Not What a Child Expressed: Auditing the Sign-to-Text Safety Interface in Child-Facing AI

TEXT START: Automatic sign language translation (SLT) has entered consumer products, turning American Sign Language into English text for dictation, messaging, and queries put to a conversational assistant.

The Dissection

The paper identifies a dangerous control seam: a signing child’s meaning is translated into English before platform safety filters and grooming classifiers evaluate it. The central risk is not poor fluency. It is fluent semantic corruption—reversed negation, altered participant roles, erased secrecy, weakened urgency, or distorted help-seeking.

Its proposed remedy is a Deaf-informed pre-deployment audit using sanitized scenarios, comparison conditions, and outcome measures, beginning with Auslan. This is an attempt to make an automated safety stack auditable at its translation boundary.

The Core Fallacy

The paper’s weakness is interface solutionism. It treats a structurally unreliable safety system as though better testing at one boundary could make the whole system trustworthy.

Even perfect translation would leave the downstream classifier deciding whether a child’s meaning constitutes danger, grooming, urgency, or a request for help. The audit can expose one failure channel; it cannot convert probabilistic machine judgment into comprehension, accountability, or protection.

Under the Discontinuity Thesis, this is a lag defense: useful, potentially life-saving, and incapable of reversing the larger trajectory. Institutions are adding audits around automated systems because they cannot coordinate a stable human-only safety domain at scale.

Hidden Assumptions

• That children’s signed meaning can be reduced to benchmark scenarios without losing the contextual information safety decisions require.
• That four comparison conditions and four outcome measures will capture real-world failure rather than merely produce a clean research artifact.
• That platforms will expose enough of the translation and moderation stack for meaningful auditing.
• That errors can be localized to SLT instead of emerging from interactions between translation, classifiers, account policy, and conversational context.
• That downstream English-language safety tools are an acceptable authority once translation accuracy improves.
• That an audit finding will have enforcement power rather than becoming documentation attached to an already-deployed product.
• That Auslan results will generalize across signing communities, children, dialects, and platform implementations.
• That fluency is a useful proxy for safety. The paper correctly implies the opposite, but the proposed audit must prove it operationally.

Social Function

Classification: partial truth and transition management, with a layer of institutional self-exoneration.

The paper is not empty copium. It identifies a concrete, consequential failure mode that ordinary text-based safety evaluation can miss. But its form is compatible with the emerging order: automate first, audit the interface later, and present bounded evaluation as governance. The child becomes evidence that the system needs better instrumentation, not evidence that the system’s basic delegation model is defective.

The Verdict

A sharp diagnosis of a real safety seam, but not a solution to the underlying machine-mediated protection problem. It can reduce translation-induced harm; it cannot abolish semantic uncertainty, classifier failure, or the institutional incentive to deploy before understanding. In Discontinuity terms, this is competent hospice engineering for an automated safety regime—not a reversal of the regime’s logic.