More and more companies are delegating decisions and processes to artificial intelligence, trusting in the idea that data alone can guide the future. Yet AI, being based on statistical induction, tends to replicate the past – and this is where a profound strategic risk lurks: predictive models ignore the exception, even when it is precisely the exception that reshapes competitive advantage.
Through a critical rereading of Hume, Popper and Kuhn, this article reveals the structural fragility of induction, especially at moments of discontinuity. And in business, it is precisely these moments that decide who wins and who disappears. In HR, AI risks excluding unconventional talents, generalists and outsiders — the very people capable of generating innovation. The paradox is that systems designed to optimise lead us to choose the same person every time.
What is needed is not a technophobic approach, but a critical one: AI is blind to the exception, even when it is precisely the exception that rewrites the rules. This is why leaders must be able to see what the algorithm cannot and create organisational spaces to accommodate it. Because AI cannot replace strategic vision — and only those who know how to use AI without being governed by it will be able to lead change.
Induction is the ability to derive general rules from a finite number of observations. It allows predictions in contexts known to be constant. A black swan, that is, an event unpredictable by induction which radically alters the rules of the system, demonstrates that induction does not always work, at least in economics, and when it does not, the issue is serious. Normal evolution does not produce radical changes because, to put it in Taleb’s words, what is normal is irrelevant. Indeed, induction rarely leads to relevant or revolutionary conclusions. On the contrary, through excessive confidence that the future will be a predictable projection of the past, induction prevents us from recognising black swans when we encounter them. In fact, it blinds us to everything outside the norm, even when we might need it.
No observation guarantees that the future will behave like the past. Any attempt to justify induction is recursive: if we say that induction has worked in the past, we are already using induction to defend induction. Hume made it quite clear that the habit of believing in the stability of things rests not on reason but on custom and psychological comfort.
Popper tried to disentangle science from induction, successfully proposing the method of conjectures and refutations. Since Popper, theories have not been verified but put to the test. They remain valid until they are refuted. But reality might change regardless, or our way of observing and describing it might change. Indeed, Goodman, using the argument of strange predicates, showed how important the observer is in describing reality. A strange predicate is a category consistent with all past observations, but which implies discontinuous and unexpected behaviour in the future.
Goodman hypothesises an emerald of the colour grue, which is green until a certain date and blue thereafter. Every emerald observed before that date is both green and grue. Past observation therefore supports both hypotheses. But we prefer green, because we are accustomed to it, whereas grue lies outside our psychological custom. Goodman argues that there is no logical justification for preferring one predicate to another when they are arbitrary.
And, reciprocally, this also applies to predicates that seem normal to us through custom, but postulate an arbitrary discontinuity. Take, for example, the ages of life: childhood, adolescence, youth, and so on through the labels. The same person changes definition according to the arbitrary criterion of successive birthdays. The person is not changing ontologically, but simply entering a new label. It is an arbitrary predicate that does not seem strange to us only because we have grown accustomed to it.
Induction, then, is a useful tool for formulating hypotheses, not for establishing truths or proofs. It seems to work, but it has a problem with discontinuities. To the point that it can lead us astray when we are undergoing a phase change rather than normal evolution.
Thomas Kuhn, in The Structure of Scientific Revolutions, his 1962 book, made it clear that science evolves not linearly, but through phases of normal science followed by revolutions that break with the dominant paradigm. The same pattern is found in other domains: Joseph Schumpeter interpreted capitalism as a dynamic of growth interrupted by waves of creative destruction, Carlota Perez described technological innovation as an alternation between diffusion and rupture, and Michel Foucault reconstructed the profound discontinuities in the history of ideas and institutions. In all these cases – and in countless others – organic growth is broken by radical leaps that reshape the realm of the possible.
And it is precisely these leaps that hit hardest those who fail to recognise them in time: those who expect everything to change at a normal pace. Those who fail to see the emergence of a paradigm shift suffer its impact in the form of decline, loss of relevance or exit from the market. Some striking examples are universally known: Symbian, BlackBerry OS, Kodak, MySpace, Yahoo!, Intel.
Ironically, Andy Grove, Intel’s long-serving CEO and co-founder, had tried to sow a culture of discontinuity within the company. In his Only the Paranoid Survive of 1996, he proposed a radically anti-inductive stance: distrust apparent stability and constantly prepare for the strategic inflection point, that moment when a technological, market or competitive change radically transforms the rules of the game. But things went into freefall for Intel precisely when it missed three inflection points: mobile, artificial intelligence and energy efficiency in chips.
We must recognise that people and organisations continually have to contend with the bias of induction. Kahneman, Tversky and Gigerenzer have thoroughly studied the fact that people judge the probability of an event according to how closely it resembles a familiar mental pattern, and that there is a law of small numbers according to which we tend to derive general rules from a few subjective data points.
Given the risks and the persistence of the inductive pattern, perhaps organisations would do well to train themselves to recognise it so as to avoid it. And it is precisely to this end that we should stress that artificial intelligence, too, is not exempt from the bias of induction: AI is a fundamentally inductive system. Predictive models learn from observed patterns. Although deep neural networks make induction flexible and sophisticated, they certainly cannot turn it into deduction or conceptual discovery. Indeed, when the context changes faster than their ability to readjust, they fail dangerously. But they can also fail because they read strange patterns without noticing their strangeness.
The discussion of bias, fairness, robustness, overfitting and generalisation in AI models is a contemporary reformulation of the same problem. It concerns, in particular, erroneous assumptions in learning processes, fairness of treatment built on predicates that are only apparently neutral, the inability to adapt to input data different from the training data, the difficulty of transferring predictive models from one context to another, and the attribution of structural value to elements that were intended to carry only relative weight in training.
AI acts on a partial representation of a reality assumed to be stable. The risk is that, in a context of unstable reality, AI may generate a useless and dangerous regression towards the mean, choosing what has already been chosen, and doing so at a moment of creative discontinuity. The effect is that, in a reality where black swans can exist, it advises the organisation to bet on tomorrow’s weather being like today’s.
In the field of innovation, this is an enormous danger. AI-based HR models – beyond the well-known gender or ethnicity biases – risk discarding precisely the people who can break the mould. Think of Steve Wozniak, the self-taught man who designed the Apple I, or Sara Blakely, the door-to-door saleswoman who invented Spanx. Or the unconventional engineer Nolan Bushnell, founder of Atari, who hired another unconventional figure, Steve Jobs, precisely because he was anomalous. Had algorithms decided who could gain access to a particular role, these profiles would have been excluded. AI would have optimised towards repetition, not towards the exception.
Karim Lakhani, in a study published in Harvard Business Review in 2007, with Kevin Boudreau, explained how the most innovative solutions often come from those working outside the domain in which the problem originated. An analysis of Eli Lilly’s InnoCentive platform – where complex technical problems are presented to a global community of solvers – demonstrates that the chances of success increase when the solver belongs to another disciplinary field. It is not specialist expertise, or a steady record of university examinations, that makes the difference, but the ability to identify the essential function of an element and relocate it in another, radically different context.
It is the same intuition that guides David Epstein when he defends the generalist. Those who have crossed different domains, who have failed while changing fields – at university or at work – who have played with many cognitive frameworks, are more likely to find a strange predicate that works just when it is needed. And they themselves are functional strange predicates. But generalists who enrolled in physics, then literature, then economics, and lost a year learning to raft, are not the favourites of AI agents dedicated to HR. And so we lose the opportunity to see what we were not looking for but which could have secured our success, or the success of a department within our business.
This is the blind spot of predictive innovation: AI would offer us the possibility of choosing the same person every time, incapable as it is of detecting what does not yet fit the pattern, the exception that changes the rules. And this directly concerns businesses. Because every so often, to survive and win, a discontinuity is needed, like a flash of insight: we are no longer in the train business, but in transport. Not flowers, but gifts. Not computers, but entertainment. Yet this requires a vision that goes beyond the data. We need to dismantle our confidence in inductive thinking.
Leaders must build critical scepticism towards artificial intelligence into the organisations they lead, creating spaces for what the algorithm is structurally unable to see. And clearly define what may have strategic implications, even indirect ones, so that it can be treated in a specific way.
To use Lakhani’s words once again: Artificial intelligence will not replace people. But people who know how to use it will replace those who do not. Recognising AI’s problem with discontinuities, and not allowing ourselves to be contaminated by it, is a fundamental requirement of knowing how to use it. This is why we cannot delegate to artificial intelligence the choices that directly affect the success of our business. Such as deciding which people will be part of it: every single choice about who joins a company is a strategic decision.
In the history of every business there are talents that AI would not have chosen who have shaped its future: this piece is implicitly dedicated to all those who have known how to put their anti-inductive creativity at the service of their organisations’ success. This piece is inspired by, and dedicated to, all these hidden champions. I would like to name some of them explicitly: Ken Kutaragi, an ordinary engineer whose personal research provided the starting point for the future PlayStation project; Art Fry and Spencer Silver, without whom 3M’s Post-it notes would never have existed; the technician Trevor Baylis, who invented the wind-up radio for sub-Saharan Africa, which played such an important role in the spread of information; and the assistant Rosalind Franklin, who took Photograph 51, so important in understanding the double-helix shape of DNA; as well as the Liquid Paper secretary Betty Nesmith Graham, who invented correction fluid; and Paul Buchheit, who conceived Gmail in his 20% time project and was the first at Google to say: Don’t be evil; and also Shigeichi Negishi of Hitachi, who invented the answering machine, and many, many others.
This article first appeared in HBR Italia in July 2025: https://www.hbritalia.it/homepage/2025/07/04/news/leccezione-che-cambia-le-regole-16310/