Why does AI give wrong directions? Common causes explained

Try planning a route across London with an AI assistant and it might confidently send you across the wrong side of the River Thames—placing the Elizabeth Tower in entirely the wrong location. This is not a hypothetical edge case.

“AI can’t yet read maps,” HERE Technologies CEO Mike Nefkens said in a recent article with Automotive World. “Maps are lines and polygons. AI reads text, code and other data features, but it can’t read lines and polygons. It hallucinates dramatically, and we have hundreds of examples of that.”

In this article, we'll explain why AI gets navigation wrong, why LLMs aren't designed to make accurate routing decisions and how high-quality, constantly updated location data helps AI make reliable real-world decisions.

Read more: Why LLMs understand language but not space

What Is AI hallucination in navigation?

In navigation, an AI hallucination happens when an AI system gives a location-based answer that sounds confident but is actually wrong. The AI doesn't realise the information is incorrect. Instead, it generates an answer—such as an address, route or distance—that appears correct based on patterns it has learned, even though it does not match reality.

The reason this happens is how LLMs are built. They are designed to predict the most likely next words in a sequence, not to understand maps or verify real-world locations. Geographic information is different, in that it consists of specific facts that can be checked against the physical world. If an AI has incomplete or unreliable information about a place, it may fill in the gaps with an answer that sounds likely rather than one that is geographically accurate.

"Asking an LLM to compute spatial relationships is like asking someone to navigate a city using only restaurant reviews and travel blogs, instead of a map and a compass," said Aleksandra Kovacevic, Senior Director, Head of Responsible AI at HERE.

Read more: How to help LLMs understand the physical world

How training data gaps lead to wrong directions

An LLM's knowledge is frozen at its training cutoff. Roads close, speed limits change and construction reroutes traffic. The model has no mechanism to reflect any of this.

AI systems with web search access can retrieve fresher information, reducing some obvious errors—such as recommending a restaurant that has since closed. But retrieval does not solve the harder problem. Knowing a place exists is not the same as computing whether it is truly nearby, meaningfully on the way or reachable under real-world conditions. That is where language understanding ends and spatial reasoning begins.

Read more: Enabling spatial reasoning for LLMs

How removing spatial computation from the LLM fixes it

Training LLMs on map data—through geographic ontologies, knowledge graphs or specialized GeoLLM models—improves how a model describes geography, but does not enable it to compute routes, evaluate real-time traffic or enforce geometric constraints. The limitation is structural.

The correct fix is to stop asking LLMs to do geography. Standard AI "grounding" gives a model external sources to check its outputs, but the model still attempts spatial reasoning. Offloading is different: the model never attempts spatial computation at all. It recognizes a spatial query, delegates it to an execution layer, and receives back a validated result. Offloading removes the source of error rather than catching it afterward.

"AI can describe the world, but it cannot reliably compute how the world works," said Christopher Handley, Senior Vice President of Product Management at HERE Technologies.

This is the principle behind HERE Location Reasoning, a geospatial, advanced grounding solution that offloads spatial computation to a dedicated engine running on HERE's continuously updated map: more than 68 million kilometers of roads across more than 200 countries and territories, sourced from more than 238 million vehicles.

A practical example: a driver asks an in-vehicle assistant, "I need to stop for fuel before the motorway section, but only somewhere a seven-meter van can access."

Answering correctly requires computing the current route, identifying the motorway entry point, filtering fuel stations by vehicle access constraints and ranking results by actual travel time—not straight-line distance. None of those steps involve language. The LLM interprets intent. The spatial engine computes the answer.

Read more: Redefining the automotive experience at the speed of agentic AI

What happens when AI agents get location wrong

As AI moves from answering questions to taking autonomous action, spatial errors compound. Across tens of thousands of queries per day, close enough and correct are not the same thing. Organizations deploying agentic AI without a spatial layer typically encounter four problems:

False autonomy: agents appear capable but depend on humans to fix spatial blind spots

Slower deployments: teams add manual guardrails to compensate, delaying pilots

Higher risk exposure: incorrect routing, non-compliance, or unsafe actions in physical environments

Missed use cases: entire classes of operational scenarios remain unaddressable

"Everything starts with a reliable, high-quality map," said Kovacevic. "If the agent doesn't understand the world around it accurately and in real time, everything that follows is guesswork."

The language model understands what a user wants. Verified map data, queried at runtime, computes where to send them.

Frequently asked questions

Can AI replace GPS or a traditional navigation system?

No. LLMs cannot perform the real-time spatial computation that GPS and dedicated routing engines carry out. Navigation requires purpose-built mapping infrastructure, not language generation.

Does web search fix AI navigation errors?

Partially. Retrieval reduces errors from stale data but does not enable a model to compute whether a location is truly nearby, on the route or reachable under real-world conditions. The spatial computation problem remains.

What is AI hallucination in navigation?

It occurs when an AI produces a location-based answer that sounds correct but does not match the physical world—fabricated addresses, wrong distances, routes on roads that no longer exist or landmarks placed in the wrong location.

What is the difference between grounding and offloading?

Grounding gives a model external sources to verify its outputs, but the model still attempts spatial reasoning. Offloading prevents this entirely: a dedicated execution layer handles spatial computation deterministically instead.

What data does HERE use to support spatial AI decisions?

HERE's map covers more than 68 million kilometers of roads across more than 200 countries and territories, sourced from more than 238 million vehicles, updated continuously so queries reflect current road conditions.