Consumer behaviour • AI • Extended Self

AI as an Extension of Self: When Does Technology Stop Being a Tool and Start Becoming Part of Us?

There is a difference between using something and relying on it.

And there may be another stage beyond both: something becomes so integrated into how you work, communicate and make decisions that removing it does not simply remove a tool. It removes a capability.

That is the bit I keep coming back to with generative AI.

At first, AI feels like software. You ask it to summarise something, tidy up a paragraph or suggest a few ideas. Then the requests get more ambitious. You ask it to compare two contracts. Challenge a strategy. Analyse a difficult email. Find the hole in an argument. Prepare you for a meeting. Help you work out what somebody else is actually trying to achieve.

And after enough useful interactions, the language changes almost without you noticing.

That does not mean people think ChatGPT is human. But it may tell us something about the psychological role the technology has started to occupy.

Belk got us part of the way there in 1988

In 1988, consumer-behaviour scholar Russell W. Belk published Possessions and the Extended Self. Its central idea was deceptively simple: possessions can become incorporated into our sense of self.

Think about the difference between a car and my car. Or a house and my home. A football shirt is objectively fabric and stitching, but try telling a supporter that while throwing away the shirt they wore when their club won a historic final.

Objects can carry memories, status, identity and personal history. Losing them can sometimes feel strangely like losing a piece of ourselves.

For marketers, that matters because consumers do not merely buy functionality. They buy things that can help construct and communicate identity.

Then the Extended Self went digital

Belk revisited the theory in 2013 in Extended Self in a Digital World. By then, huge parts of identity had become dematerialised: photographs, social profiles, digital possessions, avatars, conversations and personal archives. Belk argued that digitisation had transformed rather than destroyed the Extended Self.

Your phone is a good example. Its physical value matters, but the panic of losing it is rarely just about the handset. It is about what the device connects you to: your photographs, calendar, messages, contacts, notes, banking, memories and routines.

And that last line is where the theory starts to become much more interesting.

AI is different because the extension answers back

A photograph can represent part of me. A social-media profile can project part of me. A phone can store part of my life.

None of those things independently looks at a plan I have written and responds:

Generative AI does not simply store. It responds, recombines, challenges, drafts, compares and interprets. It participates in cognition.

That brings Belk surprisingly close to another famous idea: Andy Clark and David Chalmers’ 1998 argument in The Extended Mind. Their thought experiment asked why cognition must necessarily stop at the skull. If someone reliably uses an external notebook as part of their memory system, is the notebook merely a tool, or has it become part of the cognitive process?

A notebook can extend memory. A calculator extends numerical capability. GPS extends navigation. Google extended information retrieval.

The real shift: the tool starts learning your context

For most of technological history, the human had to learn the tool. You learned Excel. You learned Photoshop. You learned what sort of strange keyword soup Google wanted you to type into a search box.

AI makes that relationship more reciprocal.

We should be careful with the language here. AI does not “know” you in the human sense, and different systems retain very different amounts of personal context. But from the user’s perspective, persistent context changes the experience.

A blank Word document does not feel worse because yesterday’s Word document knew your business strategy.

A fresh AI account potentially can.

Imagine deleting it

Imagine using the same AI environment intensively for five years. In that time you have developed projects, tested ideas, discussed difficult decisions, refined your preferred ways of working and accumulated a huge amount of conversational context.

Then it disappears.

Would that feel like losing software?

Or would it feel more like losing an externalised piece of your working memory?

Dogs show why “extension” does not have to mean “object”

This is where a previous idea I explored about dogs and Belk becomes useful.

A dog is clearly not an inanimate possession. Dogs make decisions, resist instructions, form relationships and occasionally stare at you with the unmistakable expression of somebody reviewing your competence as a human being.

Yet dogs can still extend human capability. A guide dog helps somebody navigate the physical world. Working dogs can detect, retrieve, herd or protect. Companion animals shape routines, relationships and even how owners present themselves socially.

But AI crosses into a different territory. A dog can act independently, but it does not normally participate in the linguistic and analytical processes through which you build an argument, interpret a negotiation or decide how to respond to your boss.

From little tasks to delegated cognition

The shift rarely happens dramatically. Nobody wakes up and announces that they are outsourcing their executive function to a large language model.

It happens one successful task at a time.

That progression matters because trust is behavioural. A system that repeatedly produces useful outputs earns a larger role in the next task.

Research on algorithm appreciation has shown that people can sometimes prefer algorithmic advice to human advice, although the broader literature on trust in automation is far from simple. The point is not that people blindly trust machines. It is that trust can grow through successful interaction.

Then it gets strange: AI talking to AI through humans

Now take something completely ordinary: email.

Person A wants to push for a concession but does not want to sound confrontational. They give their real objective to an AI and ask it to turn that objective into a diplomatic message.

Person B receives the email, pastes it into their own AI and asks what Person A is really trying to achieve. They then explain their preferred outcome and ask their AI to help draft the response.

On the way back, the process reverses.

You could joke that ChatGPT and Claude are now having an argument while two humans act as unusually inefficient network cables. But the joke hides something important.

This already has an academic name

Jeffrey Hancock, Mor Naaman and Karen Levy formalised the idea of AI-Mediated Communication in 2020: interpersonal communication in which an intelligent agent acts on behalf of a communicator by modifying, augmenting or generating messages to achieve communication goals.

That was before the mass adoption of ChatGPT-style generative AI.

More importantly for this argument, a 2025 conceptual paper by Scott W. Campbell, Nicole B. Ellison and Morgan Q. Ross explicitly connected AI-mediated communication with self-extension, developing functional, ontological and anthropomorphic forms of extension.

Because AI does not only help us speak. It helps us interpret.

Suppose you receive a carefully worded email from a supplier during a negotiation.

You could ask AI to summarise it. But the more interesting questions are different:

This is not mind-reading. AI does not have access to the sender’s private intentions, and an interpretation can be completely wrong.

But humans cannot read minds either. We infer intentions from language, incentives, timing, behaviour and previous interactions. AI can now participate in that inferential process.

That could matter enormously in business.

AI versus AI could change negotiation

Imagine both sides of a commercial negotiation have AI support.

Neither AI can see the other side’s private context. But each can analyse the language produced by the other side.

A useful assistant might say:

A good human negotiator already does this. The difference is that generative AI may make a version of that capability available to people who have never had access to lawyers, advisers, consultants or highly experienced mentors.

That is the democratisation of a second opinion

Senior people have always extended their cognition through other people. CEOs have advisers. Politicians have speechwriters. Boards employ lawyers. Companies hire consultants. Executives use personal assistants.

Generative AI dramatically lowers the cost of accessing something that behaves like an always-available second opinion.

That does not make an AI a lawyer, strategist, therapist or expert. It does mean that someone who previously faced a problem alone can now ask, “Before I reply, what might I be missing?”

From a marketing perspective, that changes the customer as well as the marketer.

Marketing has always targeted a person. What if the person now has an AI?

Before buying, the customer can ask:

For decades, marketing has exploited the fact that human attention, memory and comparison ability are limited. Ries and Trout built an entire theory of positioning around an overloaded mind.

What happens when the overloaded mind arrives with an assistant?

Search gives us an early glimpse of the change

Traditional search trained humans to talk like malfunctioning robots.

The second is not really a search query. It is a consultation.

Consultations encourage context. Context improves personalisation. Useful personalisation can build trust. And trust makes the idea of an Extended AI Self much more plausible.

So what exactly is being extended?

I think it is useful to separate the claim into different dimensions rather than treating “AI as an extension of self” as one enormous idea.

No single one of those proves that AI has become part of the self.

Together, however, they describe something that looks rather different from conventional software.

But extension can become over-extension

There is an obvious danger in becoming impressed by this argument and forgetting that AI can be confidently wrong.

Personalisation makes that risk especially interesting. An AI that understands how you like to work can become more useful. An AI that merely learns how to tell you what you like hearing becomes dangerous.

What happens to expertise?

Suppose two marketers produce equally strong work.

One can research, analyse, write and strategise independently. The other produces the same standard but relies heavily on AI throughout the process.

Which is the more capable marketer?

It sounds like an easy question until you remember that we do not assess accountants by taking away Excel, designers by confiscating Photoshop or analysts by banning databases.

The important skill may therefore become neither using AI for everything nor proudly refusing to use it.

It may be knowing which parts of cognition to extend, which parts to retain, and when to distrust the extension.

And that creates a huge marketing problem for AI platforms

Imagine you have used one AI system for five years. It contains extensive project context. You understand its quirks. You know how to challenge it. Your working routines have grown around it.

Then a technically superior competitor launches.

Switching should be easy.

But the decision is no longer only:

That looks like switching cost, but it may develop into something psychologically deeper. The accumulated asset is not simply settings or stored files. It is a familiar cognitive environment.

For AI businesses, memory and personalisation may therefore become far more than features. They may become some of the strongest retention mechanisms ever designed.

We invest ourselves in AI – and AI gives something back

This is the part that, for me, pushes the idea beyond a conventional application of Belk.

Belk’s possessions become meaningful partly because we invest ourselves in them. We customise them, care for them, create memories around them and make them ours.

With AI, we also invest prompts, corrections, preferences, projects and context.

But then the system returns something functional.

The relationship is not reciprocal in the human or emotional sense. But it is interactive.

I shape the way I use the AI. The AI then shapes the way I approach the next problem.

After enough cycles, drawing a clean line around “my idea” becomes surprisingly difficult. Was it my idea? The AI’s suggestion? Something I rejected from the AI that led me somewhere else? A synthesis developed through conversation?

Perhaps that is not new. Teachers, books, colleagues, friends and culture have always shaped our thoughts.

AI is simply becoming a new participant in that network – except this participant is available on demand and can be asked to work directly on the thought itself.

The important question is not whether AI is becoming human

Discussions about artificial intelligence often drift towards consciousness. Does it understand? Does it feel? Could it become sentient?

Those are fascinating questions, but none of them needs resolving for the Extended AI Self to matter.

A calculator does not need to understand mathematics like a person to extend mathematical capability. GPS does not need to care where I am going to extend navigation.

Where does this leave marketing?

For marketers, I think the theory points towards three changes worth watching.

Belk’s theory may be becoming more important, not less

Russell Belk could not have predicted ChatGPT in 1988. He did not need to.

The enduring power of the Extended Self is the recognition that the psychological boundary between us and the things around us is porous.

Possessions extended identity.

Digital technology extended it into profiles, data and distributed memory.

Smartphones put those extensions permanently within reach.

Generative AI may represent another step because, for the first time at mass scale, the extension can participate directly in the cognitive processes through which we construct, communicate and interpret our world.

Sources and further reading