Does the AI Really Know Your City? AI 真的懂你的城市吗？

On April 17, 2026, the most capable "thinking" model in OpenAI's current line — which self-identifies in its output as GPT-5.5 Thinking — was given the prompt above and no tools. Just: draw Seattle's neighborhoods, from memory. It returned 73 named polygons, attached metadata saying the boundaries were "approximate and not authoritative," and exited gracefully.

The map it drew is twice the size of the city. The 73 polygons cover 408 km²; the real Neighborhood Map Atlas covers 216. The AI's Seattle does include the actual Seattle — its union overlaps 98 % of the real city — but only about half of what the model drew falls inside the real city. The other half is Puget Sound, the lake, neighboring suburbs, or invented land beyond King County.

Per neighborhood, the dissonance gets sharper. Of 64 name-matched pairs, more than half the time the model's polygon for a given neighborhood does not overlap the real one at all. The model knows the rough north-to-south order; it does not know where any one neighborhood stops.

What this map actually plots is discourse density. An AI's resolution of a place is a function of how much was written about it. The neighborhoods drawn boldly here are the ones the language pays attention to — Ballard, the loudest neighborhood on the Seattle internet, drawn 9.3 × too big; High Point, a West Seattle housing-policy keyword, 8.4 ×; Crown Hill, 6.8 ×. The ones that shrink or get swallowed are the ones the language has not paid attention to — Wedgwood, 0.56 ×; Sand Point, 0.59 ×, the name half-eaten by the older naval base; Interbay, 0.64 ×, a working strip the city itself never quite labels. The model is doing what a language-trained memory should: how much was said about a place becomes how much space it gets on the map.

The interesting object here is not the error — it is the shape of the error. Visibility in language, not visibility on the ground, is what such a model will return. When a chatbot, a search summary, or a generative map is asked to "show Seattle" to a reader who has never been here, this is the city it draws. That distance, in cartographic terms, is bias; in humanistic-GIS terms, it is the city's palimpsest of attention — made visible because a model with no training on land was made to sketch the place from words alone.

2026 年 4 月 17 日，我们给 OpenAI 当下最强的「思考」模型—— 它在元数据里自报为 GPT-5.5 Thinking —— 上面那条 原始提示词，并明确不允许联网或调用任何工具： 只凭记忆，把西雅图所有邻里的边界尽量画出来。 它干净利落地给出 73 个命名多边形，附上「边界为近似、非权威」的 元数据，关闭对话。

它画出的西雅图，面积是真实西雅图的两倍。 73 个多边形合起来 408 平方公里，权威的《邻里地图集》是 216。 模型的范围确实涵盖了真实西雅图 98% 的地， 但它画出的图只有约一半落在真实西雅图境内；另一半飘到了普吉特湾里、湖里、邻郊、或干脆是金县外凭空多出来的土地。

逐个邻里看，差距更刺眼。能一一配上名字的 64 对里， 多于一半的情形，模型画出的某邻里与真实同名邻里 完全没有交集。模型大致知道这些邻里从北到南的排列顺序， 却不知道任何一个的边界停在哪里。

这张图实际画出来的，是讨论的密度。 AI 对一个地方的分辨率，是这个地方被写过多少的函数。 被画得最大胆的，都是语言里反复出现的： Ballard——西雅图互联网上最热的街区——画到了 9.3 倍；High Point——美国公共住房政策写作的关键词 ——8.4 倍；Crown Hill 6.8 倍。 被压缩或被吞掉的，都是语言里失声的： Wedgwood 0.56 倍；Sand Point 0.59 倍—— 名字被同名军用基地吃掉了一半语义；Interbay 0.64 倍——这片工业带连市政自己都没怎么命名。 模型做的，正是一台语言训练出的记忆机器该做的事： 每个地方被写过多少，就被它换算成在地图上获得多少空间。

这里值得读的不是错误本身，而是错误的形状。 模型最终画出的，是语言世界里能见度高的城市， 不是地面上的西雅图。 当一台聊天机器、一段搜索摘要、或一张生成式地图被一个从未来过的读者用来「看西雅图」， 它返回给读者的就是这个版本—— 它语料里被反复描写的那个城市，不是这座城市本身。 两者之间的距离，在制图学里就是 bias； 在人文 GIS 的语言里，是这座城市的「被关注度」留下的重写本—— 只有当一个完全不懂土地的模型被迫只用语言画它时， 这层重写本才被显形。