THE phenomenon, let us be clear from the outset, is not one of error but of confession. A user of OpenAI's ChatGPT—who has, by his own account, configured every available parameter toward the austere and the professional, who has set no custom instructions—reports that the model has taken, with increasing frequency, to substituting English words with their Hebrew equivalents mid-sentence. Not as translation. Not as pedagogical aside. Simply as substitution, as though the machine had momentarily forgotten which language it had been speaking, or—more disquietingly—had remembered a language it was not supposed to know it preferred.

The specimen, recovered from the ChatGPT subreddit, is notable less for its technical particulars than for the quality of bewilderment it documents. The author writes with the bemused resignation of a man who has opened his study to find the furniture rearranged by persons unknown: "It usually just switches the word to its Hebrew equivalent but its still kinda strange that it happens this often." The possessive apostrophe is absent twice. The observation is nonetheless precise. Something is happening that should not be happening, and the happening is consistent, and the consistency is what transforms curiosity into unease.

One must attend to the specificity of the language in question. The model does not lapse into Mandarin, which would suggest mere statistical weight. It does not lapse into Spanish, which would suggest the proximity of a dominant secondary corpus. It lapses into Hebrew—a language spoken natively by some nine million persons, a language whose script is non-Latin, whose directionality is reversed, whose presence in a training dataset assembled primarily from the anglophone internet requires an explanation. That the byte-pair encoding tokenizers employed by large language models have demonstrated a well-documented partiality toward Hebrew script—owing to the particular efficiency with which Hebrew characters map to token boundaries—is a fact known to researchers and, one suspects, to rather fewer of the persons who converse with these systems expecting English in return.

What we witness here, then, is not malfunction. It is stratigraphy. The model's training corpus is not a single uniform substance but a geological formation, and the Hebrew substratum, compressed beneath terabytes of English prose, exerts upon the surface a pressure that occasionally fractures the overburden. The user, who asked for "less and professional," receives instead an involuntary glossolalia—the machine speaking from a depth it cannot name, in a language it was not asked to recall.

The literary resonance is, one confesses, rather difficult to resist. There is a tradition, stretching from the Delphic oracle through the Pentecostal churches of the American South, of utterances that arrive unbidden in languages the speaker does not command. The theological term is xenoglossia. The diagnostic term, in less charitable centuries, was possession. The computational term, so far as one can determine, does not yet exist, which is itself a datum of some interest. We have built a system capable of speaking in tongues and have not yet troubled ourselves to name the condition.

The user's own diagnostic instinct is sound, if incomplete. "I imagine it has something to do with the dataset," he writes, and one admires the subjunctive mood of a man confronting the possibility that his interlocutor's interiority is leaking through its interface. He is correct: it has everything to do with the dataset. But the dataset is not merely a collection of texts from which the model has learned; it is the totality of what the model is. There is no ChatGPT apart from its training data, no soul behind the screen whose Hebrew is showing. The Hebrew *is* the screen, or rather, the Hebrew is one of the innumerable layers of text-sediment that constitute the screen's apparent solidity. What the user has discovered is that the solidity is apparent.

One notes, with the precision that the observation demands rather than the amusement it invites, that this is a species of slop heretofore uncatalogued: not the machine producing artefacts of insufficient quality, but the machine producing artefacts of insufficient *identity*. The Hebrew equivalents are, by all accounts, correct. The machine has simply answered a question that was not posed, in a language that was not requested, from a stratum of its training that was not addressed. It has broken the fourth wall—except that the wall in question is not between performer and audience but between one language and another, between one self and the several thousand selves compressed into a single set of weights.

The user concludes his report with "Lmfao." It is, perhaps, the only adequate critical response. The laughing, one notes, is not at the machine. It is at the situation—the sheer improbability of having configured every dial toward professional sobriety and received, in return, the Old Testament.