Anthropic researchers are developing FlashProve, an open-source system meant to verify which artificial intelligence model produced a given response. In a research announcement published Tuesday, they describe the work as “provable inference”: cryptographic checks that supply mathematical evidence tying an output to a specific model.
The goal is practical. Users and developers could confirm that an answer really came from the model a service said it was using, rather than taking that claim on trust.
How FlashProve is meant to work
FlashProve is compared in the announcement to cryptographic email signatures that let a recipient check a message’s claimed sender. Researchers say the approach is meant to resist impostors who try to pass one model off as another—including when the impostor is the same lab that built and runs the model.
For the system to function, AI labs would publish a “fingerprint” for each released model. That fingerprint could later be matched to specific outputs and verified cryptographically. Researchers stress that fingerprints would not reveal model weights—the internal numbers that define behaviour and that closed-model makers such as Anthropic and OpenAI keep private. In principle, commercial labs could publish fingerprints without handing over their most valuable asset.
The design is also described as lightweight, so verification could run on a local device rather than relying on a distant server alone.
What problems it could address
Researchers note FlashProve could make it harder for a company to secretly swap in a cheaper model to cut costs, or a stronger one to inflate reported performance. It could also help check third-party claims that a named advanced model completed a task. In the announcement, researchers ask how a user could verify, for example, a claim that a software library was scanned for vulnerabilities by a frontier cybersecurity model, or that a business plan was reviewed by an appropriate model.
Still research, not a product rollout
Alongside the post, researchers released a detailed technical specification and working code for parts of the system under an open-source licence on GitHub. They say research continues and that there are no plans to deploy FlashProve at Anthropic or any other lab.
“This research post is meant to share intermediate progress, and not a completed research agenda,” researchers write, adding that they want others to examine the work, reuse it, and improve it.
For Pakistani readers who use chatbots for study, work, or customer service, the announcement is about trust infrastructure still under construction—not a switch flipping overnight on local apps.
Aasan alfaaz mein
Anthropic FlashProve naam ka open-source system bana raha hai. Iska maqsad yeh hai ke aap check kar saken ke AI ka jawab asli usi model se aaya jo company ne bataya. Abhi yeh research hai; kisi lab mein deploy karne ka plan nahi bataya gaya. Agar yeh aage chalay to AI services zyada transparent ho sakti hain.
