GENEVA—Web Manuals, the Swedish digital document management company, has announced a first customer for Amelia AI, its large-language-model-based search tool. Skyside GmbH, an aircraft management and charter company based in Vienna, Austria, will use Amelia to enhance the management of its digital documents, and is expecting to see efficiency improvements, particularly around compliance tasks.
“With Amelia, we anticipate a significant enhancement in our ability to quickly find crucial information,” Jonas Conrad, Skyside’s administrator for electronic flight bags, stated in a Web Manuals press release.
What Skyside will not be doing immediately is allowing pilots to use Amelia to search manuals in flight. That functionality will arrive eventually, Web Manuals CEO Martin Lidgard says—but there is a critically important education job to be done before that capability can be deployed. And part of that job is done by the name it has chosen for its AI tools.
“The naming is not random,” Lidgard told Aviation Week Network Show News during an interview on the company’s stand at the European Business Aviation Conference and Exhibition (EBACE 2024) here. “It’s a homage to one of the great explorers and pioneers in aviation. She was exploring the unknown. But . . . we all know the fate of Amelia Earhart. And just as you need to be mindful of how you navigate, use of AI comes with risks.”
The explosion in widespread public use of large language model (LLM) AI tools is spurring transformational change in almost every industry and society, but even as products such as Chat GPT offer huge productivity gains in certain areas, some problems are already becoming clear.
For example, LLMs have a tendency to “hallucinate,” inventing phantom facts or making up nonexistent sources. In their, as it were, eagerness to please, the systems, trained on vast sets of data,offer answers that sound plausibly convincing, yet may not contain any reliable facts.
This is already causing difficulties for college students and unwary authors—but the consequences of it happening in the preparation of a flight operations manual or in response to an urgent query from a pilot confused by an unfamiliar series of alerts in the cockpit during an emergency could be catastrophic. Web Manuals has chosen, therefore, to place strict limits on how Amelia operates.
“At this point, there is no actual training of the model,” Lidgard said. “The training it has is essentially how to construct a sentence and how to compare and combine information—but not the information itself. It does a search among the manuals that are available to that specific user who’s using it, and that is the only material that it’s getting access to. That search is more advanced than just a traditional search—then the large language model combines, or selects, the parts that are most relevant and writes a summary of those search results.
“There should be no room for hallucinations,” he adds. “That doesn’t mean that there’s no room for errors, but it’s not going to make something up out of the blue. And the way that we configure it is also exceptionally restrictive. It’s not using any information coming from anywhere else on the internet; it’s not using information that’s coming from any other customers or even from any manuals that you don’t have access to. So it’s a very tight rein.”
Skyside’s experience using Amelia will not just be of use and interest to the operator themselves but will be of great importance to Web Manuals. Lidgard said his company is not yet certain to which parts of the aviation information environment the technology will provide the most use and value for customers, and early user feedback will therefore help shape future development of the system. But as use increases, so will confidence in Amelia’s outputs, thus preparing the ground for future rollout to pilots.
“In order for people to do the right thing, which hopefully leads to safety, they need to be aware,” Lidgard said. “The challenge is that there are so many regulatory components, commercial components, technical components and operational aspects. Making sure that these massive amounts of information are made more accessible to pilots is a better option than to search for a keyword and then having to manually understand 50 or a hundred search results.”
If easier searches using conversational queries lead to better, deeper engagement between pilots and manuals, then Amelia’s deployment will lead to greater understanding and increased safety, Lidgard noted.
“Pilots, as I guess the primary user group, will be more active in interacting with the manuals, because it’s easier, more engaging. And that’s always a good thing,” he said. “And when they find themselves in a tight spot and they’re looking for that critical information, they will actually do that, rather than saying, ‘Oh, I don’t have time to do that search.’ So they will engage when they actually really need them to engage. And I think that’s something that can definitely contribute to safety.”
Getting to that stage may not take too long. “As long as we provide the proper training to ensure awareness among our users, [Amelia] is at a stage where it is already very good,” Lidgard said.