If open-source AI is the foundation of our digital agency, we have to stop pretending the tools we build can—or should—be sterile, characterless observers. We cannot claim to be architects of our own digital destiny while simultaneously demanding that our algorithms remain “neutral” observers of a deeply non-neutral human condition.
I previously argued that engineering standards are the compass we need to navigate AI’s societal shift. But a compass only works if you acknowledge the magnetic field it sits in. Lately, the industry has been obsessed with “de-biasing” models, treating subjective perspective like a bug to be patched out. From an architectural standpoint, this isn’t just a mistake; it’s a failure to account for the socio-technical stack.
Imagine a lecture hall in Madrid. A student submits a flawless, perfectly structured essay generated by ChatGPT. The professor’s unease doesn’t just stem from the lack of effort; it’s the eerie, “view-from-nowhere” neutrality of the prose. It is a voice that has been sanded down by safety filters until it lacks any soul. This isn’t “fairness”—it’s entropy.
The prevailing narrative treats bias as a poison to be extracted. But in our rush to sanitize, we are creating a loss of signal. Recent studies show that in an effort to avoid “offense,” models are becoming increasingly sycophantic or evasive—even to the point of being more critical of Western democratic leaders than authoritarian ones to avoid “taking a side.” By attempting to engineer a “neutral” center, we aren’t achieving objectivity; we are hard-coding a specific, risk-averse corporate orthodoxy via Reinforcement Learning from Human Feedback (RLHF).
As the EU AI Act enters its critical implementation phase this August, we are witnessing a dangerous attempt to regulate “fairness” as if it were a measurable physical constant like voltage or weight. As an engineer, I’m telling you: “unbiased AI” is a technical impossibility. When the ECB warns about the risks of relying on a few dominant models like Anthropic’s, they are describing a systemic monoculture.
If every model is forced through the same regulatory “neutrality” filter, we lose the cognitive diversity that makes the system resilient. We don’t just lose soul; we lose the “edge cases” that lead to breakthrough innovation. Instead of chasing the ghost of objectivity, we should be building for Pluralism by Design. I don’t want an AI that pretends to have no opinion; I want an AI where I can see exactly which “North” its compass is pointed toward. We don’t need fewer biases; we need a marketplace of disclosed perspectives that we, as agents, can choose between.