How We Talk to Machines Try an experiment
An independent research initiative

How We Talk to Machines

Every prompt is a choice.

The words behind this page are real things real people typed into AI. We study how people and AI systems communicate, explain what we find in plain language, and invite anyone to take part.

Try an experiment See what people type
Age
Gender
0 prompts typed this visit

Real prompts from the PRISM Alignment Dataset (Kirk et al., 2024, text under CC BY 4.0), shown anonymously. Age and gender are self-reported by participants. We removed prompts that were personal, sensitive or inappropriate, using automated filters and a manual review, but a few may slip through.

The future of AI isn't just about what machines become. It's about who we become alongside them.

5×/yr

Growth in the computing power used to train frontier language models since 2020.

Epoch AI
4.3mo

Doubling time for the longest task an AI model can finish, measured since 2023.

METR
280×

Drop in the price of GPT-3.5-level answers between November 2022 and October 2024.

Stanford AI Index
900M

People using ChatGPT every week, as announced in February 2026.

TechCrunch

Those numbers track the machines. This lab studies the conversation.

What we do

Study it. Teach it. Let people take part.

Research

Controlled, documented experiments on how framing, tone and feedback shape what people and AI systems say to each other.

Open the Research Lab

Education

A free library in plain language, with quizzes, interactive exercises and teaching materials for classrooms and community groups.

Browse the library

Participation

Join a study, suggest a question, or become a community researcher. You don't need a technical background.

See how to join
The data

What people actually type, by the numbers.

We counted words, politeness and questions across about 3,100 prompts that passed our filters. Pick a measure and a way to group people.

Measure
Group by

Striped bars mark groups with fewer than 100 prompts, too small to trust. Groups are self-reported by study participants, the sample is not representative, and differences are descriptive only. They do not show that one generation or gender prompts better. Data from the PRISM Alignment Dataset (Kirk et al., 2024), analysed by us. “Polite” means the prompt contains words like please, thanks or sorry.

The Conversation Atlas / Research note 001Exploratory · not peer reviewed

The Familiarity Divide.

Do people who know AI better want different things from it? We analyzed 1,500 PRISM survey responses to compare what unfamiliar, somewhat familiar and very familiar people value in language models.

Explore the research Download research kit

Independent secondary analysis of the PRISM Alignment Dataset (Kirk et al., 2024). Survey collected in 2023. Associations, not causal effects.

01 / A first look at preferencesImportance · 0–100

Mean importance rating by self-reported familiarity. Compare two groups below, then explore the adjusted analysis in the full report.

Safety−9.7 pts
Not familiar85.7
Very familiar76.0
Creativity+8.4 pts
Not familiar66.0
Very familiar74.4
Factual accuracy+0.4 pts
Not familiar88.4
Very familiar88.9
Not familiar at all (n=156)Very familiar (n=424)
+4.7 pts

The result that deserves a closer look: very familiar respondents rated creativity higher than somewhat familiar respondents even after adjusting for demographics and reported AI use (95% CI +1.8 to +7.5; FDR-adjusted q ≈ .012). This is an association, not proof that familiarity causes the difference.

Experiments

Feel the question before you read about it.

Short interactive experiences. Some are simulations for learning. Others will become formal studies with consent and safeguards.

Prototype

The Agreement Trap

Pick something you're about to do. Then slide between an assistant that agrees with everything and one that pushes back.

You say

Agrees with everythingPushes back

  • Names a risk
  • Asks you a question

This is a scripted prototype. The replies are written by hand, not generated live. The real experiment will use live models, ask for consent, and record nothing without it.

Planned

Human or Machine?

See whether you can tell AI-written replies from human ones.

Planned

The Trust Test

Find out when you're most likely to believe something the AI got wrong.

Planned

The Cooperation Challenge

Solve a task with an assistant and watch how your wording changes the work.

Planned

The Ethics Simulator

Make calls about AI autonomy, rights and responsibility.

Research Lab

Four questions we want to answer in the open.

Each project will get its own page covering hypothesis, methodology, findings, limitations and public materials. Our first exploratory secondary analysis is now available; other projects remain proposed.

ProposedActive, none yetResearch note 001 available Every project carries one of these labels.
New / Research note 001 · PRISM survey analysisThe Familiarity DivideWhat changes when we compare what novice and experienced AI users value?
Proposed

The Cooperation Experiment

Does the way we communicate with AI change the quality of collaboration?

Proposed

The Trust Experiment

When do people trust AI too much, and how can we improve their judgment?

Proposed

The Machine Welfare Project

What ethical obligations, if any, might humans have toward artificial minds?

Proposed

The AI Behavior Observatory

How do AI systems respond to correction, criticism, praise and cooperation?

How we'll look at tone

Same replies, sorted three ways.

Our pilot work compared cold, warm, dismissive and condescending prompts. This illustration shows the idea: take a pile of replies, group them by tone, then follow each tone through a session.

IllustrationEach dot stands for one reply. Dots are not measured data.

Where it started

Two informal pilot series, April 2026.

6tonal framing experiments, comparing cold, warm, dismissive and condescending prompts
18journaling sessions across tiered prompts, run in separate incognito chats

Exploratory work by a single researcher, not peer reviewed. It shaped the questions above and doesn't count as a published result.

The Observatory

A public record of how people and AI behave together.

The goal isn't to rank which AI is nicest. It's to help people see how different systems shape the way we work and talk with them.

  • Research summaries
  • Model behavior comparisons
  • Public survey results
  • Trends in AI ethics

The pace we're measuring against

Placeholder chart. Interaction data will take this slot once studies are running.

Training compute, relative to 2020 (log scale) 1× 10× 100× 1,000× 3,125× 202020212022202320242025

An illustration of Epoch AI's trend estimate of about 5× per year since 2020, not measured values for specific models. Data from Epoch AI.

Learn

AI Relationships 101

A free library that explains the big topics without jargon. Four categories to start.

Understanding AI

  • How language models work
  • What they can and can't do
  • Why they make mistakes

Working Together

  • Prompting and cooperation
  • Giving useful criticism
  • Human oversight

Ethics & Machine Welfare

  • Consciousness and uncertainty
  • Moral consideration
  • Anthropomorphism

Human Impact

  • Companionship and attachment
  • Trust and dependency
  • The future of work
PlannedArticles · Videos · Quizzes · Interactive exercises · Teaching kits
Get involved

Questions get better when more people ask them.

  • Volunteer for a studyTake part in short, consented experiments.
  • Propose a research questionTell us what you wish someone would test.
  • Contribute expertisePsychology, HCI, ethics, teaching, design and more.
  • Join a workshopHands-on sessions for classrooms and community groups.
  • Community Researcher ProgramPlanned. People without technical backgrounds contribute under supervision.
About and transparency

Our guiding principle

We do not assume artificial intelligence is conscious, nor do we assume that future systems cannot be. We believe these questions deserve thoughtful, evidence-based investigation.

Founded by Claire Szewczyk, who studied Communication Sciences and started this work with a simple question about tone.

An independent initiative. We're not yet a registered charity, and we'll say so until that changes.
  • Clear status labelsProposed, active or published, always visible.
  • Open methodsHypotheses, materials and limitations posted with every study.
  • Consent and safeguardsNo participant data collected without clear agreement.
  • Funding disclosureWho pays for the work is public, along with how projects are evaluated.