When Gemini does reasoning, can first responders rely on that reasoning? 

Researchers at Google in October 2025 posted on a new geospatial reasoning capability they’re testing that brings together different global AI models for disaster relief scenarios. They invited nonprofits to become “trusted testers” of the capability. Here’s the gist of the post:

“To solve a complex problem, you need to see the whole picture, not just one piece of it. That’s the idea behind Geospatial Reasoning, a framework powered by Gemini that now lets AI automatically connect different Earth AI models — like weather forecasts, population maps and satellite imagery — to answer complex questions.

“Instead of just seeing where a storm might hit, our latest research demonstrates that analysts can use Geospatial Reasoning to identify which communities are most vulnerable and what infrastructure is at risk, all at once. For example, Geospatial Reasoning empowers the nonprofit GiveDirectly to respond to disasters by combining flood and population density information, helping them identify who needs direct aid most.”

LLMs and chain-of-thought prompting as human-assisted reasoning

Others I’ve asked presume that Geospatial Reasoning uses chain-of-thought prompting. CoT refers to a step-by-step process to reach a conclusion. If the input data is trustworthy, the outputs can also be trustworthy. 

As I’ve said before, quality of LLMs and other statistical machine learning systems depends entirely on the quality of the input data. 

In the case of Google’s Geospatial Reasoning, Google systematically collects oodles of big data for Earth and has for over 25 years now. 

The CoT method itself just goes back to 2023 or so.

The CoT method used by LLM providers is useful for some purposes. Graphwise, for example, has evaluated the entity annotation capabilities of large language models and states that “…for common and clearly described entities like Person, Organisation, Disease, LLMs are close to human annotators and can speed up the annotation process.”

But here’s the bottom line on CoT and probabilistic-only approaches: CoT-using LLMs succeed with the pattern matching attempt turns out to be (more or less) a successful match. LLMs can’t explain why they draw the conclusions they’re asserting as answers by providing you a logical, ordered rationale that makes sense. 

As a result, you can’t trust the answers they’re giving. And because you know at least some responses will be wrong or even fabricated, but can’t predict which ones, you have to assume that every response might be useless unless thoroughly checked.

Most businesses have an additional challenge that Google Earth doesn’t: They can’t amass the data about their business in the volumes Google does about the earth for the Google Earth service.

Even when it comes to mainstream entities, Graphwise points out the necessity of checking after the LLMs you’re using for entity recognition: “There is no silver bullet: LLMs still need human validation and there does not seem to be one best model as we’ve seen that Llama-70b and GPT-4o perform differently on different tasks.” 

Graphwise recommends a blend of probabilistic methods for entity recognition and in conjunction with machine-assisted human-in-the-loop graph modeling, data enrichment and validation. The company has long experience in NLP and vector databases (See https://graphwise.ai/blog/how-far-we-can-go-with-genai-as-an-information-extraction-tool/ for more information.)

By the way, semantic knowledge graph retrieval generated automation (GraphRAG) best practices assume human-in-the-loop.

In any case, if CoT alone is the method of choice and you haven’t collected the oodles of data that Google has, you’ll need humans prompting up front as well as in the validation loop on the back end.

Why the NeSy symbolic reasoning methods used in semantic standards-based graph RAG can reason logically, but LLMs can’t

LLMs in a non-GraphRAG scenario look for a pattern to match. If there’s no solution available that matches a given problem set, a solution can’t be found. 

Now consider neurosymbolic (NeSy) methods as an alternative, such as in cases where there isn’t a vast amount of data to be tapped. Semantic standards graph RAG-based systems are NeSy by design. There is the neural net/statistical machine learning part of the system, such as the LLM and the natural language processing. But then there’s the symbolic AI side too. Symbolic logic is a long established AI method that doesn’t require to be trained on massive amounts of data the way LLMs do. 

What NeSy architecture implies is the ability for the overall system to harness both probabilistic and deterministic methods. An agent can query via an LLM. The query can be transformed into SPARQL, which calls and retrieves a trusted answer from a deterministic, contextualized knowledge graph modeled using web ontology language (OWL). OWL defines the logic. 

OWL reasoning is deterministic, symbolic logic reasoning that complements the probabilistic methods used in LLMs.

According to Herron, Ruiz and Weyde of the University of London and Oslo in a recent paper, “OWL-based knowledge graphs offer the benefits of formal, logical symbolic reasoning (e.g. reliable inference of new knowledge based on Description Logics, semantic consistency checking, extensions via user-defined Datalog rules).” 

(See Herron, D., Jimenez-Ruiz, E.  & Weyde, T.  (2023). On the Benefits of OWL-based Knowledge Graphs for Neural-Symbolic Systems. In: Proceedings of the 17th International Workshop on Neural-Symbolic Learning and Reasoning. 17th International Workshop on Neural-Symbolic Learning and Reasoning, 3-5 Jul 2023, La Certosa di Pontignano, Siena, Italy.)

Symbolic reasoning: A Sudoku example

NeSy reasoning is both deductive (on the symbolic side) and inductive. Humans do deductive reasoning all the time. Take Sudoku puzzles, for example. 

A typical 9 x 9 cell Sudoku matrix requires the solver – let’s say it’s a human for now – to determine where to place nine sets of numbers 1 through 9 in rows, columns and 3 x 3 squares within the matrix so that every row, column and box is fully populated with 1 through 9.

Stated problems in Sudoku typically provide a partial solution; the human or machine just has to complete the matrix. Here’s an example from the lecture notes for a college-level C++ programming class: 

Humans who solve Sudoku puzzles like these on their own use symbolic logic, rather than mathematical logic. Humans understand the rules and know how to apply them to find a solution. The numbers in Sudoku problems and solutions function merely as symbols; you could use letters A through I instead of numbers, for example, and the means of solving the problem would remain the same.

C++ or other conventional programming that uses recursion and backtracking is a brute force method that involves constraint-checking logic to solve Sudoku puzzles. The machine running the program doesn’t know the rules human Sudoku solvers do. Rather, it’s using a trial-and-error approach that is effective, if not efficient. It finds the solution by satisfying one rule of sorts, constraint that leads it to deduce the solution.

It’s also possible to use standard semantic graph knowledge representation and retrieval technologies (such as RDF/OWL/SPARQL) to solve Sudoku puzzles. You can find examples on the web. But semantic reasoners aren’t designed for efficient puzzle solving of this sort. They’re designed for knowledge representation and retrieval, including retrieval by LLMs or other agents.

A more efficient Sudoku puzzle solving method involves a minimum remaining values (MRV) heuristic. A heuristic is a rule of thumb that enables problem solving. Peter Norvig, a Director of Research at Google and Distinguished Education Fellow at the Stanford Institute for Human-Centered AI (HAI), posted on such a method in 2006 that uses Python, if you’d like to check it out.  (See https://norvig.com/sudoku.html.)

Norvig points out in his 2006 post that “one popular heuristic is least-constraining value, which chooses first the value [cell, in the Sudoku case] that imposes the fewest constraints on peers [i.e., other cells].”

By describing the Sudoku example, I’m pointing out how humans reason and how different computing methods can replicate what humans do reliably, in a grounded, trusted, reliable and explainable way.  It makes sense to build systems that incorporate these established methods.

GraphRAG: A gateway to more complete reasoning and a path to enriched data quality

If I had to guess, I’d say that Google Earth’s Geospatial Reasoning will become consistently useful for first responders. But I’d add the caveat that organizations–including Alphabet/Google– have to make decisions based on a wide range of variables. Not all of those variables will be informed by the kinds of vast public and private datasets that Google has amassed and curated over 25+ years for its Earth service. 

Now contrast this apparent plain RAG approach with GraphRAG. GraphRAG can be at the heart of a more complete, contextual computing environment. All sorts of heterogeneous data can be brought together, managed and reused in a uniform manner, holistically. NeSy, knowledge graph-based systems take advantage of the power of LLMs by using them as capable front end chatbots, but they add the ability to use internal business data that simply can’t contain massive numbers of examples.

Knowledge graphs use context-building metadata (data about data) to place a data source in a context or interrelated contexts where that data can be reused.

In other words, if you’re building knowledge graphs, you’re building a mirrorworld of sorts, using the data you can collect and the data you have at hand, refined and articulated so that the results can be trusted. By contrast, plain RAG won’t be effective inside business units that simply don’t generate vast volumes of examples.