AI privacy has become a peculiar bargain. People are using chatbots to solve problems, search for answers and get work done, while many also expect artificial intelligence to make their personal information less secure. A Pew Research Center survey of American adults captures this contradiction: 49 per cent said they had ever used an AI chatbot in February 2026, yet a large majority expressed concern about the technology’s future impact on personal data.
That is not evidence that the same individuals are deliberately handing over information they believe will be stolen. It is a population-level picture of adoption alongside anxiety. People can find a product convenient, distrust the companies behind it, and remain uncertain about how their conversations are handled. For publishers and technology firms, the important lesson is that usage figures are not a substitute for public confidence.
AI privacy fears have survived rapid adoption
Pew’s headline finding is that chatbot experience has spread substantially. In its 2026 survey, 49 per cent of American adults reported ever using a chatbot such as ChatGPT, Gemini or Copilot. The comparable figure shown for 2024 was 33 per cent, although Pew notes a difference in how that earlier question was presented. The broad trend is clear even without pretending the two figures are a perfect repeated experiment.
The same research found widespread doubts about the direction of AI. Many respondents believed it was developing too quickly and expected negative consequences for society, themselves or their information. Pew reported that 71 per cent expected AI to make personal information less secure over the next twenty years. Such a prediction is an attitude, not a measurement of security breaches caused by chatbots in 2026.
Both facts can be true without contradiction. A person may use AI to draft a harmless email but refuse to include confidential business figures. Another might be happy asking for recipe ideas but wary of a chatbot’s role in financial decisions. Trust is seldom a single switch that is either on or off. People make different judgements depending on what they need and what could go wrong.
The crucial question is whether people understand that distinction. A chatbot’s fluent writing style makes the interaction feel conversational and private, even when it involves a commercial service processing text remotely. Friendly presentation can make it easy to forget that data handling depends on an organisation’s terms, settings, product tier and retention rules, rather than the personal tone of the response.
The British picture adds another warning
American surveys cannot automatically describe Britain. Nevertheless, the Office for National Statistics’ June 2026 public-attitudes bulletin offers a useful second perspective. Its questions were framed differently from Pew’s, so the percentages must not be placed side by side as though they measured the same behaviour. The ONS found substantial public worry about the potential misuse of information and other risks from AI.
The UK findings help explain why privacy concerns do not vanish when technology becomes familiar. Someone who has seen convincing AI-generated content may also wonder how their own words, photographs or documents could be reused. A tool can be useful in ordinary circumstances while still raising difficult questions about information sharing, identity and accountability. Familiarity with the interface does not answer those questions.
Surveys also capture expectations about the future, not necessarily the factual security of any specific application. There is a difference between believing that AI will weaken privacy over the coming years and proving that a particular chat is currently exposed to unauthorised people. Reliable journalism needs to preserve that difference rather than turn a worrying poll result into an unsupported warning about every provider.
For ordinary users, the sensible response is neither panic nor unlimited trust. It is to understand which information belongs in a consumer chatbot and which information requires a more secure or appropriate channel. The exact answer depends on the service and the sensitivity of the material, not on a headline about the industry as a whole.
Why the chat window can encourage oversharing
Typing into a chatbot feels less formal than submitting an online form. People may write as though they are speaking to a helpful colleague, including extra context they would never put into a conventional search engine. Sometimes that context is exactly what makes the answer useful. Sometimes it includes names, account references, work documents or details about other people that were not needed for the task.
A practical distinction is between a question and the personal data used to illustrate it. Asking how to challenge an electricity bill rarely requires submitting the full account number. Requesting a plain-English explanation of a letter may not require sharing the address and date of birth printed on it. Removing identifying details can preserve much of the usefulness while reducing unnecessary disclosure.
Deletion is another area where everyday expectations can be misleading. Removing a conversation from a visible chat history is not necessarily identical to immediate removal from every backend system. Retention and deletion arrangements vary across services, and legal or security obligations may affect them. LiveAIWire has explained why deleting an AI chat may not erase every copy at once, a distinction worth checking before assuming the screen tells the whole story.
This is also why people should not use chatbot privacy claims as a substitute for reading the relevant product settings. Training controls, temporary chats, business-account protections and data retention can be materially different. A company policy can change, and a feature available to one group of users may not apply to another. The name of the model does not, by itself, specify what happens to every message.
The risk is not limited to obvious secrets
A single harmless-sounding fact may reveal little. Several small details combined can create a much clearer picture of a person. Travel plans, workplace arrangements, family routines and the content of private correspondence may become sensitive in combination. The issue is not that every such detail is dangerous, but that the user may not have consciously chosen to share the entire picture.
Workplace use makes that judgement more demanding. A meeting summary can include client names, commercially sensitive plans or opinions that participants never intended to publish. Uploading it into a chatbot may raise company-policy or contractual questions even if the user has no malicious intent. LiveAIWire’s guide to AI meeting notes and sharing permissions examines why convenience should not obscure the people whose information is involved.
The boundaries grow less obvious when assistants can connect to mail, calendars and documents. A tool may require access to a broad folder to answer a narrow question, depending on how its permissions work. Users should distinguish the benefit of the requested action from the scope of the access being granted. The easiest connection is not always the most appropriate one for sensitive records.
There is a separate question about publicly available information used in model development. That concerns the rights and expectations of people whose material was published elsewhere. LiveAIWire has discussed the OECD’s concerns about chatbot data and web scraping, which belong to the broader privacy conversation but should not be confused with the storage of a particular private chat.
How to use AI without treating it as a confidant
The safest useful habit is to start with the least information necessary. A description of a problem can be enough to obtain a checklist, explanation or suggested wording. Replace personal identifiers with generic labels while drafting, and put accurate details back only in the document or service where they genuinely belong. This approach does not make all risk disappear, but it avoids surrendering information unnecessarily.
For important decisions, establish what the chatbot can and cannot verify. An answer about consumer rights, tax, visas or medical care should point the reader towards the relevant authority, which should then be checked directly. Privacy and factual accuracy are separate problems. Removing personal details protects information; verifying sources protects the decision. Good use requires both.
Where an assistant is connected to other services, permissions deserve a second look. Ask whether it needs continuing access, whether the connection can be narrowed and what will happen to information it retrieves. Review the provider’s published documentation rather than relying on a conversational promise from the assistant. A model cannot grant contractual guarantees by simply saying it is private.
Trust will have to be earned by design
The Pew numbers should discourage a lazy interpretation of adoption. Almost half of American adults having tried chatbots does not mean almost half are comfortable with every use of them. And public anxiety does not imply people have rejected AI. It suggests that the market is growing faster than a shared understanding of where the boundaries should be.
Providers that explain retention plainly, minimise unnecessary access and make controls understandable may have an advantage as people become more discriminating. Regulators and independent researchers will continue to matter because assurances from an interested company are not the same as outside verification. Users, meanwhile, need practical explanations rather than vague claims that a platform is safe or unsafe in every circumstance.
The most useful rule is simple: give the chatbot the problem, not automatically the private life surrounding it. AI can help draft, organise and explain without being entrusted with every detail. If privacy is to catch up with adoption, both the tools and the habits around them must improve. The surveys show there is plenty of appetite for convenience, but not yet unquestioning confidence.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.