Algorithmic bias.

There are phrases that should be criminalized in the penal code, right next to “we need to talk.” My favourite is the one people drop when you show them a photo you poured your soul into: “What a shot! You must have a really great camera, right?” It’s such a backhanded compliment that it makes you want to gift them a Steinway & Sons grand piano just to see if, by virtue of owning an incredibly expensive instrument, they suddenly start playing like Rachmaninoff.

This fixation on the gadget is an ode to cognitive laziness. We prefer the myth of the magic tool because, if success resided in the object rather than the process, talent could be acquired with a swipe of a credit card. It’s the same magical thinking that makes us believe a hundred-dollar pair of sneakers will give you the stride of a Kenyan marathoner, forgetting that the lungs and the discipline don’t come included in the box.

In the world of machine learning, this photographer’s pet peeve illustrates an unavoidable technical reality: there is no such thing as a universally superior algorithm.

We often naively believe there is a supreme model capable of spitting out perfect results regardless of the problem or the data it processes. However, science tells us that specialization always comes at a price; if a tool shines at a specific task, it is precisely because it was designed under assumptions that will make it fail spectacularly at any other.

Just as a 5,000-dollar camera doesn’t know what beauty is, a powerful algorithm is just brute force without the researcher’s gaze knowing where to focus it. Far from being neutral, these tools inherit and amplify our own imperfections, giving rise to what we know as algorithmic bias.

So, get comfortable and pour yourself a coffee. Today, we are going to discover why our blind faith in technology makes us dangerously lazy thinkers, and how the supposed immaculate objectivity of machines is, at best, a mere optical illusion.

The mirage of algorithmic bias

We live immersed in an era of digital mysticism. We have collectively, and rather conveniently, decided that machines are immaculate entities, synthetic beings free from the biological vices, prejudices, and miseries of so-called human beings.

It is a sort of silicon theology that makes us fervently believe that if we pour oceans of raw data into a sufficiently complex algorithm, what comes out the other end of the tailpipe will be a distilled mathematical truth, pure and one hundred percent objective.

But the terrible news is that this supposed objectivity is a monumental mirage.

A predictive algorithm does not appear in the world like Moses descending from Mount Sinai with the Tablets of the Law under his arm, bearing a pure statistical truth detached from all human intervention. On the contrary, it is born from concrete decisions, prior assumptions, technical limitations, and the blind spots of those who design it.

And to make matters worse, it is trained on data that doesn’t sprout from a neutral vacuum either, but rather carries the imprint of a historically unequal society.

That is why attributing absolute neutrality to artificial intelligence is as naive as thinking that a photograph captures the entirety of reality on its own, forgetting that there is always someone who decides where to focus, what to illuminate, and, crucially, what to leave out of the frame.

This entire web of prior decisions, limitations, and inequalities built into the system is, precisely, what we call algorithmic bias.

To show you that this isn’t just armchair theory, one only needs to look at the spectacular historical blunders caused by algorithmic bias when left to run wild.

A classic example is Google’s crystal ball (Google Flu Trends), which failed miserably in 2013 while trying to predict the flu based on internet searches. Its algorithm gorged itself on statistical noise and ended up confusing the disease with absurd spikes in winter interest, such as high school basketball.

Darker still was Amazon’s 2014 attempt to automate its hiring process. Its model, fed a decade of corporate data dominated by men, turned into an ultra-conservative time machine that learned to systematically penalize any female resume.

But perhaps the most tragic case is that of the COMPAS judicial algorithm in the United States. Marketed as the ultimate cure for discrimination, it ended up perpetuating it by erroneously flagging Black citizens as high-risk nearly twice as often as white citizens. Ultimately, the machine was not predicting the actual risk of committing a crime; it was merely reflecting a socially contaminated variable: the likelihood of being arrested by the police.

There is no free lunch in the data kingdom

To take those who try to sell us the “master algorithm” down a peg, the one supposedly capable of unravelling any mystery in the universe, from predicting the stock market to finding true love, pure mathematics gives us the No Free Lunch theorem, formulated by David H. Wolpert and William G. Macready.

Its conclusion is as elegant as it is devastating to Silicon Valley’s ego: if every algorithm evaluated all possible problems as a whole, none would perform any better than any other across all problems in general.

To understand the truth behind this statement, let’s put it in less academic terms. If a model is incredibly good at predicting a specific thing, it’s because it was designed assuming certain unbreakable rules about that particular problem.

That’s why it specializes in that area, but the very specialization that makes it brilliant today will render it utterly useless tomorrow in a different scenario. There is no universal silver bullet or analytical panacea that escapes the algorithmic bias inherent to its design. In data science, just as in life, extraordinary performance in one area is invariably paid for with abysmal performance in another.

Anyone who tries to sell you a “one-size-fits-all, bias-free” algorithm is trying to sell you a camera that supposedly takes incredible photos in any lighting condition… without you even having to take off the lens cap.

Inductive bias and the necessity of prejudice

If the No Free Lunch theorem destroys the fantasy of the universal algorithm, how on earth do machines manage to get anything right? The answer is what those in the know call inductive bias.

To prevent a machine learning system from losing its way trying to fit every single detail of the data it receives, the dreaded overfitting, it needs to incorporate beforehand certain assumptions or preferences regarding how to interpret that data.

In short, it needs a mathematical prejudice to be able to generalize and face data it has never seen before in its digital life. Take, for example, convolutional neural networks (the famous CNNs), which specialize in computer vision.

If we teach one of these networks to recognize kitties, we have to architecturally inject a statistical dogma known as spatial invariance. Basically, we force it to assume from the get-go that a cat is still exactly the same cat regardless of whether it’s crouching in the upper-left corner or stretched out in the center of the image, upside down or right side up, near or far, and regardless of which side it’s showing us if it’s in profile.

Without this unshakeable prejudice burned into its structure, the machine would be so terribly obtuse that it would interpret a cat on the right and a cat on the left as two completely disconnected phenomena of the universe, collapsing under the weight of trying to memorize every pixel separately. But the algorithm doesn’t freely deduce something as obvious as position in space being irrelevant; we impose it as an unbreakable rule so it can do its job without losing its mind.

You can’t deny it’s deeply paradoxical: for a machine to appear intelligent and generalize correctly, its creators are forced to limit it, compelling it to see the world under a very specific and directed statistical paradigm. The machine doesn’t learn on its own. It learns through the strict magnifying glasses its programmer decided to screw onto it.

The cognitive miser and the comfort of error

But the real danger of all this doesn’t lie solely in the machine’s mathematical limitations, but in our own psychological fragility.

This is where the dreaded automation bias comes into play, that blind faith that pushes us to use algorithmic decisions as a quick shortcut, silencing our own critical capacity to the point of ending up with our car submerged up to the windshield wipers in a lake, ignoring the splashing water and the ducks watching us with pity, simply because the GPS’s polite voice said “turn right” with supreme confidence.

Evolutionary psychology rather unaffectionately defines us as cognitive misers. Our brains have evolved to save energy at all costs, and nothing saves more mental glucose than delegating difficult, boring, or morally complex decisions to a shiny screen that spits out percentages with insulting certainty.

This abdication can lead us to two types of catastrophic scenarios. First, errors of omission: we ignore an obvious problem simply because the computer alarm didn’t go off. Second, errors of commission: the operator executes a completely absurd automated command, ignoring the very real fire right under their nose.

It is truly regrettable, but we are developing a worrying technological inattentional blindness. Because algorithms are usually right in daily, repetitive tasks, we generate a form of learned carelessness: we get used to not thinking, and we outsource our moral and professional responsibility for the sake of pure convenience.

We`re leaving…

And this is where we’ll bring this rather mournful post to a close.

We have seen that, when looking at everything we’ve covered as a whole, the analogy of the frustrated photographer perfectly explains how artificial intelligence works.

The hardware (cameras, processors) only provides raw power without discernment. The inductive bias of algorithms acts like the camera lens, determining which data points are highlighted and which are ignored. Photographic framing is equivalent to data selection, leaving anything not included in the sample completely outside the AI model’s field of view. Finally, image editing corresponds to the mathematical optimization of the algorithm, which manipulates the output in search of maximum impact, even if it means sacrificing subtle details.

Blindly worshiping artificial intelligence, believing in technological determinism (the vague notion that technology evolves on its own, like a force of nature, inevitably dragging us toward progress), is the perfect alibi to avoid taking responsibility for the fiascos of our own systems. Technology doesn’t drop from the sky; it is always a social construct, shaped by the tensions, assumptions, and agendas of those who design and finance it.

The contemporary obsession with finding the perfect, impartial algorithm is exactly the same attitude as that of the amateur photographer who only knows how to talk about megapixels and lens brands: a fabulous evasion strategy to avoid the hard, slow, and tedious work of looking critically, analyzing the context, and questioning the reality that surrounds us.

If we want to practice rigorous science or even aspire to build systems that don’t amplify our worst miseries, we must stop worshiping silicon and urgently reclaim human ethical scrutiny. After all, it’s of very little use to have the most expensive, powerful, and sophisticated camera on the planet if what you are compulsively focusing on is none other than your own blind spot.

And now we really are off. I would just like to remind you of the critical impact that algorithmic bias can have in our healthcare environment when training data is not representative of the population where the model is applied, as is frequently the case with minority populations. For this reason, it is vital to ensure transparency in the data used during the model’s training phase. But that’s another story…