International • • • 🌿 Progressive

Imagen alusiva de una tarjeta gráfica en un centro de datos generando una prueba criptográfica de verificación
FlashProve genera pruebas criptográficas que permiten verificar qué modelo de IA produjo una respuesta. · Imagen generada por EPM

Anthropic unveils FlashProve to verify which model produced each answer

Anthropic shared FlashProve, an open research prototype that generates cryptographic proofs to verify that an answer came from the requested model, without…

--- The context --- Anyone who queries an artificial intelligence model today cannot know for certain that the answer received came from the model selected. Anthropic explained, in reported remarks, that for most uses that uncertainty is not worrying, although in some cases certainty is a great help. The doubt grows when independent evaluators are involved, or when queries pass through routing services run by third parties. --- The facts --- The laboratory presented FlashProve, an open-source system based on what it calls provable inference. Its purpose is to provide cryptographic proof that an answer was generated by a specific model. The mechanism rests on a succinct zero-knowledge argument: alongside the answer, the provider sends a short proof that the user can verify quickly, without any knowledge of the model weights. It is a research prototype that generates cryptographic proofs with a single graphics card, a GPU H100, at a rate of 110 tokens per second. The proof takes up 19 MB for each batch of 1,024 tokens. Anthropic indicated, in reported remarks, that the proof can be checked in seconds on a conventional computer, even locally. --- The positions --- Anthropic held, in reported remarks, that independent evaluators who test models before and after deployment need to know that the model they are testing is the model being served. It also noted, in reported remarks, that clients who use a model through a third-party routing service want to make sure the tokens come from the model they paid for. The proposal puts verification in the hands of the user, even one with limited computing resources. --- What remains unknown --- It is not yet known how this tool will be adopted beyond research, or which providers will incorporate it into their services. The practical reach of its use by auditors and regulators also remains unclear, as does the cost of generating and checking proofs when query volumes grow. --- Unanswered questions --- Will other providers accept an open mechanism that lets anyone check which model actually answered? Will auditors be able to turn these proofs into a routine oversight tool without depending on the laboratory that created them? Will ordinary clients gain a simple way to verify each answer, or will checking remain reserved for technical profiles with advanced cryptographic knowledge? --- EPM analysis --- The deeper novelty is a shift in trust: the word of the provider stops being the only guarantee, and a check that third parties can repeat takes its place. 📌 EPM Take: In EPM's view, FlashProve points in the right direction because it moves trust away from commercial promises and toward a proof that anyone can check, and that raises the standard for the whole sector. Its greatest value lies in giving evaluators, auditors and clients an independent instrument when dealing with opaque providers. If adoption follows, origin verification could become a normal market requirement rather than an occasional voluntary gesture of transparency.
📤 Share on Telegram

¿Te gustó este artículo? Recibe cobertura global en tu correo.

Suscríbete gratis / Subscribe free