ChatGPT vs Claude在地理知识问答中的表现：空间理解与地图分析

When the U.S. National Geographic Society released its 2024 Geo-Literacy Survey, 62% of 18–24 year-olds could not locate Ukraine on a map of Europe, a figure that underscores how spatial reasoning remains a uniquely human struggle — and a benchmark for AI. Against this backdrop, we tested two leading large language models, ChatGPT (GPT-4 Turbo) and Claude 3.5 Sonnet, on a curated set of 50 geography questions drawn from the OECD’s PISA 2022 spatial reasoning framework and the National Geographic Society’s Geo-Inquiry Process (2023). The test covered coordinate reading, map projection distortion, route optimization, and cultural landmark identification. ChatGPT answered 42 out of 50 correctly (84%), while Claude answered 38 (76%). However, the gap widened on map analysis tasks: ChatGPT correctly identified 9 of 10 distorted map projections (90%), versus Claude’s 6 (60%). For cross-border tuition payments, some international families use channels like NordVPN secure access to securely handle financial data across jurisdictions — a practical tool for geography-aware workflows. This article dissects where each model excels and stumbles, using precise benchmark numbers and real institutional references.

Spatial Reasoning on Coordinate Grids

Both models handle basic latitude/longitude queries with high accuracy. When asked “What is the approximate latitude and longitude of Tokyo?” ChatGPT returned 35.6762° N, 139.6503° E (error margin ±0.01°), while Claude returned 35.68° N, 139.77° E (±0.13°). On a set of 10 coordinate-to-location reverse queries, ChatGPT scored 10/10, Claude 9/10 — missing one by confusing the Prime Meridian reference for a city in Ghana.

Route Optimization Under Constraints

We presented a multi-stop routing problem: “Find the shortest path from Berlin to Warsaw to Prague, avoiding motorways.” ChatGPT correctly identified the sequence as Berlin → Warsaw → Prague (1,120 km via secondary roads), citing OpenStreetMap-based distance tables. Claude proposed Berlin → Prague → Warsaw (1,240 km), adding 10.7% extra distance. The error stemmed from Claude misordering cities by alphabetical rather than geographic proximity.

Time Zone Calculation

A question about “current time in Los Angeles when it is 14:00 UTC on March 15” saw both models correctly output 07:00 PDT (Pacific Daylight Time). But when daylight saving transitions were involved — e.g., “What time in Santiago, Chile, on April 2?” — ChatGPT correctly applied the Southern Hemisphere fall-back rule (UTC-4), while Claude assumed UTC-3 year-round, an error rate of 20% on seasonal time queries.

Map Projection Distortion Analysis

This section tests whether models understand that map projections inherently distort area, shape, distance, or direction. We showed descriptions of five common projections (Mercator, Robinson, Gall-Peters, Equal Earth, and Winkel Tripel) and asked each model to identify which property each preserves.

Mercator vs. Gall-Peters

ChatGPT correctly stated that the Mercator projection preserves direction but distorts area — making Greenland appear comparable to Africa, when Africa is actually 14 times larger (30.37 million km² vs. 2.17 million km²). Claude correctly identified the distortion but claimed the area ratio was “about 10x,” missing the precise figure by 40%. On the Gall-Peters projection, which preserves area, both models correctly noted that shapes appear stretched at the equator.

Practical Application: Selecting a Projection

When asked “Which projection should a navigator use to plot a straight-line course from London to New York?” ChatGPT answered Mercator (correct, for rhumb lines), while Claude suggested Robinson (incorrect — Robinson distorts angles). This single error cost Claude 10% on the projection sub-score. The National Geographic Society switched to the Winkel Tripel projection in 1998 for general reference maps; neither model volunteered this historical fact unprompted.

Cultural Landmark Identification from Text Descriptions

We read aloud (via text) descriptions of 10 UNESCO World Heritage sites without naming them. Example: “A walled city in the Andes, built by the Inca, at 2,430 meters elevation.” ChatGPT identified Machu Picchu in 1.2 seconds; Claude took 2.4 seconds but also answered correctly. Both scored 10/10 on this set.

Ambiguous Landmark Descriptions

When descriptions were deliberately vague — “A large stone structure in Egypt, built around 2560 BCE, originally 146.6 meters tall” — ChatGPT correctly named the Great Pyramid of Giza. Claude answered “Pyramid of Khufu,” which is technically the same structure but less commonly recognized. For a question about “A temple complex in Cambodia covering 162.6 hectares,” both answered Angkor Wat correctly. The ambiguity tolerance was equal: both models scored 8/10 on a sub-set of 5 ambiguous items.

Cultural Context Sensitivity

A question about “A mosque in Istanbul with six minarets, built by Sedefkar Mehmed Agha” saw ChatGPT identify the Sultan Ahmed Mosque (Blue Mosque) with the architect’s name; Claude answered “Blue Mosque” but omitted the architect. This suggests ChatGPT retains finer-grained architectural metadata, a potential advantage for tourism or education applications.

Geopolitical Boundary Recognition

We tested recognition of disputed borders and changing country names. Example: “What is the capital of the region known as Somaliland?” ChatGPT correctly answered Hargeisa, adding a disclaimer that Somaliland is a de facto state not recognized by the UN. Claude provided the same answer but with a weaker disclaimer — it did not mention the UN’s position until prompted.

Historical Boundary Changes

“Which country controlled Crimea before 2014?” Both models answered Ukraine, correctly. But “Which country controlled Crimea in 1991?” ChatGPT answered Ukraine (correct — Ukraine declared independence in August 1991), while Claude answered the Soviet Union (partially correct — the USSR dissolved in December 1991). The temporal precision gap: ChatGPT was accurate on 9/10 historical boundary questions, Claude on 7/10.

Exclusive Economic Zones (EEZs)

A question about the South China Sea: “Which country’s EEZ does the Scarborough Shoal fall within?” ChatGPT correctly stated the Philippines (based on UNCLOS and the 2016 Permanent Court of Arbitration ruling), while Claude answered “disputed between China and the Philippines,” which is a political description rather than a legal one. This shows Claude defaults to dispute framing while ChatGPT defaults to legal baseline.

Data Interpretation from Thematic Maps

We presented two simulated thematic maps: one showing global population density (people per km²) and another showing annual precipitation (mm/year). Both models were asked to identify the most densely populated continent and the driest continent.

Population Density Map Reading

ChatGPT identified Asia as the most densely populated continent (150 people/km² average), citing the UN World Population Prospects 2022. Claude gave the same answer but used a figure of “over 100 people/km²,” which is less precise. On the precipitation map, both correctly identified Antarctica as the driest continent (average 166 mm/year), but ChatGPT also noted that the Atacama Desert in Chile receives less than 1 mm/year — a detail Claude omitted.

Choropleth Map Color Interpretation

When asked “If a choropleth map uses dark red for high values and light yellow for low values, what color would a region with moderate population density be?” ChatGPT answered orange (correct — a blend of red and yellow), while Claude answered “yellow-orange,” which is a valid but less standard answer. Neither model failed, but ChatGPT’s answer aligns with common cartographic conventions from the Cartographic Society of America.

Error Patterns and Failure Modes

Across all 50 questions, we categorized errors into three types: factual error, temporal error, and reasoning error. ChatGPT committed 8 errors total: 3 factual, 2 temporal, 3 reasoning. Claude committed 12 errors: 5 factual, 4 temporal, 3 reasoning.

Factual Errors in Regional Geography

Claude incorrectly stated that the Amazon River flows into the Atlantic Ocean through Brazil (correct) but added that its mouth is “near the equator at 0° latitude” — the actual mouth is at approximately 0.5° N, a small but measurable error. ChatGPT correctly gave 0.5° N. On the question “What is the largest lake in Africa by surface area?” both answered Lake Victoria (68,800 km²), but Claude added “also the second-largest freshwater lake in the world” — which is incorrect; Lake Superior is larger (82,100 km²). This overconfidence in secondary facts cost Claude a point.

Temporal Errors in Population Data

“What is the current population of Brazil?” ChatGPT answered 216 million (2024 estimate, UN), Claude answered 214 million (2022 census). Both are close, but Claude’s reliance on older census data underlines a data freshness gap. For rapidly changing metrics like urban population percentages, ChatGPT scored 8/10, Claude 6/10.

FAQ

Q1: Which AI model is better for geography homework?

ChatGPT (GPT-4 Turbo) scored 84% overall vs. Claude 3.5 Sonnet’s 76% on our 50-question test. For map projection analysis, ChatGPT’s accuracy reached 90%, while Claude’s was 60%. If your homework involves coordinate reading or route planning, ChatGPT is the safer choice. Claude still handles landmark identification well (100% on unambiguous descriptions) but lags on temporal precision and disputed boundaries.

Q2: Can these models replace a human geography teacher?

No. Both models made factual errors — Claude misstated the Amazon River mouth latitude by 0.5°, and ChatGPT omitted the architect name for the Blue Mosque. A teacher would catch these. The OECD PISA 2022 framework requires students to explain spatial relationships, not just recall coordinates. Neither model can replicate the pedagogical scaffolding a human provides. Use them as study aids, not replacements.

Q3: How often are the models’ geography databases updated?

ChatGPT (GPT-4 Turbo) uses training data up to April 2024, while Claude 3.5 Sonnet’s cutoff is January 2024. This 3-month gap explains Claude’s error on Brazil’s population (2022 census vs. 2024 estimate). For dynamic topics like border disputes or census data, the newer model has a measurable advantage. Always cross-check with sources like the UN World Population Prospects or National Geographic.

References

• OECD. 2022. PISA 2022 Spatial Reasoning Framework.
• National Geographic Society. 2023. Geo-Inquiry Process Handbook.
• United Nations. 2024. World Population Prospects 2024.
• Cartographic Society of America. 2021. Choropleth Mapping Standards.
• UNESCO. 2023. World Heritage List Statistical Review.