United Airlines blocks AI bots from booking flights to prevent fraud

United's move against bots

United Airlines announced it will restrict artificial intelligence bots from interacting with its digital platforms. The company stated that allowing third-party AI assistants to access account credentials could expose sensitive personal information to unauthorized parties.

This security measure is intended to prevent unintended purchases and other types of unauthorized account activity. By blocking these bots, the airline aims to ensure that only human users are making direct changes to their travel plans and personal data.

As CNBC reported, this move mirrors a recent decision by Amazon. The e-commerce giant similarly blocked Meta's Muse agent from making purchases on its website to maintain control over transaction security.

The limits of agentic AI

While AI agents can perform complex tasks like suggesting outfits based on weather or creating travel itineraries, they still operate in a very literal sense. Shomit Ghose, an adjunct professor at the University of San Francisco, notes that bots lack the human ability to infer context from instinct, culture, or education.

Because bots treat all data as equal, they may fail to meet specific user standards. For instance, an agent asked to find a "good" restaurant might select a place with a high self-proclaimed rating on its own website but poor actual reviews from customers on other platforms.

To get accurate results, users may need to provide extremely specific guardrails. Instead of a general request, a user might need to explicitly tell the bot to check specific review sites like Google Reviews or Yelp to ensure the results meet their personal preferences.

Risks in finance and health

The risks of outsourcing decisions to AI become more significant when they involve health or wealth. Ghose suggests that while agents are good for categorizing spending or flagging unusual charges, they should not be used for moving money, making trades, or filing official forms.

In the healthcare sector, AI agents could potentially manage prescription refills. However, there is a risk that these agents might act on incomplete records and fail to recognize dangerous drug interactions or worsening medical conditions before an action is taken.

In the investment world, experienced traders acknowledge that agents can make the trading experience easier. However, they warn that letting bots make independent investment decisions can lead to significant financial losses if the human is not involved in the final execution.

Human-in-the-loop strategy

Experts advocate for a "human-in-the-loop" approach where the AI handles the heavy lifting of research while the human makes the final call. For example, an agent can research destinations based on a budget and weather, but the actual booking of flights and hotels should remain a manual task.

Data from Booking Holdings shows that customers using AI assistants to book accommodations on Agoda spent 20% more than those who did not. It is currently unclear if these users were led to pricier options or simply chose more expensive amenities on their own.

To mitigate risk, Ghose recommends that users only grant AI agents access to limited permission accounts. This means letting a bot access a news subscription rather than providing it with direct access to a personal bank account or credit card information.

Defining specific goals

For users who do want to use AI for complex tasks like investing, it is essential to clearly dictate the strategy. Instead of letting a bot decide, users should provide specific goals, such as whether they are looking for short-term or long-term optimization.

By providing this context, the AI can ingest information more effectively and treat it with the correct priorities. This ensures the agent acts as a tool for information gathering rather than an autonomous decision-maker for high-stakes activities.

Ultimately, the goal is to use AI to enhance human productivity without sacrificing privacy or financial security. By keeping the human in control of the final click, users can enjoy the convenience of AI research while avoiding the pitfalls of literal-minded automation.