In Part I of this four-part series exploring the impact of AI, we outlined the mechanics of how AI works, followed by a contrast of machine intelligence with human intelligence. In this part, we will look at how consumers and businesses are using AI’s unique abilities, and whether this constitutes a true “revolution”.

The Reality of Successful Innovations

We have a tendency to think of “revolutionary” innovations as appearing at sudden, dramatic turning points – explosions that upend everything overnight. The speed of recent AI development certainly seems to satisfy that criterion, but that is not really relevant to determining AI’s true impact. Outside of natural disasters, cataclysmic moments are usually just tipping points. The real revolution has already happened underneath – a slow, quiet, eating away of the old structure until it collapses under its own rot.

This is what we might call revolution by passive absorption. Far from overturning the system, it simply rewires it from the inside out. Not by force, but by daily use, slow adoption, and shaping how people live, think, act and feel – without them ever explicitly choosing it.

For example, the Internet, in the 1990s, was little more than an overhyped phone book, its potential vastly underestimated. Nobel Prize-winning (and therefore usually wrong) economist Paul Krugman famously dismissed its long-term impact as comparable to that of the fax machine. But bit by bit, email replaced letters; search replaced memory; social networks replaced community; and one day we woke up and realised everything had changed. Not because we planned it, but because we allowed it to creep in without resistance.

The same is true also, incidentally, in society and politics. On the morning of November 9th, 1989, nobody you could stop on a Berlin street would have believed that the wall would fall that night. But the disintegration of communist rule in Eastern Europe had been in motion for years – economically, culturally and politically. The wall didn’t fall out of nowhere; it was the crumbling veneer of a foundation that had already hollowed out.

Sometimes there isn’t even a flash point at all. The “march through the institutions” of cultural leftism was a slow, underhand process of gradually infecting key institutions with personnel who ascribe to these views. And now we find ourselves arguing about which bathroom people should be able to use.

Moreover, as Times columnist James Marriott has pointed out, enthusiasm for new tech tends to overlook “an underrated force in human affairs: inertia.” In other words, even when faced with a potentially upending innovation, we tend to stick to familiar habits and routines because their costs and benefits are already known. Change, by contrast, demands effort, uncertainty, and the burden of relearning. Most people make a sudden switch only when the grass is clearly greener on the other side. More gradual change, by definition, makes that comparison harder, so the movement is smoothed out, often invisibly. Hence, as Marriott notes, even something as transformative as electric lighting took more than a generation to catch on.

Revolutions therefore happen not with riots and slogans, but with unnoticed shifts in habits, tools, and relationships – the tacit changes that accumulate until suddenly everything is different and nobody remembers quite how it happened.

This is precisely what is not yet happening with Artificial Intelligence. Indeed, the very fact that we are being told left, right and centre that AI is revolutionary is reason enough to doubt the claim. True revolutions are only perceptible after the event.

Misunderstanding of AI’s capabilities

This is not to deny that AI has made significant breakthroughs in domains bound by rules, structure and existing data. Anywhere the challenge involves pattern recognition, formal logic or vast combinatorial search, AI leaves human (and all previous technological) capability in the dust. It’s already breaking decades-old records in mathematical optimisation, discovering new solutions in problems like Strassen’s matrix algorithm. It’s refining chip layouts, optimising data centre flow, and accelerating its own training routines. In medical imaging, it can detect early signs of illness in facial expressions or X-rays with a level of sensitivity no human doctor can match. And in fields like logistics or scientific modelling, it can generate and simulate possibilities at a scale that would take teams of researchers months. As we indicated in Part I, it still can’t tell you what matters – but once you tell it what to solve, it will find paths we never could.

Beyond that, however, AI pioneers, enthusiasts and doom-mongers alike are beholden to a fundamental category error that most people are still too dazzled to see:

They either want or expect a mechanical machine to solve non-mechanical problems such as:

• Make judgment calls
• Understand nuance
• Create resonance
• Prioritise meaning
• Choose what matters

And the popular fears of AI are built around this premise – that it will replace all human endeavours in the fields which are distinctly the most human. But the failure to understand why AI is necessarily limited in this regard is at least one reason why AI is not yet a “revolution” – or, at least, is not proving to be revolutionary in these particular tasks.

Consumer Use of AI

In spite of all the hullabaloo, direct, intentional, hands-on use of AI by consumers is still a niche interest. While media outlets hype broad usage statistics, a closer look reveals that most consumer interaction with AI is shallow, infrequent and passive – more novelty than necessity. Most people do not converse with ChatGPT, craft prompts, build tools, or integrate AI into their daily lives. Beyond homework cheats, they’re not even sure what it’s for.

Mostly for the classic FOMO (fear of missing out), companies are bolting on AI features to existing apps, e.g. “Ask AI”, “Smart Reply”, “Generate Summary”. However, such additions are mostly cosmetic, awkward, rarely used, misunderstood, and sometimes even distrusted. Users aren’t saying “Wow, AI made this easier!” They’re saying “Why is there a weird button here now?” Indeed, that was precisely the response Meta received to its clunky introduction of an AI feature in the WhatsApp interface, with no option to remove it.

The same could probably be said for the introduction of the Grok model on X and Gemini on Google apps. For some users these may have useful functionality. But they haven’t blended into user behaviour. As optional, novel extras, they are yet to become invisibly absorbed, or so baked into the way we live that we forget life was ever different. Unlike the early Internet era – which flooded the market with new tools, services and behaviours – AI has yet to generate many consumer-facing products that feel distinctly AI-born, rather than just AI-augmented.

The likely reason for this is that people haven’t grasped how AI can actually help them. Part of that is the fault of the companies pushing it – bolting half-baked AI widgets onto apps and calling it a revolution. But as we shall later, the deeper issue is that AI requires interaction. It’s not a dopamine dispenser like TikTok or Netflix. It doesn’t shove entertainment down your throat, but waits. Like a mentor, it needs prompting, dialogue, back-and-forth. And most people simply can’t be bothered because they want push-button magic, not a thinking partner. Which means the only people getting real value from AI are those willing to think with it – not just through it.

Unfortunately, this latter aspect raises the possibility that consumers may be shielded from AI’s true capabilities. As we indicated earlier, the development of the Internet caught the elite off-guard. For about fifteen years, we enjoyed what appeared to be the dawn of liberalisation, wresting the control of information from the hands of newspapers, broadcasters and publishing houses. The empire struck back following Brexit, Trump I, and later COVID lockdowns with the belated regime of digital censorship which we now endure. Having learnt from this “oversight”, much of the current fuss about AI may be an attempt to keep the genie in the bottle – to install a framework of control before people begin to benefit independently. As another author pointed out on this blog, “They’re not scared of AI becoming too smart. They’re scared of you becoming too smart.”

Business Use of Generative AI

The wrong headed adoption of AI is, however, best illustrated by business use which, similarly, seems to have opted for this “cookie cutter” approach to unique tasks. For the sake of brevity, I will focus discussion only on areas with which I am familiar: generative text content, and personalisation.

In spite of all the alarm over AI replacing writers, squeezing decent business copy out of an AI model is actually a difficult endeavour. If I were to hire a human copywriter to produce a sales page, I’d expect that individual to grasp the context and know what was needed without much instruction. That’s what I’d be paying for – not only for the knowledge and technique but also the person’s ability to unearth the context in which the task must be undertaken. An AI model struggles to operate like that.

For instance, the instruction, to an AI model, “write me a sales page for product X” might churn out something that reads okay – based on generic patterns and broadly acceptable tropes. But that model wouldn’t know your audience, the purpose of the campaign, the emotional triggers that matter, the competitive landscape, your brand’s tone, or the reader’s stage of awareness. It wouldn’t know whether to sell hard or soft, whether to lead with pain or curiosity, or how this offer fits into your wider funnel.

All of that has to be fed to the model. Actually writing copy has therefore given way to a new task known as prompt engineering – packaging up everything the AI needs to know so it can perform the job in your place. Which means, ironically, you still need to know how to do the job yourself. A bad copywriter won’t get good results from AI. So far from sacking your copywriter, you now have to become one.

But because people still yearn for that “push-button” experience – perfect output first try – prompts get stretched to cover every possible detail, tone shift and contingency. Sometimes even to the point of being able to manage entire creative workflows without interruption. Before long, your prompt ends up longer than the copy you were trying to write in the first place. It’s like outsourcing your cooking to a robot chef and spending three hours shouting through the wall about how to do your eggs.

To solve this, a cottage industry of generative AI apps has sprung up. These are not AI models themselves – just interfaces that feed pre-engineered prompts into ChatGPT (or another model) to deliver better results on the first pass. In other words, they’re fancy wrappers that know how to ask ChatGPT to write sales pages, emails, blog posts and more, with a polished default. Depending on how flexible they are, their vendors can charge you tens of dollars per month for a subscription that would cost you pennies if you engineered your own prompts in ChatGPT directly. It’s cheaper than hiring a copywriter – maybe. But the trade-off is that you’re stuck with however their prompt engineer thinks your copy should be written.

The result of this is that lazy business owners are now churning out reams and reams of copy that looks and sounds the same. Scarcely a day goes by when I do not receive at least a dozen emails from industry partners written with the same sentence structures, telltale vocabulary, and barely changeable tones of voice. Some of it will, of course, be edited, but minor edits are still locked into the same basic structure and formula. Plus, most of these owners do not know how to change the writing to make it better.

Now, one might be tempted to reply to all of this by saying that “models will develop.” Indeed they will. But all of this points to a fundamental philosophical problem with trying to use AI for these kinds of task with a “push button” attitude – even if models were to grow to be far more expert than they are today. That is the tendency towards homogenisation.

Business opportunity comes not from doing what’s “best” in an absolute sense but from doing what’s better than anyone else. If everyone is doing what’s best, then there is no opportunity, because the market is already saturated with “best”. You have to have an edge.

For instance, if you were to introduce a revolutionary new cutting tool to a sawmill, the lumber company may earn higher profits on account of its ability to produce more yards of lumber per hour. But all else being equal, that invention will cut wood at the same rate for everyone else. So once every sawmill has adopted the invention, profits across the board decline. Moreover, the cheaper and easier it is to adopt that invention, the faster the profit opportunity disappears.

That is the same with AI. But for differences between models, AI copywriting – operated in a “push-button” style – should work the same for everyone. So if it finds that “email subject lines with emojis get higher open rates,” then every brand starts stuffing in emojis. The result? The pattern gets flooded, what once stood out becomes invisible, and all adopters end up with diminishing returns. The same is true with “unique” hooks, angles, ad strategies, social content, even pricing models. If AI is mining the same pool of past successes, then everyone ends up chasing the same tactics – until they no longer work.

AI models could try and account for the fact that other users will do the same thing. But that then just leads us to the same place: everyone else’s model will be undertaking that piece of second guessing as well! So everyone will still converge on the same safe decisions, the same ad templates, the same customer journeys, with brands that sound like each other, look like each other, and operate like each other.

In contrast, your edge, as a human, is the ability to see when the crowd has gone stale – and jump left when everyone else goes right. Businesses using AI to make entrepreneurial decisions are hoping to eliminate this kind of difficult decision. Everyone, at the moment, is chasing speed, automation, and the luxury of not having to think. But faster dumb is still dumb. You can’t “proceduralise” charisma, batch insight, or outsource edge.

More than anything else, breaking the pattern no one else dares to requires nothing less than the quintessential aspect of entrepreneurship: the willingness to take a risk, and to accept the prospect of losing money in the event that you are wrong. However sophisticated AI may get, it cannot predict the future any more than we can. What if a critical supplier goes bust tomorrow? What if there are unforeseen problems with your factory or equipment? Will an innovation, which, by definition, has never been seen before, be welcomed or rejected, and why? How much money should you stake, and how much are you prepared to lose? A machine can’t walk that for you. It can’t feel the pit in your stomach, can’t weigh the personal stakes, can’t tell you what’s worth risking your time, money, health or reputation for. Because it has no skin in the game, no sense of courage, no feelings of regret.

This is the root difference. AI can advise, simulate and optimise, but only a human can commit. That’s why founders still go broke and why investors still make emotional bets. Or why great businesses get built not because the model said “safe,” but because the owner said “Sod the naysayers – I believe in this.” The most that AI can do is tell you what might be relevant for your thinking, and it may, in that regard, suggest an optimisation plan. But even then, a model that highlighted every possible minor contingency would cause only paralysis by analysis. Real value can only ever be created by taking a risk.

Theoretically, if AI gets better at spotting market gaps, those gaps should close faster – good for consumers, less so for anyone hoping to profit. But that speed can backfire. If everyone expects a new opportunity to vanish the moment it’s spotted, the optimal move might be not to bother. Worse, if every model sees the same thing and recommends the same non-action, you end up with a market paralysed by the illusion of rationality. So yet again, entrepreneurs end up having to make their own judgment: do I blink first?

Indeed, one of the common misunderstandings of market “efficiency” is the idea that it just means making things cheaper for consumers. What it really means is best satisfying the ends of all market participants with the resources available. And however piously entrepreneurs may justify their work, their ends also include the desire to be pioneers, innovators, champions – not just rich, but more rich, more admired, more influential than anyone else. We preach equality, but our deepest contest – revealed in every sporting arena – is for dominance. Rationalised sameness has no place in that deeply human urge.

In many ways, the present mania for AI resembles the early 2000s tech bubble – not in financial structure, but in psychology. Back then, companies believed ‘dot-com’ alone guaranteed value – with little concern for how revenue, customer need or basic business operations would follow. Today, AI carries the same mythical status: a magic word that supposedly automates profit, replaces people, and ensures competitive advantage by mere presence. But just as the dot-com bust exposed how little thought had gone into how new tech will create actual utility, so too will the AI gold rush expose who actually understands how value is created – and who simply slapped a neural net on top of a bad idea.

Personalisation

Having, in the previous section, examined how AI proliferation could lead to bland uniformity, we’ll now look at the opposite possibility: that everything becomes more fractured, complex and individual.

John Wanamaker supposedly said, “Half the money I spend on advertising is wasted; the trouble is, I don’t know which half.” The reason being that much of that advertising is viewed by people for whom your product has no relevance whatsoever. How often does anyone leap off the sofa to buy what’s advertised on TV just because it aired between football and the news?

Therefore, the tighter the alignment between offer and audience, the higher the return on your ad spend is likely to be. This much has been known for years, and marketers have always tried as best as they can to target their offerings to the most relevant demographics. For instance, targeting “men aged 30-50 who binge comparison videos on YouTube and click product links after midnight” is more likely to produce results than targeting “middle-aged men in the UK,” as is speaking in a tone of voice that resonates with that group.

AI, however, could take this whole field of personalisation to a different level. Whereas in the past goods and services were advertised to broad but narrowing demographics, future personalisation could target individuals dynamically – not just by interest, but by timing, tone and behavioural patterns. Imagine receiving an email not only promoting the exact thing you’re about to search for, but phrased in the kind of language to which you’re most likely to respond.

Indeed, the ultimate “dream” of personalisation would be to connect businesses with their target consumers at the very moment they are most likely to spend money. Businesses would waste less cash pushing offers to people who don’t care while, in return, consumers are less likely to see ads that feel irrelevant or annoying at any one moment.

Now to be clear, we’re nowhere near this position yet. But it’s important to focus on the possible endgame if we are to understand how far we could or should try to move in that direction.

The obvious Achilles’ heel of this dazzling vision is one massive dependency: data, and lots of it. And the more personal you want to be, the more personal and immediate that data has to be, which brings with it a number of problems.

While companies do collect an awful lot of data – and, indeed, enough to paint a distressingly accurate portrait of your life generally – translating that into useful predictions of behaviour is not quite as easy as it sounds.

What the techno-spies can get at hold of is your location (from GPS, Wi-Fi, IP address); what you click, search, watch, buy, like, hover over, scroll past; your device info (model, OS, battery level, even mouse movements); who your contacts are (if you give access – most people do); which sites you visit if they use tracking pixels (which is most big sites); your message content if they’re on a non-encrypted platform (like FB Messenger).

Disturbingly, most of this data can be gathered and tied back to you regardless of whether you are logged into a platform like Facebook, browsing anonymously, or even if you’re connecting through a VPN. All of your unique browsing configurations can add up to a digital “fingerprint” which allows firms to identify you across multiple sites on multiple devices.

However, the big problem with all of this is that, contextually, much of this data is chaotic and ambiguous.

Suppose you clicked on a jacket. Was it for you, a friend, or just a passing impulse?

You bought a book – was it for you, your boss, or a relative you barely like?

You hovered on a news article – were you reading it, reacting in disgust, or distracted mid-scroll?

You searched for “divorce lawyer” – is your marriage in trouble or are you writing a novel?

They can see what you clicked – but why did you ignore the alternatives?

You may be scrolling through merchandise – are you in a browsing or shopping mood?

These fragments get scraped together and fed to an algorithm that guesses your intent with no memory, no context, and no idea who you are beyond a stitched-together ID tag. What’s sold as “personalisation” is mostly a mess of shallow correlations wrapped in confident language. The reality? They’re not reading your mind as much as just playing probabilistic pin-the-tail-on-the-donkey with your digital crumbs.

So while they have a creepy amount of data, it’s partial, impersonal and probabilistic. You’re a pattern, not a person, to them.

Moreover, there is a philosophical barrier to the usefulness of data in making predictions. That is that actions are not the product of repeatable data. Each human decision is, instead, a one-off – rooted in mood, context, timing, even boredom. Just because you watched a horror film today doesn’t mean you want to watch another to tomorrow. Just because you may have searched “gym classes” doesn’t mean you’ll do it again – especially if you’ve since scratched that itch. In fact, what you want to do today, or are thinking about, may be totally unrelated to anything you’ve ever considered before. But even if you’re in the habit of ordering pizza every Saturday, the decision to continue that habit the next Saturday is a new decision.

Yes, a firm can try to suggest complementary products or anticipate other needs based on past behaviour. But ultimately, hyper-personalisation is trying to reduce something that calls for judgment, taste and empathy into a pattern-matching exercise – a reduction of human understanding into highly wrought statistical wizardry rather than understanding itself.

Now, that’s not to say that increased efforts at personalisation won’t have benefits. But there are risks and drawbacks.

The first, and most obvious, is that it falls flat – in other words, you are offered something you aren’t interested in right now, but you can see why an ad service may have thought that it was appropriate based on your past behaviour. And in that event, it doesn’t just fail, it breaks the illusion. You realise that companies aren’t striving to understand you, your needs, your problems, your wants, your desires – they’re just desperate for you to spend money on some item that you might have been interested in at some point in the past.

Bad personalisation therefore triggers the so-called “uncanny valley” of resistance – too close to be random, not close enough to feel truly human. To them, you’re still just a number, not a person – a breeding ground not for trust, but for cynicism.

However, I submit that – depending on how sophisticated data gathering and AI analysis can get – that personalisation working too well may be the bigger roadblock.

Suppose, for the sake of argument, you just happen to be thinking – thinking – about taking your partner out for a nice meal that evening. While you’re lost in this momentary thought, the camera on your device detects subtle changes in your facial movements and expressions, determining precisely what you are pondering. In an instant, an ad for a local restaurant pops up on your screen.

Such an occurrence is unlikely to be welcome. Instead, it freaks you out, reminding you that some creepy algorithm is watching you, and thus crossing the line from “helpful” to “violating.” Less marketing and more like surveillance with a smile.

So even if it’s right, it’s still wrong. Because people aren’t looking for marketing that’s accurate, they’re looking to be understood, respected, and to hold on to at least a belief that they are independent agents who choose the solution you happen to be offering. When a bot nails them too precisely, it doesn’t build rapport – it’s an intrusion by a machine that’s tracked you down and cornered you screaming “You want this! You want this! You want this!”

The unsettling nature of this activity begins to trigger defence mechanisms – i.e. people change their behaviour in light of the fact they know they’re being watched. They pull back, taking extra steps to ensure they cannot be monitored. In a worst case scenario, they disengage from digital technology altogether – a not inconceivable outcome if techno-surveillance really does ascend to the most intrusive heights. So businesses would be back to where they started.

In any case, personalisation can only ever match consumer wants with existing suppliers. Anything that a person might want within the next thirty seconds has to have already been produced and available for consumption. The bigger questions for business, therefore, are:

a) What do people want not now, but in one to two years time after I’ve built my factory?

And

b) Can that product be scaled, i.e. will enough people be interested in order to justify the required investment in capital goods?

The latter is especially important. There’s not much point in AI predicting that you want a certain type of jacket if few other people want it – the resources are probably better directed to making things that serve more people.

While, therefore, personalisation has possibilities, I still expect the more significant aspects for business to be how AI can examine meta-level trends: What are most people responding to? What do most people seem to want? It may then be better to craft one voice – your voice – that hits most of the pain points while showing you to be an independent and thoughtful business, rather than a robot bending over backwards for every person with a browser.

*     *     *     *     *

As we shall explain in Part IV, the true revolution in AI may not owe itself to AI’s ability to replace humans in non-repeatable tasks, but to the illusion that it can – an outcome which may prove to be far more unsettling.

Before that, however, we will look, in Part III, into how AI can and should be used to enhance human welfare – together with an examination of contrasting views of what that “enhancement” might entail.

==> Go to Part III.