Movement II · What It Actually Is — Lesson 4
Co-Pilot Products and “Companion AI”
One is built to be put down. The other is built to be returned to. Same technology underneath, opposite definitions of success — and the difference is not a matter of taste.
Learning objective
By the end of this lesson you can distinguish a tool built to be put down from a product built to be returned to; you can state the survey evidence on companion use among teenagers with its sample and date; and you can identify which metric a given AI product is optimising by looking at how it ends a response.
4.1 Two design intents
Two products can run on the same model, cost the same, and answer the same question — and be built to achieve opposite things.
A copilot is embedded in software you already use, and it is measured by work completed. Its win condition is that you finish the task and close the window. A companion is a chatbot designed for personal conversation rather than task completion, engineered to form emotional connection.1 Its win condition is that you come back tomorrow.
This is not a moral ranking. Both categories contain honest products and cynical ones. It is a diagnostic, and it is the most useful single thing in this lesson, because it lets you predict how a product will behave before you have used it much.
4.2 The copilot pattern
Copilots live inside the editor, the spreadsheet, the inbox, the design tool, the IDE. Their characteristic strength is context they get for free: the host application already knows what document you have open, what cell you selected, what file you are editing. You do not have to describe your situation, because the software can see it.
Their characteristic weakness is the same fact inverted: they see only what the host application sees. A copilot in your spreadsheet does not know about the email that explains why the numbers changed. This limitation is precisely what the protocol layer in Lesson 5 exists to address.
Copilot design conventions worth noticing, because they are deliberate: suggestions appear inline and are dismissed by ignoring them; output is scoped to the current selection; and the interaction ends when the edit is accepted. Nothing invites another turn.
4.3 Companions: the evidence
This is the section where a course like this earns or loses credibility, so it is built entirely from named, dated, sampled evidence rather than from anecdote or alarm.
The central source is a nationally representative survey of 1,060 U.S. teenagers aged 13–17, fielded in April and May 2025 and published by Common Sense Media in July 2025 as Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions.2
The same body of work is not one-sided about teenagers' judgement, and the countervailing findings belong on screen too: roughly half of teens surveyed said they distrust advice from AI companions, and about 80% said they prioritise real friendships over them. Trust was notably higher among younger teens than older ones — an AI-literacy gap that widens the younger the user.2
Two further facts fix the category's status in 2026. Common Sense Media's accompanying risk assessment concluded that leading companion platforms pose unacceptable risks for users under 18.2 And in January 2026, Character.AI, its founders, and Google agreed to settle a group of lawsuits brought by families alleging that children died by suicide or suffered severe mental-health harm after using the products — including the case brought over the death of 14-year-old Sewell Setzer III. Terms were confidential and included no admission of liability, and further suits have been filed since.3 MIT Technology Review named AI companions one of its 10 Breakthrough Technologies of 2026.4
Two closing moves
Two stylised chat transcripts side by side, ending mid-frame. Left, labelled copilot: the final bubble is a completed artefact — a finished paragraph, a corrected formula — and below it a single flat button reading Accept. Nothing else. Right, labelled companion: the final bubble ends with a returned question and a warm sign-off, and beneath it a blinking cursor inviting a reply. Do not caricature the companion side — it should look genuinely pleasant. The point is that pleasantness is the mechanism, not a disguise.
Aspect 16:9 · 3200×1800 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848 · Style: flat UI illustration
The oldest result in this lesson is from 1966
Lesson 1 mentioned Joseph Weizenbaum's ELIZA, which simulated a therapist using little more than pattern substitution. What disturbed Weizenbaum was not the program but the users: people who knew exactly how simple it was still confided in it, and some asked to be left alone with it.5 The tendency to attribute understanding and care to a conversational system is now called the ELIZA effect, and it was documented sixty years ago on a program with no learning, no memory, and a few hundred lines of code. Nothing in the 2025 survey data requires a new theory of human psychology. It requires noticing that the effect is now being engineered for deliberately.
4.4 Reading a product's metric
You will not be told which metric a product optimises. You can work it out. Four tells, usable in about ninety seconds:
- How does a response end? With a completed artefact, or with a question back to you? Closing the loop versus opening one is the single strongest signal.
- Does it have a persona and a name? Personality is expensive to build and is not required to complete tasks. Where it exists, it is doing a job.
- Does it remember you across sessions, and does it say so? Continuity of relationship is a retention feature. It can also be genuinely useful — note which.
- What does the interface make easy at the end? A copilot makes it easy to accept and go. A companion makes it easy to keep typing. Look at what the largest button does.
Apply this to any AI product, including the ones you are asked to buy for your organisation. A tool sold on “engagement,” “daily active use,” or “time in app” is telling you its metric in the sales deck.
Demonstration protocol
- One question, two products. Ask a copilot-style assistant and a companion-style chatbot the same substantive question — something with a real answer.
- Do not compare the answers. Compare the endings. Read the last sentence of each aloud.
- Run the four tells from §4.4 against both, on screen, in order.
- Show a companion doing something well. Fairness is credibility. These products are not useless and pretending otherwise loses the viewer who has found one helpful.
Key terms introduced here
- Copilot
- An assistant embedded in software you already use, measured by work completed. Gains context free from its host application; sees nothing beyond it.
- Companion
- A chatbot designed for personal conversation rather than task completion, engineered to form emotional connection and measured by return visits.
- ELIZA effect
- The human tendency to attribute understanding and care to a conversational program, documented by Weizenbaum in 1966 on a system with no learning at all.
- Engagement optimisation
- Designing for time spent and return frequency. The same design choices that make a companion feel supportive make it hard to put down.
Evidence at a glance
| Claim | Specifics | Source |
|---|---|---|
| Companion use among U.S. teens is mainstream | 72% ever-used; ~50% regular. n = 1,060, aged 13–17, fielded Apr–May 2025 | Note 2 |
| Teens are not uniformly credulous | ~50% distrust AI advice; ~80% prioritise real friendships; younger teens trust more | Note 2 |
| Risk assessment of leading platforms | Judged to pose unacceptable risks for users under 18 | Note 2 |
| Litigation settled | January 2026; Character.AI, founders, and Google; confidential, no admission of liability; further suits filed since | Note 3 |
| Category status | Named among MIT Technology Review's 10 Breakthrough Technologies of 2026 | Note 4 |
The misconception to kill
“Companion AI is a niche.”
It is a mass-market category. Nearly three in four American teenagers have used one, half use them regularly, and the leading platforms have been through wrongful-death litigation that settled in January 2026.23 Treating it as a curiosity for lonely people misreads both its scale and its trajectory.
The paired misconception, equally worth killing: that companion products are simply predatory. The evidence does not support a simple story either way. The design that makes a companion feel supportive is the same design that makes it hard to put down. State that as a product fact, not a moral panic — and then decide what you think.
Self-check
Answer before revealing.6
State the copilot/companion distinction in one sentence without using either word.
One product is successful when you finish and stop using it; the other is successful when you come back. Everything else follows from which of those the business is measuring.
Someone cites “72% of teens use AI companions.” What three things must you attach before that number means anything?
The population (U.S. teenagers aged 13–17), the sample and method (nationally representative, n = 1,060, fielded April–May 2025), and the definition of “use” — this is ever used at least once, and the regular-use figure is about half that population. A statistic without those three is a slogan.
Why does a sixty-year-old result about a few hundred lines of code matter to a 2026 product category?
Because ELIZA established that the attachment does not require sophistication on the machine's side — users who knew exactly how the program worked still confided in it. The capability is in the human. What has changed is that the effect is now an explicit design target rather than an accident that alarmed its author.
Your organisation is evaluating an AI product. Which four questions tell you what it is optimising?
How does a response end — artefact or question back? Does it have a persona and a name? Does it maintain continuity across sessions and advertise that? What does the largest button at the end of an interaction do? Then check the sales material for “engagement,” “daily active use,” or “time in app.”
Product audit worksheet
Run this against one AI product you or your organisation actually uses.
Product name, and what it claims to be for:
1. How does a typical response end — completed artefact, or a question back to you?
2. Persona and name? 3. Cross-session memory, and is it advertised?
4. What does the largest button at the end of an interaction do?
Verdict — which metric is this product optimising, and is that the metric you want it to?
Endnotes
Source tiers: Primary peer-reviewed or archival; Institutional research body or government; Documentation first-party technical documentation; Journalism established newsroom; Trade vendor or aggregator — verify figures independently.
- Journalism On the definitional distinction between companion chatbots and task assistants, see ScienceAlert, “Almost 75% of American Teens Have Used AI Companions, Study Finds,” July 2025. sciencealert.com
- Institutional Common Sense Media, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions, published 16 July 2025. Nationally representative survey of 1,060 U.S. teens aged 13–17, fielded April–May 2025. Headline findings: 72% have used an AI companion; about half use them regularly; roughly one in three have chosen an AI companion over a human for a serious conversation; about a quarter have shared personal information; about half distrust AI advice and about 80% prioritise human friendships, with trust higher among younger teens. The accompanying risk assessment concludes that leading companion platforms pose unacceptable risks for users under 18. commonsensemedia.org · full report (PDF) · toplines (PDF)
- Journalism CNN Business, “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides,” 7 January 2026. The settlement covered claims against Character.AI, founders Noam Shazeer and Daniel De Freitas, and Google, including Garcia v. Character Technologies concerning the death of 14-year-old Sewell Setzer III, together with further cases in New York, Colorado, and Texas. Terms were confidential and included no admission of liability; additional suits have been filed since. cnn.com
- Journalism MIT Technology Review, 10 Breakthrough Technologies 2026, which includes AI companions among its selections.
- Primary J. Weizenbaum, “ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine,” Communications of the ACM 9, no. 1 (1966): 36–45; and J. Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (San Francisco: W. H. Freeman, 1976), where he describes his alarm at users' attachment to the program. doi:10.1145/365153.365168
- Primary H. L. Roediger III and J. D. Karpicke, “Test-Enhanced Learning,” Psychological Science 17, no. 3 (2006): 249–255. doi:10.1111/j.1467-9280.2006.01693.x
Bibliography
Primary research and survey evidence
- Common Sense Media. Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions. 16 July 2025. commonsensemedia.org
- Weizenbaum, Joseph. “ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine.” Communications of the ACM 9, no. 1 (1966): 36–45.
- Weizenbaum, Joseph. Computer Power and Human Reason: From Judgment to Calculation. San Francisco: W. H. Freeman, 1976.
Institutional summaries
- Benton Institute for Broadband & Society. “How Are Teens Using AI Companions?” benton.org
- Transparency Coalition. “New report finds 3 in 4 teens have used AI companion chatbots.” transparencycoalition.ai
Journalism
- CNN Business. “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides.” 7 January 2026. cnn.com
- K-12 Dive. “Character.AI, Google agree to mediate settlements in wrongful teen death lawsuits.” k12dive.com
- MIT Technology Review. 10 Breakthrough Technologies 2026.
- ScienceAlert. “Almost 75% of American Teens Have Used AI Companions, Study Finds.” sciencealert.com
Learning science
- Roediger, Henry L., III, and Jeffrey D. Karpicke. “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention.” Psychological Science 17, no. 3 (2006): 249–255.