For nearly three decades, our relationship with the internet had revolved around a search box. Users interacted with search engines through short, deliberate phrases such as "symptoms of appendicitis", "best mortgage rates" or "how to file for divorce". The exchange was functional rather than personal. Most people understood that their searches were logged, analysed and used to deliver targeted advertising, creating a natural psychological distance between themselves and the platform.
Generative artificial intelligence has changed that relationship. Instead of presenting a list of links, AI assistants greet users with a blank page and an inviting prompt: "How can I help you today?" The interaction feels conversational, attentive and even empathetic. What appears to be a simple design change has fundamentally altered online behaviour.
Millions of people now disclose information to AI systems that they would never have entered into a traditional search engine.
The reason lies in what psychologists call anthropomorphism-the tendency to attribute human qualities to machines. AI remembers context, responds politely and often acknowledges emotions, creating the impression of a trusted confidante rather than a software application. As a result, users lower their guard.
People routinely paste tax returns, confidential business documents, proprietary software code, legal correspondence and deeply personal messages into AI chat windows, asking the system to edit, summarise or analyse them. In effect, the AI prompt has become a digital confessional.
Yet that perception can be misleading.
Unlike a handwritten journal or a private conversation, every prompt is processed by cloud-based systems operated by a third party. While leading AI companies invest heavily in security and offer privacy controls, the information users submit still enters an external computing environment beyond their direct control.
This shift demands a new understanding of digital privacy. The risks arise when users upload material that should remain confidential.
Personally identifiable information -- including passport details, national identity numbers, banking credentials and residential addresses -- should never be entered into a standard AI chatbot. The same applies to trade secrets, unpublished financial results, confidential customer data, medical records, legal documents and sensitive workplace communications.
Once uploaded, such information is no longer stored solely on a personal device. Depending on the service and user settings, conversations may be retained to provide chat history, reviewed for quality assurance or, in some cases, used to improve future AI models. Although providers publish privacy policies and offer options to limit data use, users should assume that anything highly sensitive deserves stronger protection.
Understanding what happens after pressing "Enter" is, therefore, essential.
Unlike a search engine, which retrieves indexed web pages, an AI system processes user input through sophisticated cloud infrastructure. Prompts are transmitted into servers, analysed to generate responses and, in many cases, stored so conversations can be accessed across devices. Each step introduces another point where information could potentially be exposed if technical failures or security incidents occur.
The consequences of a data leak involving AI conversations can be far more serious than the exposure of a browsing history.
A search history reveals isolated interests; an AI conversation often reveals an individual's thinking. It may contain business strategies, financial plans, health concerns, family problems or confidential professional discussions.
For cybercriminals, such material is invaluable. Detailed conversations enable highly convincing phishing attacks tailored to a victim's circumstances. Knowledge of a person's employer, upcoming projects or family situation can make fraudulent messages appear remarkably authentic.
For organisations, the stakes are even higher. Employees who upload proprietary software, product designs, merger discussions or commercial strategies may unintentionally place valuable intellectual property beyond their employer's direct control. If exposed through a security incident or accidental sharing, the financial and legal consequences could be significant.
Researchers have also demonstrated that anonymity is often more fragile than users assume. Even when names are removed, combining demographic details, employment history and other contextual clues may allow individuals to be re-identified. This illustrates why simply deleting obvious identifiers does not always guarantee privacy.
None of this suggests that generative AI should be avoided. On the contrary, it has become an indispensable tool for research, education, journalism, software development and countless other fields. The challenge is to use it wisely.
