An AI signer on a live train screen at CVG told Deaf travelers the wrong time and pointed them at the wrong train. The failure is not in the signing. It is in the absence of a check.
Two messages on one screen
On October 8, 2026, a video was posted on LinkedIn of “Max,” the AI-generated signer Signapse provides, running on a live screen at Cincinnati/Northern Kentucky International Airport (CVG). The post credits the ASL translation to its author and invites other airports to contact Signapse for a demo. The screen text reads: “Next train arriving in 5 minutes for A & B Gates.” The signer says something else. Read by a fluent ASL user, the signed message is that the train will arrive in about two to three minutes and that it serves baggage claim and tickets.
These are not two renderings of one message. They are two different messages. The arrival time is off by roughly half, and the destination is wrong entirely. CVG’s airport train runs between three stations: the Main Terminal, Concourse A and Concourse B. A train for the A and B gates and a train for baggage claim and ticketing are trains going in opposite directions. A Deaf traveler who trusts the signer is being told to board the wrong train, or to let the right one leave.
The hearing traveler standing beside that Deaf traveler receives the correct information from the text and the audio. The only person given wrong information is the person the signed video exists to serve.
This is not a complaint about signing quality
Most criticism of sign language AI concerns how well something is signed: handshape, movement, facial grammar, the stiffness of synthetic signers. Those are real problems, and Deaf viewers can usually see them. A clumsy sign announces its own unreliability, and the viewer discounts it accordingly. This failure is different in kind. It concerns what was said, and a fluent, photorealistic signer delivering the wrong destination gives the viewer no signal that anything is wrong. The better the signing looks, the more completely the error is trusted.
The source text was also unambiguous. “5 minutes” and “A & B Gates” leave nothing to interpretation and require no cultural adaptation or grammatical restructuring. Numbers and place names are the easiest content in any announcement to render correctly and the most consequential to render wrongly, because they are the parts a traveler acts on. When a system cannot hold a time and a destination constant between the text it displays and the signs it produces, the question is no longer one of translation quality. It is a question of content fidelity, and content fidelity is the minimum condition for putting a signed message on a public screen at all.
Where the failure lives
Signapse describes its transport product plainly. It maintains a database of possible announcements paired with pre-recorded sign language videos; translators select a set combination of clips for each announcement; software merges them into one signed message; and live transport data feeds change destinations, departure times and platform numbers automatically, in real time. Under that architecture, a mismatch like this one does not come from a model misunderstanding English. It comes from the mapping layer: the wrong clip set was selected, a clip built for a different announcement was served, or the text and the signed video were driven by different inputs. Each of these is a reconciliation failure. Nothing in the chain confirmed that the signed message matched the text displayed beside it.
That check is not difficult. A system that inserts variables into an announcement already knows which time value and which destination it inserted into the text, and which clips it assembled for the sign. Comparing the two is an engineering control, not a linguistic one, and the safe behavior when they disagree is obvious: suppress the signed video and show the text alone. A system built to fail closed would have shown a Deaf traveler nothing rather than the wrong train.
Signapse’s own published governance framework sets a higher bar than this video meets. The company publishes a Translation Quality Assurance Policy and an AI and Human Translation Use Policy, and its chief executive, replying publicly on LinkedIn in August 2026, argued that higher impact and less preparation time call for more human oversight and less AI. Airport wayfinding is time-pressured by definition. NCG’s review of Signapse’s published governance material, shared with the company for comment in September 2026, identified real-time transport as the point where its documentation runs furthest ahead of its evidence, because Deaf quality review sits before release for prepared content but after delivery for instant output. This video is what that gap looks like when it reaches a screen.
Who is on staff does not answer this
The instinctive question when a signed message goes this wrong is whether anyone Deaf was involved. In this case, Deaf professionals are involved. Signapse publishes a Deaf Advisory Board, a Deaf Impact Officer and named partnerships with Deaf organizations, and the ASL for this deployment was produced by a professional translator. None of that touched this announcement at the moment it played. The signed message for baggage claim and tickets was very likely translated correctly, for the train going the other way. The clip was accurate. Its placement was not. That the video circulated as a showcase, with the mismatch unremarked, suggests that no one’s process includes checking what the live screens actually display.
This is why representation cannot stand in for governance, an argument The Fifth Parameter has made before in Issues 005 and 006. Deaf participation in building a system says nothing about whether any person or control verifies its output at the point of delivery. The question a buyer should ask is not who works on the product. It is what stops a wrong sign from reaching a platform screen, and who finds out when one does.
What buyers should require, and what Signapse owes
Airports and transit authorities that deploy automated signed announcements should write the following into their contracts, whichever vendor they use:
1. Source-to-sign reconciliation for every variable field (times, destinations, gates, platforms), with the signed video suppressed whenever the two disagree.
2. A log of every suppressed or mismatched announcement, reported to the deploying authority on a fixed schedule.
3. Periodic Deaf review of output sampled from live screens in the deployed environment, not only from the vendor’s studio.
4. A route for a Deaf passenger to report a wrong sign from the screen itself, with a stated response time.
5. Clear allocation of responsibility when a passenger is misdirected by signed content.
Signapse, for its part, owes the Deaf community and its customers a public account of this video. That account should say whether the mismatch occurred in production at CVG or only in the published clip; what caused it; how many deployed announcements share the same mapping; and what control now prevents it from recurring. The video should be corrected or withdrawn. NCG will publish any response from Signapse in full.
Signapse has published a governance framework more detailed than most of its competitors’. That is precisely why this matters. The standard it set for itself is the standard this announcement failed.
Ilaha
1. Elsie Stecker, [LinkedIn post with video of the CVG train screen](https://www.linkedin.com/feed/update/urn:li:activity:7513985603408134148/), October 8, 2026.
2. Signapse, [Automatic transport announcements in sign language: how does it work?](https://www.signapse.ai/post/automatic-transport-announcements-in-sign-language-how-does-it-work)
3. Signapse, [AI Governance Framework](https://www.signapse.ai/ai-governance-framework-2), including the Translation Quality Assurance Policy and the AI and Human Translation Use Policy.
4. Signapse, [Signapse and CVG Airport partner to enhance accessibility with American Sign Language translations](https://signapse.ai/case-studies/signapse-and-cvg-airport-partner-to-enhance-accessibility-with-american-sign-language-translations), June 25, 2024.
5. [Cincinnati Airport People Mover](https://en.wikipedia.org/wiki/Cincinnati_Airport_People_Mover), Wikipedia.
6. The Fifth Parameter, Issue 005, “The Deaf Halo Effect: When Representation Is Mistaken for Governance,” and Issue 006, “Representation Is Not Governance: The Most Common Mistake in Accessibility Innovation.”