FDA employees who meanwhile spoke to Stat News, called the tool “hurried” and said that the capacities were transferred by civil servants, including Makary and those at the Department of Government Efficiency (Doge), which was led by controversial billionaire Elon Musk. In its current form, it should only be used for administrative tasks, not for scientific, the employees said.
“Makary and doge think that AI can replace the staff and shorten assessment times, but it is certainly not possible,” an employee said. The employee also said that the FDA was not established to set up guardrails to use the tool. “I am not sure in their hurry to get it out that someone thinks through policy and use,” said the FDA employee.
According to Stat, Elsa is based on Claude LLM from Anthropic and is being developed by the Deloitte consultancy. Since 2020, Deloitte has paid $ 13.8 million to develop the original FDA documents database from which Elsa's training data is derived. In April the company received a $ 14.7 million contract to scale up the technology on the entire agency. The FDA said that Elsa was built in a GOVCLoud environment with a high security and a “secured platform for FDA employees offers to access internal documents, while all information remains within the desk.”
Previously, every center within the FDA worked on its own AI pilot. After the cost savings in May, the AI pilot was originally developed by the FDA Center for Drug Evaluation and Research, called Cder-Gpt, chosen to be scaled to an FDA-wide version and renamed Elsa.
FDA employees in the Center for Devices and Radiological Health said NBC News that their AI pilot, CDRH-GPT, Buggy is, is not connected to the internet or the internal system of the FDA and has problems uploading documents and allowing users to submit questions.
