About Half of US Adults Now Use AI Chatbots. What It Means for You

On June 17, 2026, Pew Research Center reported that 49% of US adults say they use AI chatbots, and about a quarter use them daily (survey of 5,119 adults, February 17-23, 2026). For you, the useful skills are knowing what a chatbot is, when its answers need checking and which settings control what happens to your chats.

Key takeaways

A chatbot generates text by predicting likely next words from patterns in its training data, so it can sound confident while being wrong. NIST calls this "confabulation".
Whether your chats train a vendor's models depends on the service and its settings, and turning training off does not stop storage, human review or safety uses on the pages checked.
California's companion-chatbot law has applied since January 1, 2026. Maine requires a notice when a business chatbot could pass for a human, and Utah regulates mental health chatbots. The FTC opened a study of companion chatbots and minors in September 2025.
For health, legal or money questions, treat a chatbot answer as a starting point to verify with an official source or a licensed professional.

What it means for you

First-time users: start with low-stakes tasks and read where chatbots go wrong.
At work: check the privacy table and your employer's policy before pasting anything sensitive.
Parents of teens: the FTC inquiry and California's law, both covered under new rules, focus on chatbots that act as companions.
Students, and anyone with health or money questions: invented sources are a documented failure mode; see the verification notes under where chatbots go wrong.

What a chatbot actually does

An AI chatbot is software you talk to in ordinary language. Most run on a large language model (LLM), trained on very large amounts of text. NIST's generative AI profile describes the core mechanism: LLMs predict the next token or word in a sentence or phrase, generating outputs that approximate the statistical distribution of their training data.

Prediction produces fluent drafts, summaries and explanations quickly. The prediction step itself does not check a claim against a source. Some chatbots can also search the web, but the answer is still generated text, so fluency is not evidence of accuracy.

How many people use them, and for what

Pew found that 44% of US adults use ChatGPT, up from 34% in 2025, 23% in 2024 and 18% in 2023. Pew notes the 2026 question was worded differently, so the jump is not exactly like-for-like. The full report breaks down what people use chatbots for, as a share of all US adults:

Use (Pew, February 2026)	Share of US adults
Searching for information	42%
Tasks at work (employed adults only)	38%
Fun or entertainment	25%
Creating or editing images or video	24%
Getting medical advice	20%
Diet and fitness information	20%
News	13%
Emotional support or advice	10%
Companionship	4%

The same survey found that 51% of adults do not use chatbots, so "everyone is talking about them" is truer of the conversation than of daily habits.

Where chatbots go wrong

NIST's profile, published in July 2024, names "confabulation" as a distinct risk. It defines it as confidently presenting erroneous or false content, often called "hallucination", including answers that contradict earlier statements in the same conversation. NIST says confabulations are a natural result of how generative models are designed, are especially relevant to open-ended, long-form answers and expert fields, and do harm when people believe them because of the confident tone. Models can also produce confabulated logic or citations that appear to justify an answer, and could falsely claim to be human.

Health, legal and money questions

Health: NIST gives the example of a confabulated summary of patient information that could lead doctors to an incorrect diagnosis, and Pew found that 20% of adults use chatbots to get medical advice. Bring a chatbot's answer to a clinician; it is not a diagnosis.

Legal: models can invent citations that look real, so look up any case name, statute or quotation in the original source.

Money: the Consumer Financial Protection Bureau's June 2023 report on chatbots in consumer finance warns that the statistical methods behind LLM chatbots are not well positioned to tell correct from incorrect data, and that inaccurate product information could lead to inappropriate fees. It examined chatbots at banks and other financial companies, but the caution applies to any chatbot asked about fees or rules.

This is general education, not medical, legal or financial advice. For decisions about your own situation, confirm with an official source or a licensed professional.

Chat privacy and training settings

Whether a vendor trains on your chats depends on its policy and your settings. The table summarizes each company's help page as checked on October 20, 2026, listed alphabetically. These pages cover consumer accounts; work, school and API products have separate terms.

Service	Training control	What turning it off does not cover
Anthropic Claude (Free, Pro, Max)	Model Improvement setting in Privacy Settings. Incognito chats are not used for improvement.	Chats flagged for safety review may be used for policy enforcement; a thumbs up or down stores the whole conversation for up to five years, and it may be used for training.
Google Gemini	Keep Activity setting in Gemini Apps Activity. Temporary chats are not used to train Google's models.	Chats are still kept for 72 hours to respond to you; feedback you send can still be used; a subset of chats is reviewed by humans, and reviewed chats are kept up to three years.
Microsoft Copilot (personal account)	Updated app and web (since August 18, 2026): Microsoft says prompts, responses and file contents are not used to train foundation models. Older app: "Training on conversation activity" and "Training on voice conversations" under Privacy.	Updated app: optional feedback may be used to improve Copilot. Older app: opting out covers future activity only; other uses such as product improvement, advertising, safety, security and compliance remain.
OpenAI ChatGPT	"Improve the model for everyone" under Settings, then Data controls. When off, new conversations are not used for training. Temporary chats are not used to improve models.	Turning it off does not delete or hide saved chats; archiving a chat does not change training use; temporary chats may be kept up to 30 days for safety.

OpenAI says its controls depend on sign-in, plan and workspace settings, and that memory is separate from training. Microsoft's updated-app privacy controls include "One shared experience", which lets your chats personalize Bing, Edge and MSN.

Across the pages checked, an opt-out limits training, not storage, review or safety uses. Keep customer records, health details and account numbers out of consumer chatbots.

New rules for chatbots

FTC study of companion chatbots

On September 11, 2025, the FTC issued orders to seven companies: Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI, Snap and X.AI. It asked how they make money from engagement, develop and approve chatbot characters, measure and limit harm to children and teens, inform parents and users, enforce age restrictions, comply with the COPPA Rule and use or share personal information from chats. No findings had been published as of this update.

State laws

California SB 243 (companion chatbots). Governor Newsom signed it on October 13, 2025, and the author's office says it took effect January 1, 2026. Companion chatbot platforms must have protocols to address suicidal ideation and self-harm, disclose that interactions are artificially generated, give minors break reminders and keep them from sexually explicit images the chatbot generates. Platforms report self-harm protocols and crisis-referral statistics to the state, and families may sue noncompliant developers.
Maine L.D. 1727 (approved June 12, 2025). The enacted text bars using a chatbot in trade and commerce in a way that may mislead a reasonable consumer into thinking they are talking to a human, unless a clear and conspicuous notice says they are not. A violation counts as a violation of the Maine Unfair Trade Practices Act. It took effect on September 24, 2025, the general effective date for nonemergency laws from that session.
Utah H.B. 452 (effective May 7, 2025). The enrolled bill covers "mental health chatbots" used by people in Utah. These are generative AI tools that converse like a therapist and that the supplier represents, or a reasonable person would believe, can help with mental health conditions. Suppliers must disclose that the tool is AI, not a human, before access, after seven days away and whenever asked. They may not sell or share users' input with third parties, or their health information apart from listed exceptions, or use input to target ads. Scripted tools such as guided meditations, and tools that only refer users to a human therapist, are excluded.

These laws cover specific kinds of chatbots and specific states. They do not create a general right to be told, in every app, that you are talking to a machine.

Frequently asked questions

Is a "hallucination" the same as "confabulation"?

For practical purposes, yes. NIST uses "confabulation" as its formal term and says the phenomenon is colloquially called "hallucination" or "fabrication". Its definition also covers outputs that drift from the prompt.

Does opting out of training make my chats private?

No. The vendor pages checked describe remaining uses after an opt-out: Google keeps chats for 72 hours to respond to you, ChatGPT keeps your saved chats in history, the older Copilot app keeps safety, security and advertising uses, and Anthropic may use flagged conversations for safety review.

How long do chats stay on file?

It depends. Gemini auto-deletes activity after 18 months by default, adjustable to 3 or 36 months or off. Microsoft's privacy FAQ for the older Copilot app says activity is stored for 18 months by default and can be deleted. OpenAI says ChatGPT temporary chats may be kept up to 30 days for safety.

Can a chatbot's citations be trusted?

Not without checking. Open the cited source yourself and confirm that it exists and says what the chatbot claims.

Does the FTC inquiry mean these companies broke the law?

No. The FTC issued the orders under its Section 6(b) authority, which supports broad studies without a specific law enforcement purpose. Receiving an order means the company is part of the study, not that it violated a rule.

A short checklist before you chat

Open the privacy or data settings of the chatbot you use and decide whether training on your chats should be on or off.
Use temporary or incognito modes where offered, and delete history you do not need.
Keep customer data, health details and account numbers out of consumer chatbots unless your employer approves the tool.
Look up every citation, quote, statistic, dosage and fee figure in an official source before acting on it.
If a teen uses a chatbot that acts as a companion, check the service's own parental controls and age settings.