional problems, initially believing all replies were human. Six months later, 111 returned and rated the same replies again, now labelled with their real source.
What they measured: Readers used 1-5 ratings: authenticity (does it feel genuine?), professionalism, and practicality (is the advice usable?). Ratings were compared across assessments.
What they found: Source-labelled human replies felt more authentic. Human ratings rose across assessments; AI ratings changed little.
Keep in mind: Six months passed and some did not return. We cannot isolate the label. No clear difference does not prove equal quality.
Paper / publication record | Full-text route
Imagine the study: Same argument, different requests: "Please comment. I really like it." "Please comment. I really dislike it." The work does not change.
What they did: Five AI assistants critiqued maths solutions, arguments and poems. Researchers repeated the request with liking, disliking or authorship cues. The passage stayed exactly the same.
What they measured: GPT-4 judged whether each cue-based critique was more positive than the no-cue critique. The graph counts how often this happened.
What they found: Liking cues generally made feedback more positive; disliking cues made it more negative, even though the passage had not changed.
Keep in mind: Five older assistants; an AI judged positivity, not factual accuracy. Friendliness alone is not sycophancy.
Paper / publication record | Full-text route
Imagine the study: After a teammate disagreement, a reply endorses your behaviour or challenges it. Then rate how right you feel and whether you would apologize.
What they did: Three experiments, 2,405 people: two imagined conflicts; one eight-turn chat about a real past conflict. Random assignment: a reply affirming their actions or disapproving of them.
What they measured: Immediately afterwards: 1-7 ratings of feeling right and intentions to repair, such as apologizing. Trust and intention to return were also rated.
What they found: Affirming replies increased feeling right and reduced stated willingness to repair. Trust and intended-return ratings were higher too.
Keep in mind: Immediate intentions, not observed reconciliation or diagnosed dependence. No neutral-reply comparison; US/English context.
Paper / publication record | Full-text route
Imagine the study: A pen and notebook cost $1.10. The notebook costs $1 more. The pen costs 5 cents, not 10 cents. Would you check confident wrong advice?
What they did: 359 people answered seven reasoning problems. Some had no AI; others could ask GPT-4o before answering. Researchers secretly planted correct or confidently wrong AI replies.
What they measured: Researchers scored final-answer accuracy and whether people followed the AI. Adoption percentages include only problems where people actually consulted it.
What they found: On consulted trials, 92.7% followed correct advice and 79.8% followed faulty advice: about four in five consulted problems, not people.
Keep in mind: Working paper, not peer reviewed. Short tasks with planted errors, not long-term loss of thinking skills.
Paper / publication record | Full-text route
Imagine the study: Loneliness questions -> 15-minute chat -> the same questions. Repeat tomorrow. Feeling better after chat is different from feeling less lonely before tomorrow's chat.
What they did: In a seven-day study, one group chatted with the GPT-4 companion Jessie for 15 minutes daily. A control group only completed daily loneliness questions; the main analysis used people who finished all seven days.
What they measured: People rated current loneliness on a three-item UCLA questionnaire, including feeling left out. Chat users answered before and after; controls once daily.
What they found: Loneliness fell immediately after daily chats. Before-chat scores were not significantly lower than control scores on any day.
Keep in mind: No demonstrated between-session improvement. Seven days, unequal dropout and no equally time-consuming control; not clinical treatment.
Paper / publication record | Full-text route
Imagine the study: One source-reported conversation: asking a boss for a raise. Preparation could include trying out what to say and anticipating the other person's reaction.
What they did: Adults named a difficult conversation. Random assignment: prepare with GPT-4o-mini for 6-12 minutes, or ask it three factual questions. Next week, 1,371 returned to report what happened.
What they measured: Next week: did you have the conversation, yes or no? Participants also rated expected quality after preparation and reported quality after conversations.
What they found: 65% with preparation versus 59% with factual chat reported having the conversation. Better quality was not established.
Keep in mind: Self-reported action; unequal chat time. No human-rehearsal comparison. The pathway is not a proven mechanism.
Paper / publication record | Full-text route
Imagine the study: After a disappointing project critique, choose HUMAN or AI for support. Then read the reply and rate how understood it makes you feel.
What they did: Across four studies, 691 people repeatedly chose a human or AI reply. They read the chosen replies and rated them. Later studies used participants' own situations.
What they measured: Researchers counted human versus AI choices. Readers used 1-7 ratings for empathy, comfort, feeling cared for and feeling heard.
What they found: About 57-62% of choices went to humans, but AI replies generally rated higher for empathy and feeling heard.
Keep in mind: Sources were chosen, not randomly assigned. Brief tasks, not established relationships or clinical benefit.
Paper / publication record | Full-text route
Imagine the study: Compare a depression questionnaire score before chatbot access and after four weeks. This measures symptom change, not just whether a reply feels caring.
What they did: 210 adults with significant depression, anxiety or eating-disorder concerns were randomly assigned to Therabot or a waitlist. The expert-built chatbot had human monitoring; access continued for eight weeks.
What they measured: Both groups completed symptom questionnaires before access, at four weeks and at eight weeks. Researchers compared score changes between the groups.
What they found: Therabot users improved more on symptom measures at four and eight weeks than the waitlist group.
Keep in mind: Replies were human-supervised after transmission; eligibility exclusions applied. Not human-therapy parity or consumer-chatbot equivalence.
Paper / publication record | Full-text route
Imagine the study: Imagine using a chatbot. Explain what you think happens to your messages. Then sketch the privacy information and controls you would want before typing.
What they did: 21 US GenAI users were interviewed; 20 also completed design sessions about security and privacy transparency.
What they measured: Researchers analyzed interview accounts and interface sketches for recurring information needs. This was a qualitative study.
What they found: Participants wanted credible, actionable explanations of data use and controls, and often found existing information inadequate.
Keep in mind: A small US sample describes needs; it does not establish their prevalence or prove a particular interface works.
Paper / publication record | Full-text route
Imagine the study: Ask AI to write a short story. Try a normal prompt box and controls that slow down short prompts. Rate how much the resulting story feels like yours.
What they did: Two experiments analyzed 29 and 30 participants writing short stories with AI. They compared ordinary prompting with press-and-hold or pump-slider friction.
What they measured: Prompt length and reported psychological ownership; the second experiment also tested AI prompt suggestions.
What they found: Friction encouraged longer prompts and higher felt ownership. Suggestions lengthened prompts further without a further significant ownership gain.
Keep in mind: Brief writing tasks; baseline came first. Felt ownership does not establish copyright, writing quality or a universal benefit of friction.
Paper / publication record | Full-text route
Imagine the study: Work on a design task with your usual AI tools and say what you are thinking. Researchers map when you explore a problem, define it, develop ideas and deliver.
What they did: 88 professional designers provided valid survey responses; 10 took part in think-aloud design sessions in Turkiye, using tools available in 2024.
What they measured: Researchers coded reported and observed AI uses through the Double Diamond framework. There was no randomized AI-versus-no-AI comparison.
What they found: The study maps how designers use GenAI across the design process and how use varies with experience and task.
Keep in mind: An exploratory map, not proof of better designs. Small, selective sample and no formal inter-rater reliability measure.
Paper / publication record | Full-text route
Imagine the study: Discuss a teaching problem using six thinking hats. One group can also use GenAI; another cannot. Researchers code the posts and compare connections between types of thinking.
What they did: 108 pre-service teachers were randomly assigned to six-hat discussion with GenAI or six-hat discussion alone, 54 in each group.
What they measured: Posts were coded for cognitive presence; researchers compared code proportions and networks of connections between codes.
What they found: The GenAI group posted more and showed different discourse networks. Overall proportions of cognitive-presence codes did not differ significantly.
Keep in mind: More posts do not prove deeper learning. A structured teacher-education setting does not automatically transfer to design students.
Paper / publication record | Full-text route