An artificial intelligence company contracted by the U.S. government says its AI systems produce near-perfect results when analyzing large datasets.
PrimerAI announced on October 14th that updates to its AI platform will reduce the incidence of hallucinations to nearly zero. As a result, the company believes the results will have far-reaching implications for the Department of Defense and the defense industry as a whole.
In AI parlance, the term “hallucination” refers to a model spitting out false results.
“In high-stakes environments where accuracy and timelines are critical, Primer’s enhanced platform emerges as an innovative solution,” the company’s press release states.
PrimerAI CEO Sean Moriarty explained what’s important about the system changes in a phone interview with Military Times.
Moriarty said that while many AI platforms have a 10% hallucination rate, PrimerAI has reduced that to 0.3%.
The biggest benefit of this update was the ability to fact-check your own results.
The company says this unique system, which catches more than 99% of errors before they reach the user, is called the Search Enhancement Production Verification System. Large-scale language models (LLMs), which are AI systems that can understand and process human language, already use search expansion generation when given a prompt. ChatGPT is a prime example.
According to Cindy Ma, Senior Product Manager at PrimerAI, the novelty of PrimerAI’s system is that when it generates a response or summary, it generates a claim against that summary and supports that claim with the source data.
Marr, who demonstrated the system in person Oct. 14 at the Association of the U.S. Army’s annual conference in Washington, D.C., said this additional layer of correction dramatically reduces errors.
PrimerAI believes that providing a large buffer against inaccuracy is of paramount importance to the Department of Defense, as small hallucinations can cause dramatic reactions.
“Imagine a world where the LLM says it has five times as many aircraft carriers as the enemy actually has,” Moriarty said.
Moriarty acknowledged that despite striving for perfection, there is always room for error. Even though high-quality reference data is the foundation of AI system analysis, there are still pieces of the puzzle that can be distorted. The data itself can be contaminated by human judgment.
“We strive for zero defects, but the world we live in today is not a zero-defect world,” Moriarty said.
Riley Cedar is an editor at Military Times, where she covers breaking news, criminal justice and humanitarian stories. He previously worked as an investigative intern at The Washington Post and contributed to the ongoing Badge Abuse investigation.
