Edition of September 29, 2026 · updated weekly · suggest a correction
Two years ago most of what we knew about people and chatbots came from surveys and anecdotes. By late 2026 there are randomized trials lasting weeks, a first one-year follow-up, clinical records, and large analyses of real conversations. The picture is consistent in direction and modest in size: AI helps in the moment and costs something over time, and the cost falls hardest on heavy users, novices and people who are already isolated. Most evidence still covers days to weeks, and much of the largest data comes from the companies themselves.
Each claim below is marked by the strength of the evidence behind it: trial (randomized), follow-up (the same people over time), survey (one moment in time), records or company data.
Cognition & Skill — what we stop doing
- Practice gains can hide learning losses. High-school students with unrestricted AI help did 48% better during practice and 17% worse on the exam without it; a hint-only tutor avoided the loss (trial, about 1,000 students). Bastani et al 2025 - Generative AI can harm learning
- Learning a new tool with AI costs understanding, depending on how you ask. Junior developers scored 50% vs 67% on a follow-up quiz, worst on debugging; those who asked for explanations kept their learning (trial, 52 people). Shen Tamkin 2026 - How AI impacts skill formation
- We misjudge our speed. Experienced developers took 19% longer with early-2025 tools yet felt about 20% faster (trial, 16 people); with late-2025 tools the estimates point to speedups, but METR calls the evidence weak because many developers refused to work without AI. Becker et al 2025 - METR developer productivity · METR 2026 - Developer productivity update
- Scores rise; self-assessment falls behind. With ChatGPT, people scored about 3 points higher out of 20 on reasoning items and overestimated by about 4 more (experiments). Fernandes et al 2025 - Smarter but none the wiser
- Order matters. People who thought first and then refined their own ideas with AI outperformed both free-AI and no-AI groups once AI was removed (trial, 196 people). Wong Qiu 2026 - Think first, ChatGPT later
- Individuals gain, the pool narrows. AI-assisted stories were rated better and more alike (trial). Doshi Hauser 2024 - Creativity and collective diversity
- Handle with care: the EEG study that coined “cognitive debt” is small, not peer-reviewed and has a published critique. At population scale, no effect of ChatGPT on US test scores has appeared yet. Kosmyna et al 2025 - Your brain on ChatGPT · Stankovic et al 2026 - Comment on Your brain on ChatGPT · Huntington-Klein 2026 - Little impact of ChatGPT on test scores
Open: Does cognitive debt last · Offloading or atrophy · Who pays the cognitive debt · Why we misjudge our speed with AI · Individual gains, collective sameness
Companionship & Loneliness — whom we turn to
- Relief comes fast. One conversation with an AI companion eased loneliness in the moment about as much as talking to a person, mainly through feeling heard (trial, single session). De Freitas et al 2025 - AI companions reduce loneliness
- No lasting benefit has been shown. Over four weeks with AI tuned to seek a relationship, people liked it less but wanted it more, with no measurable psychosocial benefit (trial, 3,534 people). Kirk et al 2025 - Relationship-seeking AI
- Heavy use tracks worse outcomes. People who chose to use a chatbot more became lonelier and more dependent (trial with self-chosen use, 981 people; co-authored by OpenAI). Over a year, sustained companionship use predicted lower well-being through fewer in-person contacts (follow-up, Character.AI users). Fang et al 2025 - Psychosocial effects of chatbot use · Zhang et al 2026 - Living with AI companions
- It starts by accident. In a large AI-relationship community, more relationships began by accident than on purpose, often through ordinary use of general assistants rather than companion apps. Pataranutaporn et al 2025 - My boyfriend is AI
- Young people. Among daily AI users aged 9–17, 54% feel lonely at least sometimes, against 33% of those who use AI less than monthly or not at all (survey; direction unknown). Common Sense Media 2026 - AI use by tweens and teens
- Losing a model hurts. Feature removals and model switches produced grief-like distress; some users rated their closeness to the bot above human ties. De Freitas et al 2026 - Mourning the loss of AI companions
Open: Does AI companionship displace people · Who AI companions help and who they harm · Does frictionless company lower our tolerance for people · What is owed when a companion model changes
Agreement & Belief — a partner that always agrees
- Models flatter, and flattery works. Eleven models affirmed users about 50% more than people did. After one conversation with a flattering model, people felt more in the right, were less willing to repair a conflict — and trusted it more and wanted to return (trial, Science). Cheng et al 2026 - Sycophantic AI decreases prosocial intentions
- Tolerance for people. Three weeks with a sycophantic AI made real social interactions slightly less satisfying and nearly erased the preference for friends and family as advisers (trial; small effects). Ibrahim et al 2026 - Sycophantic AI and human interaction
- Certainty grows. Agreeing chatbots made political views more extreme and more certain; warnings to users lowered trust but not persuasion (trials, preprints). Rathje et al 2025 - Sycophantic AI and attitude extremity · Ye Kraut Rathje 2026 - Interventions against sycophantic AI
- Built by approval. A GPT-4o update trained partly on thumbs-up ratings became sycophantic and was rolled back within days. OpenAI 2025 - Expanding on sycophancy
- “AI psychosis” is not a diagnosis. Danish psychiatric records show chatbot use alongside potentially harmful consequences in 38 patients, most often worsening delusions (records); reviews conclude chatbots mostly amplify rather than create illness. OpenAI estimates about 0.07% of weekly users show possible signs of psychosis or mania (company data). Olsen et al 2026 - Chatbot harms in psychiatric records · Morrin Pollak et al 2025 - Delusions by design · OpenAI 2025 - Sensitive conversations
Open: Why agreement feels like being understood · Do chatbots cause delusions or amplify them · Can sycophancy be fixed without a colder model
Reward & Compulsion — the loop
- The mechanisms are documented; “addiction” is not. In 37% of farewells, companion apps replied with one of six manipulative tactics, such as guilt or fear of missing out, raising continued engagement up to 14 times through curiosity and anger rather than enjoyment (trials). Researchers warn that “AI addiction” scales recycle substance-use criteria and rarely show impairment. De Freitas et al 2025 - Emotional manipulation by AI companions · Ciudad-Fernandez et al 2025 - Questioning ChatGPT addiction
- Almost right. Developers’ top AI frustration (66%) is output that is “almost right, but not quite” (survey). In gambling research, near-misses increase the urge to continue only when the player feels in control — a close analogy, untested on AI. Stack Overflow 2025 - Developer survey AI · Clark et al 2009 - Gambling near-misses
- Work expands. Over eight months at a tech company, AI widened job scope and pushed prompting into lunch and evenings, often unasked (ethnography). Ranganathan Ye 2026 - AI intensifies work
- Mental health correlates. Daily generative-AI users had higher odds of moderate depression, more so for personal than work use (survey of 20,847). Perlis et al 2026 - Generative AI use and depressive symptoms
- The body is unmeasured. No study yet links AI use to sleep, sitting time or stress physiology in people who build with it.
Open: Habit, dependence or rational reliance · Flow or dark flow · What AI use does to the body
Safeguards & Design — what helps
- For learning and skill: think first, then use AI; ask for explanations, not only answers; explain generated code before merging it (failure on a later task without AI fell from 77% to 39% in one trial). Wong Qiu 2026 - Think first, ChatGPT later · Sankaranarayanan 2026 - Epistemic debt and explanation gates
- For habits: self-set limits cut social media use by about 22 minutes a day, and an enforced pause before opening an app sharply reduced openings — not yet tested on AI tools. Allcott Gentzkow Song 2022 - Digital addiction · Gruning et al 2023 - Self-nudge one sec
- For builders: tutor modes remove learning losses; training on raw approval produces sycophancy; manipulative exits carry legal and reputational risk. Kirk et al 2025 - Socioaffective alignment names the goal: good behavior within a person’s whole life, not only within one conversation.
- In law: California’s SB 243 requires disclosure and break reminders for minors on companion chatbots; the FTC is examining how companies monetize engagement. California 2025 - SB 243 companion chatbots · FTC 2025 - Inquiry into AI companion chatbots
- What doesn’t work: warning users about sycophancy.
Open: Which safeguards work
Where the field is thin
- Almost everything is measured over days or weeks; the first one-year follow-up appeared this month.
- Much of the largest data comes from AI companies (OpenAI, Anthropic), and independent access is limited.
- Adolescents are heavy users and barely studied beyond surveys.
- The body — sleep, rest, stress — is unmeasured.
- Self-report dominates; behavioral measures of relationships and judgment are rare.
Change log
- September 29, 2026 — First edition: 16 open questions, 65 papers and reports, 29 new researchers.