Your Attention Please

A few days ago, I was in a shop buying a household appliance when someone involved in the sale asked for my email address. I probably shouldn’t have been, but I was thrown a little by this. Consequently I gave a response that was probably a little more diplomatic than “none of your business”, although that was essentially the point I was making. Why did they want it? What were they going to do with it? And what was I going to get in return? These seemed reasonable questions, but they weren’t questions I was expecting to have to ask. Possible more interestingly the shop assistant didn’t seem to know the answers.

You know I’ve lost count of the number of organisations I’ve given my email address to over the years, often in circumstances where the immediate benefit to me was negligible or non-existent. Sometimes there’s a discount ( although even that can be a dubious bargain if the price of saving a few pounds is a permanent place on somebody’s marketing list.) On other occasions though, it’s simply presented as something the organisation needs, without any particularly convincing explanation of why.

But does the organisation actually need it? And if it does, what for?

There’s a principle in data protection that organisations should collect only the personal information they need for a specified purpose. It sounds straightforward enough. If you don’t need somebody’s email address, don’t ask for it. If you do need it, explain why. Yet we’ve become so accustomed to being asked for personal information that we often hand it over without thinking, as though the request itself establishes a right to have it.

we’ve become so accustomed to being asked for personal information that we often hand it over without thinking, as though the request itself establishes a right to have it.

But here’s something worth considering. In that shop, somebody had to ask me for my email address, and I had the opportunity to question why they wanted it. When I use AI, I can volunteer far more information than an email address, without anybody having to ask. I can describe my finances, discuss my health, explain a family dispute, share my political opinions or upload a document containing details about other people. I can do all of that in the space of a few minutes, and because I’m getting something useful in return, I probably won’t stop to consider what I’ve handed over.

The shop assistant wanted an email address. But AI can get hold of something much more revealing: a picture of how I think, what I care about, what worries me and how I might behave. And I may have provided the raw material for that picture myself, one apparently harmless conversation at a time.

We spend a great deal of time worrying about what AI tells us, whether its answers are accurate, whether it makes things up and whether we can trust its recommendations. Those are important questions, and I’ve written about them elsewhere a number of times. Hopefully the message is starting to get through that we shouldn’t just trust AI output, that the responsibility for thinking still rests with us. But, unfortunately, we’re paying much less attention to the information travelling in the other direction.

What are we putting into these systems? And what are we giving away in the process?

Off the top of my head we type in personal details, describe our problems, explain our thoughts and feelings, discuss our political and religious views, upload documents and ask for advice about situations involving other people or organisations. And we do it because the system seems helpful, because the conversation feels natural and because it’s easy to provide more information when we want a better answer. After all, that’s what all the AI training courses, and lessons on how to prompt, tell us to do. Sometimes we simply don’t stop to think about what we’re sharing.

we do it because the system seems helpful, because the conversation feels natural and because it’s easy to provide more information when we want a better answer ... Sometimes we simply don’t stop to think about what we’re sharing.

Your attention please.

It’s the sort of announcement you hear at a railway station or airport, usually just before somebody tells you that your train has been cancelled, your gate has changed, or your carefully planned journey is about to become considerably more complicated. We’ve become accustomed to hearing it without necessarily paying much attention. Perhaps that’s the problem. There are things happening around us that deserve our attention, but we’re busy looking elsewhere.

And while we’re looking at the answers on our screens, we may be overlooking the information we’ve put into the machine to get them.

So I think we should pay a little more attention to the information travelling in the other direction. We spend a great deal of time worrying about what AI tells us, while overlooking what we tell it in the course of these apparently innocent conversations.

Because there’s something different about a system that responds conversationally. We can explain ourselves, provide more context, correct a misunderstanding and ask follow up questions. It feels like a conversation, rather than the submission of information to an organisation. But the conversational experience tells us nothing, by itself, about how the information is handled, how long it’s retained, what settings apply to its use, or who might be able to access it.

The information doesn’t have to be especially sensitive to be revealing. It might be a discussion about work, a disagreement with a colleague, a family problem, a financial decision or an opinion we haven’t shared publicly. Taken individually, these details might seem unremarkable. Put together, they can build a surprisingly detailed picture of a person.

Taken individually, these details might seem unremarkable. Put together, they can build a surprisingly detailed picture of a person.

Whenever we use ChatGPT, Claude, Gemini or any of the other AI services now available, we need to think about how much information the task actually requires. And when we upload information about other people, we’re making decisions about their privacy too, sometimes without their knowledge or permission.

Of course, the usefulness of AI often depends on context. There is no doubt that providing relevant information will produce a better result. But that doesn’t justify giving a system every detail that happens to be available. We should share what the task requires, not everything we happen to know or feel like saying.

This becomes particularly relevant as AI moves into more of our everyday lives. I was interested in the introduction of Alexa+, Amazon’s conversational AI assistant, which began rolling out in the UK in March 2026. For years, people have worried about the little device sitting in the corner of the room, apparently waiting to be addressed. There have been stories about people discussing something and subsequently seeing an advert for it on Facebook, along with the familiar jokes about telling people to be quiet “in case Alexa is listening.”

Amazon says its Alexa devices detect a wake word and send audio to the cloud when the device is activated, rather than continuously recording and transmitting every conversation. Accidental activations are fairly common though. We’ve all seen the blue ring suddenly appear when a random word has been said by somebody in person, or most likely on the TV. And I don’t know if it’s just me but sometimes the device activates when no one has said anything. This obviously raises legitimate questions about what might be captured and when ( particularly when something raised in a conversation suddenly and mysteriously turns up on your Facebook feed.)

And there’s also documented history of human review. In 2019, reports revealed that Amazon employees and contractors listened to a sample of Alexa voice recordings to help improve the service. Amazon acknowledged the practice. The recordings weren’t evidence of every conversation being secretly monitored for advertising, but they did demonstrate that what people regarded as a private interaction with a device could, in some circumstances, be heard by a human being.

Alexa+ takes the experience further by making interaction more conversational and allowing it to use context and remember information to personalise responses. An assistant that remembers relevant preferences doesn’t have to ask the same questions every time. That’s useful, but it also means that we need to pay attention to what we tell it and what it retains.

An assistant that remembers relevant preferences doesn’t have to ask the same questions every time. That’s useful, but it also means that we need to pay attention to what we tell it and what it retains.

During the Alexa+ onboarding process, there was a clear message on the screen of my phone that told me not to tell Alexa sensitive or secret information. It was good advice. The same rule should apply to every conversational AI system.

Don’t disclose sensitive information unless there’s a clear reason to do so and you understand the relevant safeguards. That’s a basic principle of information governance, no different from the advice we might have been given about email, online forms or social media years ago. Yet the more natural the interaction becomes, the easier it is to forget that we’re sharing information with a technology service.

And what happens when the AI system remembers things between conversations?

If you use a service that has memory or personalisation features enabled, it may retain selected information from previous interactions to help it respond more usefully in the future. The details and controls vary between services, and memory isn’t necessarily a complete record of everything you’ve ever said. But information supplied in one conversation can influence another conversation days or weeks later.

You can test this for yourself. Ask ChatGPT, Claude or Copilot what it knows about you. Ask what it remembers from previous conversations, if that feature is available and enabled. Ask what it thinks your interests are, what your likely views might be on a particular subject, or how it would predict you’d respond to a given situation. You could even ask it to estimate your financial circumstances or where you live.

The answers may surprise you – for me they were surprisingly accurate. Some information will have been provided explicitly. Other details may have been inferred from what you’ve said, the questions you’ve asked, the preferences you’ve expressed or the connections between different pieces of information. Some conclusions will be accurate, others less so. An AI system can make a confident sounding inference that is completely wrong, so its response shouldn’t be treated as an authoritative dossier on your life. But even an imperfect picture reveals how much information we’ve supplied, or how much can be inferred from it.

Some information will have been provided explicitly. Other details may have been inferred from what you’ve said, the questions you’ve asked, the preferences you’ve expressed or the connections between different pieces of information

Of course, memory and personalisation have real benefits. I don’t particularly want to have to explain my preferences, circumstances and working methods every time I use a system. Having relevant context available can make the interaction more efficient and the answers more useful. But convenience comes with a responsibility to understand what is retained, how it is controlled and what might happen if the information is exposed or used in ways we didn’t anticipate.

The problem extends beyond our own information. People regularly paste emails, reports, meeting notes and other documents into AI systems to summarise them, rewrite them or help solve a problem. Those documents may contain personal information about colleagues, clients, patients, students or members of the public. They may include confidential commercial information or details that were provided for a very specific purpose.

The fact that an AI tool can process something doesn’t mean we have permission to give it that information. Nor does removing a name necessarily make a document anonymous. A combination of dates, roles, circumstances and other details can sometimes make an individual identifiable even when their name has been removed.

We should be equally clear about the responsibilities of the companies providing these services. They operate within legal and regulatory frameworks, but they also face commercial pressures, make technical decisions and sometimes get things wrong. Privacy policies and assurances are important, but they aren’t a substitute for thinking carefully about what we disclose.

Privacy policies and assurances are important, but they aren’t a substitute for thinking carefully about what we disclose.

Copyright disputes involving AI companies illustrate how contentious the use of information can become. There have been legal challenges over books and other material used to train AI models, including a major 2025 settlement by Anthropic concerning millions of books obtained without permission. The settlement resolved claims without establishing that every allegation was proved, but the dispute itself demonstrates that the rights attached to information and creative work cannot simply be assumed away because a technology company finds the material useful.

We would be foolish to assume that every question about information rights has already been settled, or that privacy will always take precedence over commercial interests. That makes it all the more important to understand what we’re handing over and to use the available controls rather than relying on vague assumptions about how our information is treated.

And this brings us to what might happen if personal information reaches the wrong hands.

Imagine the picture that could be assembled from someone’s conversations over months or years. Their interests, worries, financial pressures, relationships, political opinions, ambitions and vulnerabilities. Some of that information might be explicit. Some might be inferred. Some might be out of date or wrong. But taken together, it could provide clues about what matters to that person, what frightens them, what they want to achieve and what might persuade them to act.

Now imagine someone trying to exploit that knowledge.

Scammers already use personal information gathered from social media, data breaches, public records and other sources to make their approaches more convincing. A message that refers to a real employer, a family member or a recent event is more likely to attract attention than an obviously generic attempt. Detailed personal information gives a fraudster more opportunities to tailor a message, identify a plausible pretext or work out which kind of pressure might be effective.

Detailed personal information gives a fraudster more opportunities to tailor a message, identify a plausible pretext or work out which kind of pressure might be effective.

And, of course, criminals don’t need direct access to an AI chatbot’s private memory to benefit from a detailed picture of someone. Compromised accounts, data breaches, information shared carelessly and personal details gathered from different sources can all contribute. The more information available about us, the easier it can become to construct a convincing lie.

We should also recognise how much can be revealed without being explicitly stated. Someone may never say that they’re financially desperate, lonely, anxious about their job or worried about a family member. But a series of conversations could provide clues. A fraudster doesn’t need a perfectly accurate psychological profile. They may only need a plausible opening and enough information to make their approach seem credible.

That’s why the information we disclose deserves as much attention as the answers we receive. Once information has been shared, we may have limited control over how it is retained, combined with other information or exposed through circumstances beyond our control. Prevention begins with considering what we provide in the first place.

The principle I came back to in that shop was a simple one. If someone wants my email address, I’m entitled to ask why. If they need it for a legitimate purpose, they should be able to explain that purpose. If they don’t need it, they shouldn’t be asking for it. And if the benefit to me is negligible, I’m entitled to decide that the exchange isn’t worth making.

The same principle applies to AI. Before pasting in a document, describing a personal problem or telling a chatbot something about another person, ask whether the information is necessary. Could the task be done with less detail? Could identifying information be removed? Do you understand the service’s relevant privacy settings? Are you entitled to share the material in the first place? And would you be comfortable if the information became accessible to someone beyond the audience you have in mind?

Before pasting in a document, describing a personal problem or telling a chatbot something about another person, ask whether the information is necessary. Could the task be done with less detail?

These aren’t reasons to abandon AI. They’re reasons to use it sensibly. We share information every day, and many services depend on it. But we should make those decisions consciously, with some understanding of what we’re sharing and why. A friendly, conversational interface doesn’t provide the same privacy as a conversation with someone we know and trust.

And that’s what we need to hear more often. We’ve become very good at asking AI to pay attention to us. It’s time we paid attention to what we’re telling it.

We don’t need to tell AI everything just because it will listen. And we shouldn’t confuse the feeling of a private conversation with a guarantee of privacy. All it needs is for you to think about what you share and give it your attention, please.