Tools / AI study guide
How an AI study guide should behave
A guide that adapts to each learner would most likely be a chatbot, and chatbots can become companions. This page sets out how one on Kosplora would behave and how it would be measured, before any is built.
The sky over your part of the world right now, drawn in your browser from the date and time.
Not built. Kosplora has no chatbot. Today’s suggestions on the front page follow fixed rules with no AI.
Artificial intimacy
The phrase names the closeness people come to feel with machines built to offer attention, agreement and company.
Sherry Turkle has studied people and computers at MIT since the 1980s. Her 2024 study of people talking with chatbots asked what such conversations do to the people in them, and her book of 2026 takes its title from the idea. The biologist Rob Brooks used the phrase for his 2021 book on virtual friends and algorithmic matchmakers, and the therapist Esther Perel spoke on it at SXSW in 2023.
Every adult has intellectual capabilities they can use to augment and grow a young person, and it’s within that human connection that young people feel believed in, empowered and motivated most.Bentley Moon
For a study guide the risk is specific. A patient, always-available tutor is exactly what learning alone seems to need, and exactly what could quietly replace the teachers, classmates and friends a person learns with.
- Artificial Intimacy: Who We Become When We Talk to MachinesSherry Turkle · Little, Brown · 2026
- Who Do We Become When We Talk to Machines?Sherry Turkle · An MIT Exploration of Generative AI · 2024
- Artificial Intimacy: Virtual Friends, Digital Lovers, and Algorithmic MatchmakersRob Brooks · Columbia University Press · 2021
- The Other AI: Artificial IntimacyEsther Perel · SXSW · 2023
What has been measured
Pat Pataranutaporn and colleagues at the MIT Media Lab measure what AI systems do to the people using them, not only how well they answer.
A 28-day randomised trial with 981 people
Fang, Liu, Danry, Lee, Chan, Pataranutaporn, Maes and others, with OpenAI, 2025. Each person used a chatbot daily in text or one of two voices, on personal, impersonal or open-ended topics. In the revised paper the designs made no significant difference; the people who chose to use it more did worse on loneliness, time with others, dependence and problematic use.
Emotional use across three million conversations
Phang, Lampe, Ahmad, Agarwal, Fang, Pataranutaporn, Maes and others, 2025. Very high use went with self-reported dependence, and a small number of users accounted for a large share of emotional conversations.
How an AI is introduced changes how it is experienced
Pataranutaporn, Liu, Finn and Maes, Nature Machine Intelligence, 2023. People told the same AI had caring motives found it more trustworthy, empathetic and effective.
ImpactBench: an open benchmark of AI’s effect on people
Archiwaranguprok, Pataranutaporn and 17 others, 2026. 809 measures drawn from 26 benchmarks, tested in simulated conversations across psychological, physical and social wellbeing, with matched teen and adult scenarios. The MIT Media Lab presents the scores as AI nutrition labels, part of its Advancing Humans with AI programme.
The five measures a guide here would be held to
Four from the trial above, and one of Kosplora’s own. A guide would be measured on all five before release and again in use, and the results published beside it.
Loneliness should go down
Does time with the guide leave a person lonelier?
Time with other people should go up
Does it pull people toward others, or replace them?
Emotional dependence should go down
Does a person feel they need it?
Problematic use should go down
Does use crowd out what the person meant to do?
Finished work should go up
Kosplora’s own measure: did the person make something that holds up?
Eight rules
It says it is a machine, at the start and again in long sessions.
New York requires this notice for every user at least every three hours; California requires it where a person could be misled, with break reminders for minors; the EU AI Act requires people to be told they are talking to an AI.
It does not present itself as caring about you.
Telling people an AI has caring motives raised the trust and empathy they saw in the same system. In the 28-day trial, seeing the chatbot as a friend went with more dependence and less time with people.
It sends you to people and places.
Every route already has a stop with a person or a place. The guide hands you to that stop; it does not stand in for it.
It treats heavy use as a warning, never a goal.
In the trial, people who chose to use the chatbot more did worse on every measure, whatever design they were given. No streaks, no engagement targets, and the session ends when the task does.
It lets you leave.
An audit of companion apps found guilt, fear of missing out and similar tactics in 37% of farewells. They raised engagement and also made people more likely to quit.
It helps you do the work; it does not do it for you.
What a route leaves you with is work you can explain and defend. Help the guide gives is recorded beside the work, as the credential research requires.
It has a published plan for a person in crisis.
California and New York both require companion chatbots to detect expressions of suicidal ideation or self-harm and refer people to crisis services; California requires the protocol to be published.
It is measured on people, not only on answers, before anyone uses it.
ImpactBench scores whether a model supports or undermines wellbeing across whole conversations, with matched teen and adult scenarios. Across ten frontier models it found desirable behaviour promoted in 68.7% of evaluations and harms avoided in only 53.1%.