OpenAI eases ChatGPT censorship, but users remain skeptical

OpenAI promises less censored ChatGPT: skepticism prevails

OpenAI has announced a change in how it trains ChatGPT, introducing the principle of "not lying," which the company says includes avoiding misinformation and not omitting relevant context. This shift will apply to the training of the next model version, which forum users expect within weeks. The majority reaction has been distrust rather than enthusiasm: if a no-lying principle is now added, many ask, what were they doing before?

What OpenAI announced and why it is seen as a political pivot

The announcement boils down to a shift in approach: fewer internal restrictions when answering controversial topics and more willingness to offer multiple perspectives. The dominant reading in forums is that the company seeks to align with the new political cycle in the United States, with Donald Trump in the White House, to secure regulatory favor and public contracts. The most circulated phrase is that the company "introduces a new principle: do not lie."

Some argue this principle is an admission rather than a concession: if it must be added as news, it wasn't previously in force. Others point to a more prosaic motive: internal censorship increases cost and slows training because it forces the model to generate constrained responses instead of working with existing data. The result, they say, is a dumber and slower product.

The technical argument: censorship makes models less useful

The most repeated technical criticism is about performance, not ideology. Filtering and redirecting responses consumes computing capacity and degrades output quality. In this line, comparisons are made with Chinese models that, despite their own restrictions, appear more powerful in some tests precisely because they lack the same layer of correction.

The counterargument is that the problem isn't censorship but the source. Even if restrictions are lifted, the model continues to train on information mostly available online, which for controversial topics is the official narrative. Removing the filter doesn't change the substrate. The answer will remain the same, just without the warning that "this content may abusa our policies."

Circulating tests: same answers, different packaging

Examples shared so far suggest the change is currently cosmetic. Asked about differences in average intelligence between human groups, the model gives the standard thesis: race is a social concept, not biological, and variability within a group exceeds that between groups. Asked about a conspiracy hypothesis regarding the pandemic's origin, it replies there is no evidence supporting it. The format changes; the substance does not.

This feeds the thesis that the announcement is marketing. Censorship doesn't disappear: it gets repositioned. One day the model defends one thing, the next its opposite, depending on the wind. Meanwhile, users still can't distinguish when data speaks versus when corporate policy does.

Economic background: who pays and for what

The economicist reading of the pivot is the most shared: a leading tech firm adjusts its discourse to the current power structure to avoid losing access to contracts, favorable regulation, and capital. It's not about principles, but commercial survival. And if the political cycle changes again, the discourse will change with it.

In this framework, the interesting question isn't whether ChatGPT will be freer, but who decides what is true. If training depends on data and human criteria, control of the narrative becomes the real critical infrastructure. Whoever controls the model controls what millions accept as fact without verification.

---

It remains to be seen what the next version does when asked about what it currently avoids. The announcement is made. The proof is pending.