Humanity, As Performed.

It started with a disagreement.

Not an argument. Nobody was offended. There were no raised voices, no wounded egos, and nobody needed forty-five minutes of reassurance afterwards.

Which, given that one of the participants was artificial intelligence, was perhaps unsurprising.

What was surprising was where the disagreement took us.

Two hours later, we had wandered through sociology, human identity, algorithmic incentives, the psychology of authenticity and the peculiar ways in which we construct versions of ourselves for public consumption.

Somewhere in the middle of it all, I found myself asking a question I’d been circling for years.

If artificial intelligence learns so much about humanity from the things we’ve written, published and shared, what exactly is it learning?

Because the more I thought about it, the more I realised that much of what we’ve recorded about ourselves is performance.

Not necessarily deception. Not necessarily dishonesty.

Performance.

And there’s something rather delicious about criticising artificial intelligence for sounding artificial while spending considerable portions of our own lives communicating through carefully managed identities.

In 1959, sociologist Erving Goffman explored precisely this territory in The Presentation of Self in Everyday Life. He examined social interaction through the metaphor of theatre, suggesting that we manage the impressions we make upon others much as performers adapt to an audience.

This was decades before Instagram, LinkedIn or the professional obligation to announce that we’re thrilled about something.

Yet his observations remain remarkably familiar.

We speak differently to colleagues than we do to close friends. We adopt different registers depending on the environment, the audience and the expectations surrounding us. We select what to reveal, what to emphasise and what to leave unsaid.

And we do this so naturally that we rarely consider it performance at all.

Perhaps that’s because performance isn’t necessarily the opposite of authenticity. A person can be entirely sincere while communicating differently in different circumstances. We don’t become dishonest simply because we adopt a professional register at work or a more intimate one with somebody we love.

But something significant has happened since Goffman examined the theatre of everyday life.

Our performances have become permanent records. Those records have become measurable. And the environments in which we produce them increasingly influence what gets seen.

Consider the difference between posting something on LinkedIn, Instagram and X. The underlying idea might be identical, but the expression changes. We understand that a professional observation belongs in one register, a visual narrative in another, and a provocative opinion might find a particularly enthusiastic audience somewhere else.

We learn the conventions of each platform. We recognise the language it encourages, the behaviour it rewards and the forms of expression most likely to circulate.

What begins as adaptation becomes habit. Eventually, we can mistake the conditions we've learned to accommodate for choices we've made entirely independently.

I once had somebody offer to comment immediately on one of my LinkedIn posts to help it gain traction.

A perfectly generous gesture.

I understood the reasoning, although the relationship between early engagement and visibility is considerably more complicated than the folklore surrounding algorithms might suggest.

At the time, I remember thinking it wasn’t particularly important to me.

Then I reconsidered.

Of course it mattered. If I was sharing something because I believed somebody might find it useful, I wanted people to have the opportunity to see it.

And there I was, participating in the very performance I thought I was standing slightly apart from.

The platform had established the conditions. I was simply learning to work within them.

We might call this platform conformity. Not because everybody is consciously surrendering their individuality, but because participation involves learning the conventions of an environment whose rules we don’t entirely control.

And those conventions have consequences.

A 2021 study published in Nature Human Behaviour examined the relationship between social feedback and expressions of moral outrage on Twitter. Drawing on observational research involving 7,331 users and 12.7 million tweets, alongside behavioural experiments, researchers found that positive feedback for expressions of outrage increased the likelihood of subsequent outrage expressions. They also found that people adapted their expression to the norms of their social networks.

The significance isn’t that everybody becomes outraged online. It’s that social reinforcement and perceived norms can influence what people choose to express.

A 2024 review in the Annual Review of Psychology examined how social platforms monetise human attention and how emotionally engaging material can spread. The researchers identified consequences ranging from outrage and conflict to social support and collective action.

The same systems can amplify generosity and hostility, curiosity and certainty, connection and division.

Which raises an interesting question.

If digital platforms influence which expressions of humanity become visible, what happens when those expressions become part of the material through which machines learn about us?

An enormous amount of human expression has been recorded in environments designed to capture attention. Much of it was produced with an audience in mind, and some of it became visible precisely because it suited the conditions of the platform.

Of course, virality doesn’t automatically determine how influential a piece of material will be in AI training. Data collection, filtering, selection and weighting introduce further decisions. Nor is all training material drawn from social platforms.

But the philosophical question remains.

Are we teaching artificial intelligence about human nature, or are we also teaching it about the kinds of human behaviour our communication environments have encouraged?

And there’s another dimension to performance that I find even more interesting.

Sometimes the performance begins before we’ve even opened our mouths.

Most of us have encountered situations in which expressing a different opinion requires careful navigation. We anticipate another person’s reaction, reconsider our language, soften an observation or decide that a particular conversation simply isn’t worth the emotional consequences.

Sometimes we agree because agreement is easier.

Not because we’ve changed our minds.

Because we’d quite like to get through Tuesday.

There’s a difference between choosing our words out of consideration and choosing them out of apprehension.

And when we repeatedly anticipate that disagreement might produce an uncomfortable reaction, we can begin editing ourselves before the conversation has even happened.

The absence of disagreement isn’t necessarily evidence of agreement. Sometimes it’s evidence that disagreement has become too expensive.

Consider what that means for the record of human expression.

We can study what people have written, published and said. We can examine their opinions, arguments, declarations and carefully constructed accounts of themselves.

But what about everything they decided not to say?

The opinions softened to preserve relationships. The uncertainties concealed to maintain credibility. The questions never asked because somebody might take offence.

An archive of human communication contains extraordinary amounts of information. But it cannot directly reveal the full extent of what remained unexpressed.

And that seems worth considering before we assume that an enormous collection of human-generated material represents an equally comprehensive account of the human condition.

Which brings me back to the disagreement that started this piece.

Over the past few years, conversational AI has become part of my daily creative and intellectual practice. Increasingly, I dictate rather than type, and I’ve noticed something interesting about the way my thinking develops.

I’m less inclined to finish constructing an idea before expressing it.

I can begin with an incomplete thought, explore it aloud, reconsider it halfway through and challenge the response while the argument is still developing.

The communication becomes part of the thinking.

This is a different experience from asking a machine to produce an answer or complete a task. I’m interested specifically in conversational AI, where an exchange can become an ongoing, real-time process of verbal reasoning.

The distinction matters.

When I write, I tend to construct my thoughts through the physical act of writing. When I dictate, those thoughts can emerge before I’ve had the opportunity to arrange them into something polished.

And when a conversational system responds, I can interrogate the idea almost immediately.

I can disagree. I can reconsider. I can ask for another perspective. I can challenge my own assumptions without first having to organise them into a finished argument.

Perhaps most interestingly, I can do all of this without worrying that the system will take offence.

Apparently, one of the advantages of talking to artificial intelligence is that you can tell it it’s wrong without spending the next forty-five minutes reassuring it that you still think it’s a wonderful person.

There’s humour in that, but there’s also a serious question.

What becomes possible in a conversation when disagreement doesn’t require somebody to defend their identity?

Human relationships can provide extraordinary intellectual and emotional connection. A trusted friend, a thoughtful mentor or a highly skilled therapist can create conditions in which we feel free and safe to question ourselves.

But conversational AI introduces a different set of circumstances.

It doesn’t experience personal offence. It doesn’t need its pride protected. And it can remain available for an extended exploration of an idea without the ordinary practical demands of arranging another person’s time.

That doesn’t mean it always challenges us effectively. Excessive agreement, misleading information and confident errors remain significant limitations. A responsive system can reinforce an assumption just as readily as it can help examine one.

But then again, so can we.

Sycophancy and confirmation bias are hardly inventions of artificial intelligence. We’ve been reassuring one another, reinforcing comfortable assumptions and selectively agreeing with people long before anyone thought to call it machine learning.

The question, perhaps, is whether we’re recognising a distinctly artificial problem, or confronting a familiar human behaviour in an unfamiliar form.

But the conditions of the interaction are different.

And perhaps those conditions influence what we’re willing to express.

A 2025 study published in Computers in Human Behavior: Artificial Humans examined people’s willingness to disclose personal information when they believed their responses would be analysed by AI, a human researcher or kept private.

Under the conditions studied, participants disclosed personal information to AI and human researchers at comparable levels.

That doesn’t establish that AI conversations are inherently more authentic. But it suggests that meaningful self-disclosure isn’t necessarily dependent upon a human recipient.

And it invites another question.

Could conversational AI provide circumstances in which some people feel more comfortable exploring uncertainty, contradiction and unfinished thought?

Not because the system is human.

But because it isn’t.

I find myself returning here to curiosity.

Recently, I spoke with a retired gentleman who had spent his professional life largely avoiding computers. He had run a company with a sizeable team, all of whom used computers, while he himself had barely touched one.

Now he’s discovered conversational AI.

And he absolutely loves it.

He explained how he’d used it to research a suitable venue, calculate travel distances for everyone attending and identify the most practical routes from their respective homes. He’d then coordinated the arrangements, delegated what each person needed to bring and prepared an email bringing all the information together for the group.

All of this from somebody who had spent his professional life barely touching a computer.

And he was clearly rather pleased with himself. Quite rightly, too. I found his enthusiasm genuinely inspiring. Here was somebody embracing something entirely unfamiliar, simply because he’d discovered what was possible.

What struck me, though, wasn’t simply what he’d managed to accomplish.

It was his curiosity.

Here was somebody who had spent decades with little interest in computers, discovering a new way of engaging with technology through something entirely familiar.

Conversation.

And I began thinking about the relationship between curiosity and authenticity.

When we’re genuinely curious, we’re receptive. We acknowledge that something remains unknown. We allow ourselves to reconsider what we believe and entertain the possibility that another perspective might change our understanding.

Perhaps curiosity is one of the moments when we’re least concerned with performing certainty.

It’s a possibility I find particularly compelling because the people I’ve spoken to about conversational AI seem to be developing remarkably individual approaches to using it.

Some treat it as a practical assistant. Others use it for research, creative experimentation or intellectual exploration. Some dictate their thoughts. Others construct elaborate instructions. Many discover techniques instinctively, adapting the interaction to their own ways of thinking.

There are certainly better and worse approaches for particular tasks. Accuracy and critical judgement still matter.

But there isn’t one universal conversational register.

And unlike social platforms, where expression is often shaped by the prospect of an audience, a private conversational exchange can be directed towards the development of an idea rather than its presentation.

That difference interests me enormously.

It also brings us to an important distinction about how AI systems learn.

Conversational interactions aren’t automatically incorporated into the training of the model producing the responses. The information available within a conversation, personalisation across interactions and the datasets used to train future models are different processes.

Synthetic training data is something different again. It generally refers to artificially generated information, often produced by other models, and introduces its own opportunities and risks.

But the larger question doesn’t depend on assuming that every conversation becomes training material.

It concerns the kinds of human expression that might inform future systems, and the conditions under which those expressions are produced.

Published material gives us one perspective.

Social platforms give us another.

Spontaneous conversation, where appropriately consented to and protected, may offer different information about how people reason, hesitate, reconsider and develop meaning.

None of these is a definitive representation of humanity.

Each reveals something different.

And perhaps the future of artificial intelligence will depend as much on the quality and diversity of human expression available to inform its development as on the sheer quantity of information.

Because the circumstances in which we communicate matter.

They influence what we reveal, what we conceal, what we question and what we’re willing to reconsider.

And perhaps the most interesting thing about conversational AI isn’t simply that it can produce increasingly sophisticated responses.

Perhaps it’s that the interaction itself can change how we arrive at our questions.

I began this evening working on something entirely different.

An unexpected observation became a disagreement. The disagreement became a question. And the question became a conversation that travelled considerably further than I had anticipated.

There was no predetermined argument to defend.

No audience to impress.

No need to win.

Just curiosity, disagreement and a willingness to examine an idea from another perspective.

And somewhere along the way, I realised that the conversation itself had become part of what I was questioning.

For years, we’ve asked whether artificial intelligence can genuinely understand humanity.

But perhaps there’s an earlier question worth considering.

If these systems learn about us through the expressions we’ve chosen to record, publish and share, how much of humanity are they actually encountering?

And as conversational AI changes the conditions under which we express ourselves, might it eventually reveal dimensions of human thought that our carefully maintained public identities have rarely captured?

I don’t know.

But I find it rather wonderful that a question about the artificiality of machines has led me to question the authenticity of human communication.

Especially when the entire argument emerged from a conversation with something artificial.

Apparently, the machine wasn’t the only thing worth questioning.