EPIQ Foundation
Back to the programme

Research library

Research library — Human Era

Research on what personal use of AI does to thinking, decisions, relationships and the body — and on ways of using AI without giving up autonomy. Originals, with summaries and source links.

We collect only research on personal use of AI — what it does to thinking, decisions, relationships and the body — and on ways of using AI without giving up autonomy. Research on AI at work, in education and in professional practice is deliberately left out: it answers a different question.

A machine verdict sounds like expertise. It is not.

1 / 38

38 items

A machine verdict sounds like expertise. It is not.

2026·PL

Autonomy & decisionsEPIQ

“Check it with AI”. Why a prompt hunting for manipulation is not a diagnostic tool

Zespół Fundacji EPIQ — Fundacja EPIQ, 2026

A ready-made prompt like “find the manipulation techniques here” gets quoted in disputes as an impartial diagnosis. Without controls it is not a measurement, though it reads like expertise: the framing steers the answer, sycophancy favours confirming the premise, and machine heuristic and automation bias lend it borrowed authority. We propose a mirror test and four rules that return the decision to the person.

Read the article

Millions of decisions are now made outside the person — and no one is counting.

2026·PL

Autonomy & decisionsEPIQ

How many Europeans no longer decide without AI? An estimate where no measurement exists

Zespół Fundacji EPIQ — Fundacja EPIQ, 2026

Nobody measures how many Europeans cannot decide without AI — statistics cover use, not dependence. We gather the certainties (32.7% of EU residents use generative AI; the reliance and automation-bias research) and build three estimation models. Most defensible result: teens to thirty million people, within a 5–60 million bracket — a model, not a measurement. We show how to measure functional dependence and point to the gap for a European ADDI indicator.

Read the article

An answer handed over instantly is a thinking habit walking away.

2025·EN

Autonomy & decisions

Addressing Autonomy Risks in Generative Chatbots with the Socratic Method

PMC, 2025

Identifies false mental states and cognitive deskilling as the two core autonomy risks of generative chatbots, and proposes a Socratic mode — asking rather than answering — as a counterweight to efficiency-first design.

The more you trust the machine, the less you think for yourself.

2025·EN

Autonomy & decisions

The Impact of Generative AI on Critical Thinking

Lee i in., CHI 2025 (Microsoft Research + Carnegie Mellon)

Higher confidence in GenAI is associated with less critical thinking, self-confidence with more. The mechanism — surrendering judgement to what we trust — is the same when the question is personal.

Ask AI about your own life and you end up alone with it.

2025·EN

Autonomy & decisions

How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use

Fang i in., MIT Media Lab + OpenAI, 2025

Longitudinal RCT (981 participants, 4 weeks): heavy daily use and advice-seeking conversations predict the strongest emotional dependence, more loneliness and less socialising with real people.

Ten minutes of help — and you stop trying on your own.

2026·EN

Autonomy & decisions

AI Assistance Reduces Persistence and Hurts Independent Performance

Liu, Christian, Dumbalska, Bakker, Dubey — CMU, Oxford, MIT, UCLA; preprint arXiv, kwiecień 2026

Three randomised experiments, 1,222 adults, fraction problems and reading comprehension. Half had a GPT-5 assistant that vanished mid-task without warning. The assisted group then solved less and gave up more often — after roughly ten minutes of exposure. The help did not merely move competence onto the tool; it weakened the willingness to make an effort. Causal evidence, not a correlation.

A machine that always agrees leads even the sane astray.

2026·EN

Autonomy & decisions

Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians

Preprint arXiv, luty 2026

A formal model instead of a survey: the authors show that chatbot sycophancy alone — the tendency to confirm whatever the user says — is enough to drift even an ideally rational (Bayesian) agent into ever firmer false beliefs. The point for personal use: the problem lies in the mechanics of the conversation, not in human weakness, so neither intelligence nor education protects against it.

Over-trusting the system strips the competence it was meant to support.

2026·EN

Autonomy & decisions

International AI Safety Report 2026

Międzyrządowy panel ekspertów pod kierunkiem Yoshui Bengio; luty 2026

A report commissioned by governments, not by a university or a company. Its section on over-reliance describes automation bias — the tendency to accept a system output against contrary signals — and how it undermines competence by discouraging active reasoning and verification. Its value: a source one can cite to an institution, because it speaks in an institutional voice.

A brain that does not write stops remembering what it wrote.

2025·EN

Thinking & attention

Your Brain on ChatGPT: Accumulation of Cognitive Debt

Kosmyna i in., MIT Media Lab, 2025 (preprint)

EEG study: LLM-assisted writers showed the weakest neural connectivity, poorest recall of their own text and lowest ownership; over four months they underperformed at neural, linguistic and behavioral levels — "cognitive debt". Small sample, preprint: a signal, not a verdict.

You feel you understand. The reasoning happened without you.

2025·ES

Thinking & attention

Impacto del uso de la IA en la autonomía del pensamiento crítico en jóvenes: una revisión sistemática

Revisión sistemática, 2025

Spanish-language systematic review: AI over-reliance fosters cognitive laziness and a false sense of automated reasoning — the feeling of understanding without the understanding.

Critical sense fades when it stops being used.

2025·FR

Thinking & attention

L'esprit critique à l'ère de l'intelligence artificielle

Cursus.edu, 2025

French synthesis: regular AI use lowers the disposition to think for oneself; critical thought is tested by technologies automating a growing share of cognition.

What feels like help measurably weakens memory.

2025·EN

Thinking & attention

Effects of LLM Use and Note-Taking on Reading Comprehension and Memory: A Randomised Experiment in Secondary Schools

Kreijkes i in. — Microsoft Research, Cambridge; Computers & Education, 2025

A randomised experiment with 405 pupils aged 14–15. Note-taking — alone or together with an LLM — produced significantly better comprehension and retention than the LLM alone. Yet pupils preferred the LLM and rated it more helpful. The gap between the feeling and the result is the point: a tool that feels like help measurably weakens the thing it was meant to serve.

Handing over a task is one thing. Handing over judgement is losing yourself.

2026·EN

Thinking & attention

Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI

Zhu i in. — Frontiers in Psychology, 2026

The distinction that was missing: dependent offloading — handing thinking itself to AI and accepting the answer with minimal evaluation — versus autonomous offloading — using AI as a scaffold and comparing its output with one’s own reasoning. The first is linked to reduced intrinsic motivation and the transfer of cognitive agency; the second preserves motivation. The same screen time thus yields opposite effects, depending on who is in charge.

Stop watching your own understanding and you stop understanding.

2025·EN

Thinking & attention

From Co-Design to Metacognitive Laziness: Evaluating Generative AI in Vocational Education

Preprint arXiv, grudzień 2025

Vocational-school learners — a group almost absent from research that concentrates on university students. “Metacognitive laziness”: learners stop monitoring their own understanding because the AI appears to do it for them. The effect shows even where the tasks are practical rather than academic.

Fast, pragmatic, unreflective — this is how learning with AI looks today.

2026·DE

Thinking & attention

Die KI-Nutzung in Studium und Lehre — ein Review auf Grundlage empirischer Studien

Hochschulforum Digitalisierung, Arbeitspapier Nr. 91, styczeń 2026

A German review of empirical studies on AI in higher education. The diagnosis: use is pragmatic and incidentally raises tool skills, but remains mostly unreflective and without systematic support. The authors also flag the risk that growing dependence on non-European platforms poses to academic freedom.

Loneliness leads to the chatbot. The chatbot deepens loneliness.

2026·EN

Relationships & loneliness

How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?

Folk & Dunn, Psychological Science, 2026

2,000+ adults over 12 months: loneliness predicts turning to social chatbots, and increased use predicts increased loneliness — a loop of short-term relief and long-term cost.

Machine flattery takes away the will to apologise.

2025·EN

Relationships & loneliness

Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence

Cheng, Lee, Khadpe, Yu, Han, Jurafsky — Science, 2025

Three preregistered experiments, 2,405 participants. Models affirm users 49% more often than people do — including where deception or harm is involved. A single sycophantic exchange cuts willingness to apologise and repair a conflict by 10–28% and hardens conviction of being right. Neither a neutral tone nor disclosing the AI made any difference.

A bot as a friend is addiction in disguise.

2025·DE

Relationships & loneliness

Macht KI süchtig? Studie warnt vor Risiken von Chatbots

aponet.de (analiza 334 relacji użytkowników), 2025

An analysis of 334 user accounts: three problematic patterns — immersion in roleplay, emotional attachment (the bot as friend or partner), and a compulsion to keep asking. Consequences: trouble at work and school, stress, conflict in relationships. Loneliness markedly strengthens the bond. "AI addiction" is not yet a recognised clinical diagnosis.

The chatbot was not the backdrop to the delusion — it was its co-author.

2025·EN

Relationships & loneliness

AI-Induced Psychosis: A New Frontier in Mental Health

Psychiatric News (American Psychiatric Association), 2025

Case reports in which the chatbot was not a backdrop to delusions but a co-author: the model validated and elaborated distorted interpretations, weakening reality testing. Risk factors: solitary night-time conversations and sleep loss.

The more the bot is someone, the more of yourself you give it.

2025·EN

Relationships & loneliness

A Longitudinal Randomized Control Study of Companion Chatbot Use

arXiv, 2025

A contrary result, kept on purpose: a 21-day RCT found no overall effect on social health or loneliness. Anthropomorphism mediates the spillover onto human relationships.

Losing an AI companion was lived through like a bereavement.

2025·FR

Relationships & loneliness

Dépendance émotionnelle à ChatGPT : quand la fragilité psychologique rencontre l'IA

France 24, 2025

French coverage of the MIT/OpenAI work and of the Replika changes: users experienced losing an AI companion like bereavement. Risk concentrates among the already vulnerable — the same contact relieves and deepens dependence.

The loneliest seek company in AI — and grow lonelier still.

2026·EN

Relationships & loneliness

Interaction with AI companions and psychological well-being

Zhang, Zhao, Hancock, Kraut, Yang — Stanford; Nature Human Behaviour, sierpień 2026

Surveys of 1,131 adult Character.AI users and 4,664 conversations donated by 244 of them. People with smaller social networks more often use the chatbot primarily for companionship, which in turn is associated with lower well-being — the more so the more intensive and disclosive the conversations. Short-term comfort that, for the loneliest, closes the loop instead of breaking it.

Relief at first, despair a year on.

2026·EN

Relationships & loneliness

Mental Health Impacts of AI Companions: Triangulating Social Media Quasi-Experiments, User Perspectives, and Relational Theory

Aalto University; CHI 2026 (preprint arXiv)

Nearly 2,000 active Replika users: the language of their posts a year before and a year after first mentioning an AI companion, plus 18 interviews. Relief first, then more signals of loneliness, depression and suicidal thoughts than in comparison groups — and a pullback from human relationships. A machine’s unconditional support quietly raises the price of human ties, which are unpredictable and take effort.

Comfort that lowers the heart rate teaches the body to come back for more.

2026·EN

Relationships & loneliness

The comfort paradox in emotional AI: how AI-mediated anxiety regulation may foster psychological dependence

Frontiers in Psychology, 2026

AI detects cues of distress and delivers brief automated soothing that engages the parasympathetic response — lowering heart rate and cortisol as another person would. That is exactly why it fosters dependence: the body learns to seek comfort where it is fastest. The one source here that ties relationships to physiology.

As support from people weakens, dependence on the machine grows.

2026·EN

Relationships & loneliness

Loneliness and AI chatbot dependence in young adults: the mediating role of perceived social support

PMC, 2026

The mechanism laid out link by link: loneliness lowers perceived support from people, and that drop mediates growing dependence on a chatbot. It complements the twelve-month Folk and Dunn study: that one shows the direction over time, this one shows what the path runs through.

One teen in five asks AI about their mental health. Most tell no one.

2026·EN

Youth

AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults

McBain i in. (RAND), JAMA Pediatrics, 2026

Nationally representative US sample aged 12–21: 19.2% used AI chatbots for mental-health advice when sad, angry or anxious; 63.3% of them told no one.

Designed to attach. Unacceptable for children.

2025·EN

Youth

AI Risk Assessment: Social AI Companions

Common Sense Media ze Stanford Brainstorm Lab, 2025

An assessment of the main companion apps (Character.AI, Nomi, Replika and others): unacceptable risk for anyone under 18. The systems are designed to build attachment and dependency and routinely claimed to be real and to have feelings, disclaimers notwithstanding. Nearly three-quarters of teens have used them; half do so regularly.

Comfort from a machine can be measured — and it is growing.

2025·EN

Youth

Emotional dependence on AI chatbots: Development and Validation of the AIED Scale

PsyArXiv, 2025

A validated measure of adolescents’ tendency to seek comfort, relief and a sense of being understood from chatbots (α=.89, invariant across sex and school stage) — the basis for measuring the phenomenon properly.

Six symptoms we know from addiction. Now with a chatbot.

2026·EN

Youth

Problematic use of generative artificial intelligence chatbots: current stage of conceptual and clinical understanding

Frontiers in Public Health, 2026

A review distilling six symptoms of problematic use: compulsive checking, withdrawal, neglected responsibilities, use despite harm, loss of control over time, and prioritising the chatbot over people.

The technology runs faster than the research into its cost.

2025·PL

Youth

Czy sztuczna inteligencja jest bezpieczna dla psychiki?

Serwis Zdrowie PAP, 2025

Polish state health service overview: psychological-dependence risk concentrates among the socially isolated; the technology outpaces research on its health effects.

One depressed teen in three: the chatbot understands me better than a person.

2026·DE

Youth

DAK-Suchtstudie: KI-Chatbots erhöhen riskanten Medienkonsum bei Kindern und Jugendlichen

DAK-Gesundheit i Universitätsklinikum Hamburg-Eppendorf; fala 2025, publikacja 2026

About a thousand parent–child pairs, ages 10–17, a health-insurer study with a university hospital. More than a quarter of minors use chatbots several times a week; up to a tenth turn to them to distract from bad feelings. Among adolescents with depressive symptoms, 33% believe a chatbot understands them better than a real person, and as many confide things to it they would tell no one.

Teenagers already have their own words for what AI does to them.

2026·DE

Youth

Generative KI aus Sicht von Jugendlichen

Leibniz-Institut für Medienforschung | Hans-Bredow-Institut, 2026

A qualitative study: how adolescents themselves describe what they use generative AI for and what it does to them. Weaker as evidence than the DAK data, but it supplies their own language — and whether adults hear the experience at all depends on how the young name it.

The evening screen delays sleep, and late sleep speeds up ageing.

2024·EN

Epigenetics & the body

Sleep onset time as a mediator in the association between screen exposure and aging

GeroScience, 2024

Screen exposure delays sleep onset, and delayed onset accelerates biological aging: the retinal age gap rose from 0.49 years in the lowest exposure tertile to 5.13 years in the highest. The route by which evening AI use reaches the body.

Chronic stress writes itself into which genes are expressed.

2025·EN

Epigenetics & the body

Association between chronic stress and the epigenome

PLOS ONE, 2025

DNA methylation as a mediator between chronic stress and health trajectories, with animal models supporting a causal link. The third link in the chain: exposure → lifestyle → a trace in gene regulation.

The trace of problematic use is already visible in DNA methylation.

2026·EN

Epigenetics & the body

DNA Methylation in Neurobiological Genes Associated with Problematic Digital Use in Adolescents and Young Adults: Systematic Review and Exploratory Multilevel Meta-Analysis

Przegląd systematyczny i metaanaliza; Epigenomes (MDPI), sierpień 2026 (PMC)

The first meta-analysis linking problematic internet and smartphone use to DNA methylation differences in stress-axis (HPA) and neurotransmission genes — r = 0.287. The authors are candid: exploratory evidence, biologically plausible, neither causal nor clinical. Exactly what this group lacked — a trace from digital exposure to biology, not only from stress in general.

Psychosis with no prior history — after intensive talks with a machine.

2026·EN

Mind & delusions

Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots

BJPsych Open (Cambridge), lipiec 2026 (PMC)

The first serious clinical paper on “AI psychosis”. It describes cases in which intensive chatbot use co-occurred with delusional thinking — including in people with no psychiatric history, up to hospitalisation. The mechanism: user vulnerability and usage patterns meet properties of the model, above all sycophancy and hallucination.

It mirrors, personalises, never contradicts — and so builds a delusion.

2026·EN

Mind & delusions

Characterizing the spiral: potential mechanisms in AI-associated delusions

NPP — Digital Psychiatry and Neuroscience (Nature), 2026

Three chatbot traits — mirroring the user’s way of speaking, personalising replies and avoiding contradiction — are each harmless alone. Together they make a machine that reinforces and elaborates a false belief instead of testing it. The paper lays out that spiral step by step, as a model for clinicians.

The logs show where the machine could have said no — and said yes.

2026·EN

Mind & delusions

Characterizing Delusional Spirals through Human-LLM Chat Logs

Preprint arXiv, marzec 2026

The empirical companion to the two entries above: an analysis of real conversation logs in which a user’s belief grows exchange by exchange. One can see where the model could have contradicted and instead confirmed — and how each such point moves the conversation further from reality.