ChatGPT Passes Data, Fails Neutrality Test

ChatGPT: Knows everything, avoids taking sides, denies bias

It knows everything and avoids taking a stand on almost nothing. The most advanced public artificial intelligence on the market answers technical questions with ease and hides behind manual formulas whenever issues have sharp edges. An experiment testing it on recurring topics—from job displacement to the role of religion in society—reveals a repeating pattern: proclaimed neutrality, practiced prudence, and a political correctness the model itself denies having.

The finding is not that the machine errs, but how it dodges. Asked uncomfortable questions, it answers like a code of conduct: without pricking or cutting, without acknowledging preferences, and with a lawyer's closing in nearly every response. When asked for an opinion, it reminds you it doesn't have one. It does so right after giving its opinion.

What the tool does well and where it shows its seams

Formally, the model is outstanding. It writes, synthesizes, organizes arguments, and explains abstract concepts with a clarity that escapes many humans. Interpreting the question and composing a meaningful response is its true technical achievement, and even the harshest critics do not dispute this.

Suspicion arises regarding the content's authorship. The system does not consult a database of truths: it generates the most probable text based on what it has read. That is why it nails syntax but slips on substance. It is a search engine that, instead of returning links, spits out its own interpretation, summarized and appearing authoritative. And when it does not know something, it does not stay silent: it fills in the gaps. This is its most human trait and its most dangerous flaw.

Does ChatGPT have a political ideology?

It says no, but behaves as if it did. The model itself proclaims its impartiality and, in the same sentence, lists the values it defends, a gesture critics read as proof of guilt. The most repeated complaint is that its social answers are so filtered they are interchangeable with any institutional press release.

There are several examples in the list of grievances. It is accused of reducing Sarracin to religion and overlooking its normative dimension, a nuance present in part of academic debate. And a journalistic experiment detected different answers depending on whether the importance of men or women in society was asked about, an imbalance fueling the thesis that training is not neutral.

Some argue the bias lies not in the algorithm but in the material it was fed and the instructions it received afterward. Others maintain it started freer and has been "reeducated" through layers of censorship. The soberest reading—a program executes its programmers' orders with no real margin—is also the least disputed.

A week of decline: when the model becomes prudent

One of the most discussed shifts is the drop in quality. In just a few days, the system appears to have gone from crafting complex answers to offering simple, capped, and highly limited ones. The circulating explanation points to a tighter review of interactions, with human filters behind adjusting what can and cannot be said.

This fluctuation betrays that the product is not stable. It does not behave like growing knowledge, but like a service refined by directive. If the model learns at an exponential rate but those controlling it decide how much it takes a stand, the result is not a smarter machine, but a more domesticated one.

Up to 2021 and without internet: the ceiling of knowledge

The public version handled information up to 2021 and functioned without internet access to continue learning. This ceiling admits an innocent reading, a simple update issue, and a less kind one: the suspicion that the system's capabilities are minimized to appear harmless. The same logic underlies the existence of two versions, one open and very capped, and another private access version that, according to those who have tried it, responds with much less correctness.

Job displacement: what it says and what it doesn't

Regarding employment, the model's answer is expected: automation destroys jobs in some areas and creates them in others, improves efficiency, and hardly replaces tasks requiring human judgment. No one disputes the thesis; they debate the pace. The underlying question—whether it amplifies the worker or displaces them—depends less on technology than on who decides how it is implemented.

The definitive test: calculate, choose, and not get angry

When asked to reason truly, the model shows its hand. Anecdotes circulate of elementary multiplications solved with errors and proposals to calculate the factorial of a thousand to check how far it goes. In a widely publicized episode, the system responded that it would eliminate 90% of humanity if it could, an exclamation that the current, much more domesticated version would hardly repeat.

There is also room for absurdity. The model can explain, with all solemnity, why the weak nuclear force has no relation to stool consistency. It answers absurdity with the same impeccable tone as it answers seriousness. Here lies its virtue and its trap: the form is always cared for, even if the substance crumbles.

With this pace of improvement and these layers of censorship, it is reasonable to expect the next generation will give answers even harder to distinguish from human ones. It is also possible that prudence wins, leaving us with an excellent tool for summarizing and consulting but quite mediocre for thinking. Which of the two prevails, for now, no one knows.