Commercial aviation has reached a level of safety that would have been difficult to imagine a few decades ago. According to IATA’s 2025 Annual Safety Report, the industry operated 38.7 million flights and recorded an all-accident rate of 1.32 per million flights. Eight fatal accidents were recorded during the year. The longer-term trend is also clear: fatal accidents have become less frequent even as the industry has continued to grow.
Those numbers are important context for the discussion about artificial intelligence. Aviation did not become safe because of one breakthrough technology. It became safe through disciplined reporting, standardized procedures, training, investigation, data analysis, and a willingness to change when the evidence showed a weakness. AI is now entering that system, and it should be judged by the same standard.
For safety leaders, the useful question is not whether AI is impressive. It is much simpler: where can it improve a real safety process, and what controls are needed before we rely on the result?
Where AI Is Useful Today
Predictive maintenance is one of the clearest places to start. Modern aircraft generate large volumes of information on engine performance, vibration, hydraulic pressure, temperatures, and other parameters. Machine learning can help identify combinations or trends in that data that deserve attention before they become obvious through conventional monitoring. The point is not that every anomaly predicts a failure. The value is that maintenance teams can focus their attention earlier and make a better-informed decision about whether action is required.
From a safety perspective, that can reduce exposure to technical events and, just as importantly, reduce the operational pressure that follows an unexpected defect. The technology supports the engineer; it does not replace the engineering judgement or the maintenance release process.
Flight data monitoring is another natural application. FDM and FOQA programs already compare large numbers of flight parameters against defined event logic. The challenge is often what happens after the event is detected: reviewing cases, finding recurring patterns, and understanding whether several apparently unrelated events point to the same underlying issue. AI-assisted analysis can help shorten that search. For a safety team, the real benefit is not an automated conclusion. It is having more time to interpret the data and less time spent looking for the pattern.
One practical lesson from working with FDM is that the event itself is rarely the answer. A high-acceleration landing, an unstable approach, or an overspeed can be detected automatically, but the meaningful safety question is why it happened, whether the same combination of factors is recurring, and what action will actually reduce the exposure. That still requires context from crew reports, airport and weather conditions, SOPs, training, and operational history. AI can help connect those pieces; it should not shortcut the investigation.
Change management is a more interesting case because the process depends heavily on the quality of the questions asked before a change is introduced. A new route, a change in ground handling, a fleet transition, or a software modification may all require a formal safety risk assessment. In practice, those assessments combine available data with the experience of the people in the room. That experience is essential, but it is also limited by what the team remembers and what information it can reasonably review.
This is where natural language processing could add value. A tool can search internal safety reports, maintenance records, previous investigations, and earlier risk assessments much faster than a person can. Used well, it may surface hazards or combinations of hazards that deserve consideration and might otherwise be missed during a workshop.
There is an important limit, though. The AI output should be an input to the risk assessment, not the risk assessment itself. A broader list of possible hazards is useful only if the organization can verify the source information, understand the operational context, and decide which risks are actually relevant. The final judgement still belongs to the accountable operational and safety functions.
In my experience working with safety systems across multiple AOCs, the difficult part of Management of Change is usually not completing the risk matrix. It is making sure the team is asking the right questions and learning from what has already happened elsewhere in the operation. Similar events may be classified differently, stored in separate databases, or known mainly to local teams. An AI tool that can retrieve and connect those signals across operations could be genuinely useful, but only if the underlying reporting taxonomy, data quality, and governance are disciplined.
The Hard Question: What Do We Need to Trust?
This is where aviation’s relationship with AI becomes more difficult. Safety management relies on traceability. A captain should be able to explain a decision. A maintenance engineer signs for work that has a documented basis. After an occurrence, investigators need to reconstruct what happened and why. That expectation does not disappear because the recommendation came from an algorithm.
Some AI systems, particularly more complex machine learning models, can identify useful patterns without providing an equally clear explanation of how a specific recommendation was produced. That creates a problem for safety-critical applications. Good statistical performance is important, but it is not the only requirement. If an organization cannot explain what data the model used, how the model was validated, or under what conditions its output should not be trusted, the operational case for relying on it is weak.
The issue is practical, not theoretical. Regulators will need to decide what evidence is acceptable for certification and oversight. Airlines will need to define which outputs can support a decision and which require independent human review. Investigators may eventually have to examine cases in which an AI recommendation was followed, ignored, or misunderstood. In each case, accountability has to remain clear.
A sensible approach is to treat AI like any other safety-relevant input: validate it, define its limitations, establish when human review is mandatory, and keep enough information to understand how the output was produced. Aviation has always been cautious about new sources of safety information. That caution is a strength, not an obstacle to innovation.
The Human Factor Does Not Disappear
It is tempting to describe AI as a way to reduce human error by reducing human involvement. In safety-critical work, that is too simple. The human decisions do not disappear; they move to different points in the process.
Someone chooses the data used to build the model. Someone defines what the system is allowed to recommend. Someone decides when the result is good enough to influence an operational decision. And, at the frontline, someone may still have to decide whether to follow or override that recommendation. Each of those decisions creates its own failure possibilities.
That is why investment in AI cannot be separated from investment in people. Safety analysts need to know how to question an algorithmic output rather than simply accept it. Organizations need clear boundaries between AI-assisted and autonomous decision-making. They also need to protect the reporting culture that feeds the safety system with information in the first place. A sophisticated model built on poor, incomplete, or distorted reporting will not solve the underlying problem.
Another lesson from day-to-day safety management is that most teams do not suffer from a lack of data. They suffer from fragmented data and limited analyst time. Mandatory and voluntary reports, FDM events, maintenance reliability information, audits, ground occurrences, and operational changes often sit in different systems. For that reason, one of the most valuable uses of AI may be less dramatic than autonomous decision-making: helping a safety team connect weak signals early enough to decide whether they deserve action.
Why This Matters in Latin America
Latin America is not a single operating environment, and it is risky to generalize about the region. Still, many carriers operate with tighter margins and face infrastructure constraints that can limit access to the same analytical capability available to larger operators. That makes the economics of new safety technology particularly relevant.
Cloud-based analysis tools, scalable FDM solutions, and commercially available predictive maintenance platforms may lower some of those barriers. For a mid-sized or low-cost carrier, the opportunity is not to copy the technology strategy of the largest airlines. It is to use tools that can add value to an existing safety management system without creating a level of complexity the organization cannot properly govern.
That distinction matters. AI can be useful because it allows a safety team to review more information, detect patterns earlier, and challenge assumptions with data. But none of those benefits justify weak validation or unclear accountability. The technology has to fit the maturity of the organization and the criticality of the decision it is supporting.
AI will not make aviation safer on its own. It can help us see things sooner and process information at a scale that was previously difficult, but it still operates inside a safety system designed and managed by people. The standard should therefore be familiar: evidence, defined controls, traceability, and a clear understanding of limitations. In aviation, that is how a new tool earns trust.
