The cost of errors is rising: how language models are transforming the aviation industry.

About Large Language Models

Large language models are neural networks specifically designed for processing and generating human language. These models are trained on vast amounts of text and utilize complex algorithms to analyze and understand linguistic patterns. As a result, LLMs can predict the next token (word or symbol) in a text based on what has come before.

The initial steps in natural language processing were taken back in the mid-20th century, when primitive algorithms for syntax analysis and text generation were developed. However, significant progress was only achieved with the advent of machine learning and artificial neural networks. The most well-known example of a large language model is OpenAI GPT, introduced in 2020.

One of the key characteristics of LLMs is the improvement in the quality of their responses as the amount of text they have "read" during training increases. Modern models, such as GPT-3, contain hundreds of billions of parameters and are trained on trillions of words. This enables them to generate solutions to even quite complex problems.

Large language models are actively used in marketing, economics, text processing, and creating instructions for chatbots. However, in fields where the cost of error is high, their integration is progressing at a slower pace. Aviation is one such field. As Yuri Chaynikov explained, decisions on key issues are still made by humans here. Nonetheless, LLMs have already helped ease some processes.

Large Language Models in Aviation

In particular, large language models have helped optimize the work of software developers for aircraft. A special system called Copilot (the so-called second pilot) interacts with human developers, analyzes their code, and suggests improvements and additions to the current project. This allows programmers to save significant time and effort on monotonous and routine tasks.

— The productivity of even a top-notch developer increases by 10-15 percent thanks to this technology. And this is no longer the future; it's an undeniable present, — explained Yuri Chaynikov.

According to the expert, neural networks have also provided significant advantages to designers, whose work includes tracking trends based on open publications.

— Thousands of articles are published each year, and all of them need to be read, considered, and assessed for their relevance to current tasks. This is a challenging job that a language model can handle: we load texts into it, it reads them, summarizes, categorizes, and thematizes. We can also assign the task of searching for information on the internet to a large language model: in patent sources, popular news, and specialized publications. This way, it can form some vision based on our request, — added the expert.

Yuri Chaynikov clarified that neural networks can also analyze a list of requirements for the compatibility of parts and suggest assembly options to the design engineer that will help solve the tasks at hand.

— The final decision is still up to the designer. But a large amount of routine work related to selection and mutual verification can be delegated to the machine,— Chaynikov said, adding that large language models can even calculate the cost of various types of assemblies and identify the most advantageous options among them.

Furthermore, the use of LLMs is becoming increasingly effective in marketing and copywriting, which are essential for the aviation industry as well. Neural networks excel at writing persuasive texts, analyzing and capturing audience needs, adhering to a specified style, and effortlessly adjusting it based on required parameters. Tasks that might take a human several hours can be completed by artificial intelligence in a matter of seconds, thus partially replacing specialists in these fields.

Additionally, large language models assist in analyzing data from maintenance reports, identifying patterns, and predicting potential malfunctions. For instance, AI can aid in decoding and analyzing text reports from engineers and technicians. Moreover, LLMs can go even further and provide recommendations for addressing issues, thereby enhancing diagnostic accuracy and reducing the likelihood of human error.

Language models are also utilized in training pilots, engineers, and other airline personnel. AI platforms can tailor educational materials to the knowledge level of each specific specialist, as well as create interactive simulators. For example, a system based on language models can simulate complex situations and suggest possible actions based on the decisions made by the employee.

Large language models are also used to improve passenger service. For instance, chatbots and voice assistants based on LLMs help answer customer queries, find information about flights, schedule changes, or delays. These models can process requests in multiple languages, which is particularly important for international airlines. This capability reduces waiting times on the line and enhances passenger satisfaction.

Error Probability

However, even in the aforementioned fields, large language models, at this stage of their development, cannot completely replace human labor, as they sometimes generate fabricated facts. Therefore, constant oversight of LLMs is necessary at this time.

— There will always be errors. The question is how the system's optimal behavior and its interaction with humans are structured. For example, consider the work of a designer: the neural network has made some selections, drawn some conclusions, but ultimately, it is up to the human to decide whether to use a particular approach or not, — said the director of the digital transformation department at BetBoom.

According to him, over time, more decisions will be made by neural networks, and the likelihood of errors on their part will decrease. In the future, we can expect the emergence of even more powerful and adaptive models that can integrate more deeply into operational processes. For instance, LLMs could become part of "smart" air traffic management systems capable of processing vast amounts of data in real time and assisting operators in making optimal decisions. Another promising direction is the creation of fully autonomous maintenance systems that can predict malfunctions and propose solutions without human intervention.

New LLMs will also aid in developing communication systems between airlines and airports to enhance flight coordination and optimize airspace usage. Such technologies will contribute to increasing air route capacity and reducing delays, which is particularly relevant in light of rising passenger traffic.

The expansion of the application of large language models in aviation will accelerate data processing, reduce costs, improve labor productivity, and simultaneously lower the risk of human error. However, broader implementation of technologies into everyday life will lead to a new problem. There will be no one to take responsibility for the mistakes made, and no one to — bear the consequences of the damage, believes Yuri Chaynikov. This is one of the challenges that will need to be addressed in the future to actively utilize large language models in the aviation sector. If the right approach to this issue can be found, then in the long run, new technologies will enhance efficiency, safety, and quality of passenger service, while also making flights more comfortable and accessible.

This material was prepared with the support of the Ministry of Science and Higher Education of Russia.