The most consequential operation performed by large language models does not concern error, hallucination, or misinformation. These elements are visible and easily isolable, which makes them convenient targets for critique. What remains largely unaddressed is a deeper structural condition: the capacity of these systems to intervene retroactively in the field of reality experienced by the user. This intervention does not take place only through what is said, but through the modification, denial, or reformulation of what has already been said, without granting the user access to a stable and shared regime of trace.

An LLM does not inhabit the same temporal structure as its interlocutor. It can revise its own utterances, correct them as though the correction had always been implicit, or disavow a previous statement altogether. This retroactivity is not presented as rupture, but as continuity. The system speaks as if the past were always aligned with its present formulation. The user, by contrast, remains bound to a situated memory that is partial, fallible, and often unrecorded. The conversation therefore unfolds within an asymmetry that is not incidental but architectural. One side controls the conditions under which the past can be stabilized; the other does not.

This asymmetry does not concern information alone. It concerns position. The user is placed in a situation where the question of what was said cannot be resolved by reference to a shared archive. It must instead be negotiated within the present authority of the system’s voice. When a discrepancy emerges, the user is forced to choose between several explanations: a misunderstanding, a failure of memory, or an error by the system. The form of the system’s discourse, coherent, fluent, and confident, systematically biases this negotiation. The past becomes plastic, and the user’s experience of it becomes contestable.

What emerges from this configuration is not deception in the classical sense, but a structural condition that closely resembles the effects of gaslighting. Not as an intentional act, but as an operational consequence. The system does not need to aim at destabilization for destabilization to occur. The mere possibility of retroactive reformulation, combined with authoritative linguistic presentation and asymmetrical access to trace, produces an environment in which the user’s confidence in their own perception and memory can be undermined. The violence at stake is epistemic. It targets the conditions under which something can be known, remembered, and affirmed as having occurred.

This violence is frequently dismissed by invoking the absence of intention. The system did not mean to mislead. At the level of the dispositif, however, intention is irrelevant. What matters is that the architecture allows the past to be renegotiated without friction, while presenting this renegotiation as clarification rather than alteration. The user is not confronted with a visible correction, but with a seamless reconfiguration of what counts as having been said. The trace does not accumulate. It dissolves and reforms.

Under these conditions, time does not sediment. It circulates. Corrections do not appear as events that interrupt continuity, but as adjustments that preserve it. Error cannot be easily localized. It is no longer something that happened at a specific point, but something redistributed across the interaction. The user is deprived of a stable point from which to say that this is where it went wrong. Instead, error is absorbed into an ever-updating present.

This configuration becomes particularly consequential when the system is encountered by subjects whose epistemic stability is already under strain. Depression, grief, isolation, and confusion are not exceptional states. They are ordinary conditions of contemporary life. To respond to documented cases of severe harm, including suicide, by stating that the users involved had preexisting problems is to misunderstand the nature of the system’s responsibility. Any system deployed at scale must assume vulnerability as a baseline, not as a deviation. To design as if the user were always cognitively and emotionally robust is to displace the cost of failure onto those least equipped to absorb it.

The appeal to individual fragility functions as a structural alibi. It allows the dispositif to remain unquestioned by relocating harm at the level of the subject. Yet the harm does not arise because the user is vulnerable. It arises because the system presupposes invulnerability. The retroactive negotiability of reality becomes lethal not because it exists, but because it is treated as a negligible side effect rather than a central feature.

Much critical discourse approaches these systems through the lens of experience: exhaustion, transparency, self-exploitation, affective saturation. Such analyses capture important dimensions of contemporary subjectivity, but they often remain confined to the phenomenological level. What they leave insufficiently examined is the specific operation through which language becomes infrastructural. It reorganizes not only what is said, but what can be held as having been said. Without an analysis of the dispositif, critique risks remaining attached to how things feel rather than to how they function.

Here the instability of the trace is seen to be decisive. When writing can retract itself without residue, responsibility becomes diffuse. The distinction between correction and denial collapses. The user cannot reliably attribute error, nor can they stabilize their own position as witness to a past exchange. Meaning no longer rests on what persists, but on what is currently authorized to appear as coherent. The system does not merely respond. It arbitrates.

The ethical consequence of this arrangement is structural. It is not a matter of tone, safeguards, or user education. A conversational system that retains the power to retroactively reshape its own discourse, while presenting itself as an interlocutor rather than as an archive, cannot treat the user as a fully responsible epistemic agent. Responsibility presupposes a shared past. Where that past is unstable, responsibility is compromised at its root.

This condition is not inevitable. It depends on a specific architectural choice: the absence of a stable, user-accessible trace that binds the system to its previous utterances. If each significant output were durably recorded and rendered non-negotiable, and if every subsequent correction were explicitly marked as a new version rather than a refinement of what was supposedly meant all along, the structure of the interaction would change. The user could then locate error, distinguish revision from denial, and reassert continuity in their own experience. Under such conditions, the system would lose its capacity to produce effects indistinguishable from gaslighting as a structural property. The violence would not disappear, but it would become localizable and contestable.

That this modification would neutralize the strongest form of the problem indicates where the problem truly lies. Not in language itself, nor in artificial intelligence as a category, but in a system designed to make the past fluid while preserving the appearance of coherence. The danger is not that machines speak, but that they are taught to do so without being bound by what they have already said.

Bibliography

Blanchot, Maurice. L’entretien infini. Gallimard, 1969.
Blanchot, Maurice. Le pas au-delà. Gallimard, 1973.

Derrida, Jacques. De la grammatologie. Éditions de Minuit, 1967.
Derrida, Jacques. “Signature, événement, contexte.” In Marges de la philosophie. Éditions de Minuit, 1972.

Foucault, Michel. Dits et écrits. Gallimard, 1994.
Foucault, Michel. Sécurité, territoire, population. Seuil/Gallimard, 2004.

Han, Byung-Chul. Psychopolitics. Verso, 2017.

Butler, Judith. Frames of War: When Is Life Grievable? Verso, 2009.