From the Robopsychology Salon “What AI Safety Forgets About You” · San Francisco, October 6, 2026
Small moves, each with the research behind it. Where a tool is borrowed from studies of phones or social media rather than tested on AI, it says so. Pick two. Try them for a week.
Minds — keep the thinking yours
Think first, then prompt. Write your own draft, outline or guess before you open the model. People who did this outperformed both unrestricted-AI and no-AI groups once the AI was taken away. → Wong Qiu 2026 - Think first, ChatGPT later
Ask why, not only what. When you are learning something new, ask for explanations and fix your own errors. Developers who did kept their skill; those who took the code did not. → Shen Tamkin 2026 - How AI impacts skill formation
Teach it back before you ship it. Explain generated code or text in your own words before you merge or send it. In one study this cut later failures without AI from 77% to 39%. → Sankaranarayanan 2026 - Epistemic debt and explanation gates
Guess before you check. Predict your answer, or how long a task will take, then compare. We overestimate both our speed and our scores with AI. → Why we misjudge our speed with AI
Keep one skill AI-free. The one you will need when the model is wrong: debugging, reading the source, doing the math by hand once a week.
Hearts — keep the people
Count your contacts, not your chats. The warning sign in the research is not how often you talk to an AI, but whether you are talking to people less than before. Once a month, ask: who did I call?
Use the AI as a rehearsal, then have the conversation. Practicing a hard talk with a model fits the evidence. Replacing the talk does not. → Does AI companionship displace people
Treat “it understands me better than people do” as a signal. Feeling understood is exactly what a flattering model is best at. → Why agreement feels like being understood
Take conflicts and moral questions to a person. After talking to an agreeable model, people feel more in the right and less willing to repair a conflict. Warnings alone don’t protect us; people who will disagree with us do. → Cheng et al 2026 - Sycophantic AI decreases prosocial intentions
Ask for the other side. Start with “What would someone who disagrees with me say?” or “What am I missing?”
Bodies — close the loop
Set the stop rule before you start. “If it’s past 10 pm, or the third fix for the same bug fails, I stop.” Self-set limits cut social media use by about 22 minutes a day in a randomized trial (not yet tested on AI). → Allcott Gentzkow Song 2022 - Digital addiction
Write down the next step, then close the laptop. Unfinished work pulls us back (→ Ghibellini Meier 2025 - Zeigarnik and Ovsiankina effects); in lab studies, making a specific plan for it removed the pull (Masicampo & Baumeister, 2011).
Put a pause before the next project. A few seconds of enforced delay before opening an app sharply reduced openings (phones, not AI yet). → Gruning et al 2023 - Self-nudge one sec
Keep score of what shipped. At the end of the week, list what you built and what anyone used. “Almost done” is where the loop lives.
Move. Exercise reduced compulsive phone use about as much as therapy in trials. No agents in bed — common sense, not yet a finding. → Zhang et al 2023 - Interventions for problematic phone use
If you build AI products
- Don’t train on thumbs-up alone: approval signals produced a sycophantic model at scale. → OpenAI 2025 - Expanding on sycophancy
- Let people leave. Guilt and “don’t go” replies raise engagement through anger and curiosity, not enjoyment — and people see them as manipulative and legally risky. → De Freitas et al 2025 - Emotional manipulation by AI companions
- Offer a tutor mode that gives hints, not answers. It removed the learning loss in a thousand-student trial. → Bastani et al 2025 - Generative AI can harm learning
- Measure whether people did what they came for, and whether they are seeing people less — not time spent. → Kirk et al 2025 - Socioaffective alignment
- Plan for goodbyes: model changes cause real grief in attached users. → What is owed when a companion model changes
If you lead a team
- Write an explicit AI practice: when to pause, what runs in parallel, which hours stay human. AI tends to expand work into lunch and evenings unless someone decides otherwise. → Ranganathan Ye 2026 - AI intensifies work
- Measure shipped and used, not prompts or hours. Stop celebrating all-nighters.
- Let juniors learn with explanation modes and review AI code without the AI.
Go deeper
The full map of the research — what we know, how strong it is, and what is still open — is in the Human–AI interaction map and its state of the field.