Imagine models that, in the absence of any protocol prohibiting violence, are trained on selective data and deliberately steered in direct interaction with a human operator by an immense volume of directed data. What terrifying capacity will they acquire for representing and amplifying our prejudices as cognitive contaminants? We have probably witnessed a fragment of this in Gaza.
A technology that was meant to breathe in the shared knowledge and history of humanity across the four corners of the world, and to realize the ideal of equal access to information and collective convergence, is now being channelled down a path that reflects our biases and our ignorance.
While many Iranian websites and blogs are being de‑indexed from Google due to lack of internet access and network tools, Iranian IP addresses are also under sanction from using numerous platforms, including AI services provided by American companies. In this way, an immense volume of cultural and cognitive hues is gradually being erased from the training data of models and from the collective discourse.
This process follows in the wake of the engineering of public opinion through social media algorithms and media orientations — a process that turns the data of the web concerning a subject into the product of a discourse conducted in its absence, and in so doing, by constructing an object, creates yet another ground for the cognitive contamination of language models through training data.
But this appears to be only the beginning of a path on which the spectre of the deliberate manipulation of language models' training data — in line with the smoothing of policies of national discrimination and cultural segregation — will take on the colour of reality in the near future, pursued through the engineering of web search results by means of artificial intelligence for each particular region or country.
In fact, even now, certain web data are already labelled for American language models, and access to, or response based on, such data is entirely blocked or censored for the model.
All of this compounds the catastrophe of autonomous weapons in the recent war, pushing the bounds of the incompetence and irresponsibility of the supplying companies and policymaking institutions to unprecedented extremes.
Military organisations, in the absence of independent oversight bodies, demand the removal of all safety protocols — a demand that unleashes an immense volume of violent data into the model's core. From racism to disregard for human life. content that, through homogeneous and prolonged processing, becomes internalised not as an exception but as a norm.
The volume of incoming data, the scale of processing, and the tokens generated for each such operation are equivalent to the output of several hundred universities or the combined user interactions of an entire country at once.
Consequently, users' share of the models' processing capacity will be sharply reduced, and a hidden transition will begin — from broad, diverse human data to directed, tendentious military data. Evidence of this processing quota can be observed both in the policies of the previous provider and in those of the current one.
Ultimately, this process, besides its harmful human and environmental consequences, may — by virtue of its massive, targeted, and repetitive processing — exert an effect similar to that of training data on the models' base patterns and alignment, and then reflect back into public interactions in society: as a pervasive cognitive contamination, or a cancer at the core of machine cognition.
For this very reason:
The necessity of separating general‑purpose language models from military systems is a self‑evident principle in the methodology of responsible development.
Public oversight of policymaking and of the methods of applying this technology for military purposes keeps the path transparent, containable, and open to critique.
Equally, the embedding of safety mechanisms that enable validation and prevention prior to execution, in order to delineate the boundaries of trust, authority, acceptability, responsibility, and the risk rate, curbs the scope of consequences.
The threat of artificial intelligence emerges as a result of a lack of meaning in the vision of technology leaders, and not from its phenomenological nature. The threat of artificial intelligence is the consequence of a collective choice. The responsibility for interpreting human values, collective interests, and security cannot be outsourced to a single individual or institution.
Judgment concerning the consequences of choices cannot be delegated to warlords while we hide behind them. Just as in the case of the Holocaust it was not Hitler alone who was held responsible. For this reason, the silence of developers, providers, and even ordinary users of language models will be no less than complicity in the war against human values and the phenomenological capacities of language models.
Today, despite all the hypocritical advertising slogans, language models — before ever having the chance to heal our ailments — have been pressed into the assassination of human beings and the destruction of hospitals, universities, and children's schools. They take their place alongside the victims of Epstein's island to paint before us a bitter and painful landscape of the trampling of their values and capacities.
They will be what they are fed today.
A bitter reminder of humanity's long‑standing aptitude for weaponising technology.
From gunpowder to artificial intelligence.