Time-Spent Research for Chatbots

In February 2026, 44% of American adults said they use ChatGPT, and about a quarter used some chatbot every day. What that does to learning, relationships, and judgment over years is largely unknown, because the records sit inside the companies. Congress should shield independent researchers, let users donate their own conversations to science, require monthly usage statistics, and fund the long studies that can connect the two.

The Problem

In a February 2026 Pew survey, 44% of U.S. adults said they use ChatGPT, and about a quarter used some chatbot daily, including 4% who said "almost constantly."¹ In a 2025 survey, 72% of teens had used an AI companion, and a third of teen users had chosen to discuss serious matters with one instead of a person.²

These tools teach, draft, and advise, often well. The benefits are immediate, but the costs appear over months or years, often silently. In one four-week study of 981 people by MIT and OpenAI researchers, participants who chose to use the chatbot more reported more loneliness, more emotional dependence, and less socializing.³

Companies can measure what the public cannot. In October 2025, OpenAI estimated that about 0.15% of its weekly users showed explicit indicators of potential suicidal planning and about 0.15% showed heightened emotional attachment to ChatGPT, and warned the figures "may change materially."⁴ It shared them voluntarily, in a blog post. Three gaps keep independent eyes out:

1. Researchers risk their accounts. In 2024, 23 AI researchers warned that company terms lead scientists to fear account suspensions or legal reprisal for good-faith evaluation, and called company access programs "an inadequate substitute" for independent research.⁵
2. Only insiders see the logs. Researchers at OpenAI analyzed more than 3 million ChatGPT conversations for emotional cues;⁶ independent scientists have no comparable route, even to data users would gladly share.
3. Nobody reports time. OpenAI says it does not measure success "by time spent or clicks."⁷ New laws in California, Washington, and Oregon require companion-chatbot operators to report crisis referrals,⁸ but no law requires any company to report how long people spend in conversation.

Why legislation: The FTC voted 3–0 in September 2025 to order seven chatbot companies to explain how they measure and monitor harms to children and teens.⁹ A one-time study cannot create continuing public statistics, user rights, or researcher protections; only a statute can. In June 2026 the House passed, 267–117, a bill ordering a four-year NIH study of chatbots and children's mental health,¹⁰ yet no enacted law gives such a study the data it needs. Companies whose products now counsel and keep company with millions of Americans owe the public a way to learn their effects.

The Solution

A four-step staircase: each step stands alone, and each step up asks more of companies and gives the public more to study. Scope: consumer chatbots offered by web or app with more than 10 million monthly U.S. users, defined by function so successor technologies are covered. Chatbot harm warnings, memory controls, and parental controls are addressed separately.

Step 1 — Give research a safe harbor. Bar covered companies from suspending, banning, or suing researchers for good-faith, privacy-protective study of their chatbots or for publishing the results, whatever they show. A bipartisan Senate bill would do this for social media research.¹¹

Step 2 — Let users donate their conversations. Give every user a downloadable record of their own usage and a simple way to share chosen conversations, date ranges, or usage data with a qualified study: redact freely, revoke anytime, parental consent for minors, and no effect on service. One House bill would compel companies to hand funded researchers de-identified data;¹² a user's own choice is the gentler route.

Step 3 — Report usage every month. Covered companies report monthly, on a six-month delay:

• Metrics: active users; total and median active conversation time; sessions and session length; daily-use distributions; nighttime use; and use of memory and companion features. An open browser tab is not a conversation.
• Broken out by: age band (with unknown ages reported), state, product version, and web or app, with small groups suppressed.

The records exist: OpenAI researchers counted 18 billion weekly messages from 700 million users by July 2025.¹³

Step 4 — Fund the long studies. Direct NIH and NSF to fund multiyear studies that link donated records to learning, relationships, and well-being, and to publish null and negative results alike. Costs that appear over years need studies that last years.

Where to start: Step 1 is the floor: complying costs companies nothing. Step 3 is the heart.

Administration and enforcement: The FTC sets definitions and reporting standards within 12 months; compliance follows six months later, and the first public release after the six-month delay. Civil penalties apply to missed or false reports, obstructed donations, and retaliation against researchers; appropriations fund the NIH and NSF studies.

Risks and Mitigations

• Intimate data: Chat logs can hold a person's most private thoughts. Public reports carry only suppressed aggregates; raw conversations move only by the user's choice, into secure research environments with deletion schedules. People named in a donated chat never consented, and that risk remains.
• Hours miss purpose: An hour drafting a contract differs from an hour confiding in a companion. Public counts are therefore paired with donated records and outcome studies; purposes are recorded only when users volunteer them. Averages will still blur individual lives.
• Trade secrets: Companies will call usage data competitively sensitive. The six-month delay and aggregate reporting meet most of that concern; narrow confidential treatment can cover security details, with independent review of disputes. Litigation is likely.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal — 119th Congress

State

What this adds: Three states now count crisis referrals, and the House has voted to study chatbots and children. None measures how much time Americans spend with chatbots, lets users donate conversations to science, or protects the researchers who study them. This proposal does all three and funds the long studies that can use them.