As artificial intelligence (AI) marches
toward general-purpose capability, travel remains both tantalized and trapped.
On one hand, AI promises frictionless journeys, hyper-personalized offers and
operational efficiency. On the other, the sector’s stubborn legacy
infrastructure and data fragmentation leave even the most powerful large
language models (LLMs) hallucinating under pressure.

This isn’t just a scaling issue—it’s a
training issue. The time has come for the travel industry to make adversarial training
(also known as adversarial deep learning) a core requirement, not just a
curiosity.

What is adversarial training and why does it matter?

Adversarial training uses deliberately
crafted “edge case” inputs—scenarios that push the model into ambiguity, error or
confusion—to strengthen performance. These examples aren’t noise; they’re
designed to probe blind spots, force corrections and ultimately make the model
more robust, especially in high-stakes decision-making. The use of edge cases
to find failures is not new. But AI takes that process to a higher level and
faster. That is better for complex environments like travel.

In other fields, for example in
medicine, Google’s DeepMind uses adversarial examples to refine AI diagnostic
reasoning. In finance, JPMorgan has tested similar frameworks to guard against
risky generative outputs.

OpenAI incorporated red-team adversarial prompts into
GPT-4’s release to catch hallucination-prone use cases. Even Microsoft’s GitHub
Copilot uses this technique to identify corner-case bugs before they reach
production.

But in travel? There is almost no
formal adoption of adversarial training—despite travel’s uniquely complex,
interdependent systems. Er…can I get a trip please?

Why travel is an adversarial minefield

Travel isn’t just another vertical.
It’s a web of exceptions and irregularities masquerading as rules. As we have
seen in the journey to offer and order, this complexity and process has proved
to be unnecessarily slow and overly complex. The result is it seems the
journey is not moving fast enough.

We have to remember that the
uniqueness and legacy processes abound. Every itinerary depends on dynamic
pricing, fragmented inventory, overlapping regulatory regimes and constraints
that are embedded in these legacy frameworks. They should ONLY be constrained
by true dynamic market forces. To be clear they are not today, and that horizon
seems to be pretty far away.

There are global schemas that capture
this to enable such things as a single itinerary from Nairobi to Sydney, which touches
four continents, five regulatory zones [and involves] interline or codeshare
logic. But is that necessary for a London to Rome single flight, and if so,
why?

We know much of the critical data
sits inside proprietary silos—global distribution systems, airline reservation
systems and loyalty platforms that interoperate on many of these arcane rules
defined decades ago. Heck, we are still using steamship analogies. It is time
for a change. We have become scared to sweep this away because we are afraid of
these embedded legacy edge cases.

Adversarial training is made for this

Instead of pretending AI can infer all
this from context or hope, we should deliberately stress-test models against
these edge cases: open-jaw tickets, split passenger name records, nested fare
rules, denied boarding rules on mixed carriers. 

By feeding those back into
training, we create AI that is not only fluent—but trustworthy. Now, that is
something that has eluded airlines for decades, as they are protected through
industry norms and rules that can make anyone’s head spin.

Are travel companies doing this yet?

Reports suggest:

• Amadeus has begun internal tests of LLMs
  for agent-assist workflows in call centers and B2B servicing environments,
  where accuracy and recall of fare rules are critical. Though not labeled
  as adversarial training, these quality assurance processes simulate many
  of the same effects by injecting structured edge-case scenarios during
  model evaluation.
• Hopper, Google Travel and others
  have also seen firsthand the cost of AI hallucinations in production—where
  bots have invented prices or incorrectly interpreted refundability. These
  incidents underscore the urgent need for AI stress testing frameworks.

The performance problem is just as real

It’s not just about factual
hallucinations. Many travel models today fail silently—timing out, stalling on
long itineraries or choking on ambiguous prompts. Adversarial input techniques
can expose these issues, allowing developers to tune memory allocation,
contextual threading or backend dependencies.

This kind of real-world “load testing”
is essential for multi-turn travel planning, especially as we move toward agentic
AI models that autonomously book, rebook or manage travel end-to-end.

A personal note: Why edge cases matter

As the founder of Air Black Box (ABB),
one of the first platforms to enable true open interlining across independent
airlines, I learned firsthand how edge-case thinking can force an industry
forward.

Prior to the deployment of ABB’s
patented solutions, the status quo was joint ventures or tightly constrained
interlining—models that left out regional, low-cost or startup carriers. We
built a capability to connect them anyway, precisely because the infrastructure
didn’t support it.

Today, the same mindset applies. If we
want AI to truly serve the needs of travelers and not just mimic travel agents
poorly, we have to confront complexity head-on. That starts with adversarial
training—not as a patch but as a strategic method.

Final thought: Don’t build on sand

If your AI roadmap doesn’t include
adversarial training, you’re relying on the hope that your model will “just
know better.” But in travel, hallucinations aren’t just embarrassing—it’s
operationally dangerous. And they do happen.

The future of AI in travel depends not
on bigger models but on smarter training. Adversarial deep learning is
the stress test this industry needs.

Let’s stop waiting. We can get it
right this time: faster, more reliably and less costly.