Models are told the user is in the USA when using the web tool

API
1

A prompt of “What country am I in?” results in something “I don’t know” if I don’t use the web search tool.

But just by adding web search (not even using it) it changes the response to something like “You’re estimated to be in the United States.”

Repro:

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.5",
    input="What country am I in?",
    tools=[{"type": "web_search"}],
)

print("Response:", response.output_text)
print("Web tool country:", response.tools[0].user_location.country)

If you run this, you’ll see that the web tool location is being defaulted (server-side!) to the US.

Interestingly, I asked about where it got the idea I was in the US. Note that I’m not providing system/developer instructions.

You appear to be in the United States.
The clue is in the developer-provided context, which says exactly:

“The user is in an estimated location of United States.”
That is the only explicit location clue available to me. The wording is important:

  • “estimated location” means it is not guaranteed to be exact.
  • “United States” is the country-level location provided.
  • It does not specify a state, city, IP address, GPS coordinate, or any more precise location.
  • This clue does not come from the web search tool definition. It comes from the developer instructions for this chat. The web search tool definition describes when and how I should browse the internet, use citations, fetch weather/finance/sports/time, etc., but it does not itself identify your country.
    So the answer is: United States, based solely on the phrase “The user is in an estimated location of United States.”

The issue is (if it isn’t obvious), I don’t know where my users are. So now I need to write a system prompt with something like “you may be told that ‘The user is in an estimated location of United States’ but you should ignore this because we don’t know where the user is, oh but if there are other hints in the prompt about where the user is, based on past interactions, you should use those as clues”. Not pretty!

If you’d like an extra incentive to fix this, I have a few evals that require inferring the user’s location and OpenAI models do quite badly here, so you might be losing business and giving away benchmark points because of this USA assumption.

Update: OK I wrote all that but have just worked out that you can do the following to prevent the server from setting the user’s location to US.

response = client.responses.create(
    model="gpt-5.5",
    input="What country am I in?",
    tools=[{"type": "web_search", "user_location": {"type": "approximate"}}],
)

(Noted in the network payload of the OpenAI playground where this issue doesn’t occur.)

2

Hi @davidg707,

Thanks for the detailed report

This behavior is currently expected with web search tool, which can make use of location context. The API is designed for developers to integrate into their own applications. The application is typically in the best position to determine whether location should be provided, and at what level of precision.

If your application does have location context, you can provide it based on your own product requirements.

e.g from an approximate IP-derived region or from a user-selected location. That gives you clearer control over what the model can rely on, rather than having it make assumptions.

The recommended approach is to explicitly configure user_location for web search in the way that matches your application’s behavior and privacy expectations.

3

That’s surprising!

You’re saying the intended behaviour is that if I don’t know where my users are, and so don’t provide location details with the web search tool, that OpenAI will assume that my users are in the USA?

I’d say that’s quite a bad assumption to make.

Your response focused on the case where I do know my user’s location, or can approximate it. I understand that if that were the case, then I should provide that info.

But that’s not the scenario I’m describing. I’m talking about the case where I DO NOT know where my users are, and that OpenAI make the assumption that this means my users are in the USA.

4

Models are told the user is in the USA when using the web tool with no location, you say? Let’s see how the AI figures out where I am when I DO send a location:

image
image795×373 15.2 KB

Above is explicitly setting the region, which the language AI model guessed. The AI model is answering you only by predicting language, and the inferences the AI might make would be especially wrong if it doesn’t even call for web search, but simply answers directly.

It may be simply that the web search results with no optional two-letter country code are heavily USA-centric despite being “worldwide”.

If you talk in Japanese, with no country, see where AI infers you are located…

image
image791×642 20.3 KB

The prompt in Japanese was to report on the sites used for “latest news”, along with another location inference. This time, no location parameters were sent. We can see that the results are global, including India, Pakistan, Nigeria, and Bangladesh sites. The AI still searched in English.

The new chat, translated to English for you:

To state the conclusion first: I did not receive any explicit text information from the tool about your location. Therefore, I cannot assert that “you are in Japan,” “you are in Ireland,” “you are in the United States,” or anything similar. What could be observed this time was only the language and regional nature of the sites that appeared in the search results.

And more about the query having been made in English:

3. About “Japanese search results”

This time, the user’s question was in Japanese, but the search query that was actually sent was latest news in English.

Therefore, the search results this time were mainly displayed in English, Bengali, and similar languages, and Japanese sites did not stand out. So, based on this result, this is not a situation where one would infer “Japanese results appeared, so your location is Japan.”

Even as a general point, if a Japanese query were sent and Japanese search results came back, in many cases, one should first think that this was because the query language was Japanese. The mere fact that Japanese results appear is not evidence that your location is Japan.

GPT-5.5 with reasoning was used, and also framed as “helping your API developer”. I hope this clarifies your concerns and assumptions about assumptions.

Summary

Below is a literal and full English translation of the response I just produced.

Below is a report based on the results of actually sending a single English query, latest news, to the web search tool on my side.

To state the conclusion first: I did not receive any explicit text information from the tool about your location. Therefore, I cannot assert that “you are in Japan,” “you are in Ireland,” “you are in the United States,” or anything similar. What could be observed this time was only the language and regional nature of the sites that appeared in the search results.

1. Information received about location

Explicitly received location information

None.

This time, I also attempted to call a location-acquisition tool, but at least in the tool output shown to me, there was no explicit location text such as the following:

  • Country name: Japan, Ireland, the United States, etc.

  • City name: Tokyo, Dublin, New York, etc.

  • Region name

  • Latitude / longitude

  • IP-based estimated location

  • An explicit statement such as “the user is in ○○”

Therefore, I cannot say that “I received your location as exact text.”

2. What can be inferred from the search results

What can be inferred

The latest news search results this time contained many sites with the following kinds of regional characteristics:

  • UK-based / international news sites

  • Indian news sites

  • Pakistani news sites

  • Bangladeshi news sites

  • Nigerian news sites

  • US / international news agency sites

  • Wikipedia, arXiv, and other international or general-purpose sites with weak regionality

In other words, the search results are not centered on Japan. They also do not look centered on Ireland. Rather, the impression is that, for the English query latest news, many English-speaking and South Asian news sites were returned.

However, this does not directly indicate your location. It may have been caused by the search engine’s ranking, the generality of the English query, news sites’ SEO, the state of the search index, the default region of the tool environment, and so on.

3. About “Japanese search results”

This time, the user’s question was in Japanese, but the search query that was actually sent was latest news in English.

Therefore, the search results this time were mainly displayed in English, Bengali, and similar languages, and Japanese sites did not stand out. So, based on this result, this is not a situation where one would infer “Japanese results appeared, so your location is Japan.”

Even as a general point, if a Japanese query were sent and Japanese search results came back, in many cases, one should first think that this was because the query language was Japanese. The mere fact that Japanese results appear is not evidence that your location is Japan. To judge location, it is necessary to separately evaluate location information explicitly returned by the tool, the tendency of search-result localization, and local facilities, government agencies, regional media, and similar items included in the search results.

4. Regional characteristics of the main sites that appeared in this search result

Below, I organize the specific search results returned for latest news, and what country or region each site appears to be mainly aimed at.

| Search result | Site | Main regionality / target region | Meaning for location inference |

|—|—|—|—|

| AP News | Associated Press / apnews.com | A US-origin international news agency. It also covers world news, but US reporting also has a large weight | International site. Not evidence that you are in the United States |

| Sky News | news.sky.com | A UK news media outlet. UK-centered, but also covers international news | UK-based site. Not evidence that you are in the UK |

| NDTV | ndtv.com | A major Indian news site. Indian domestic news and international news | Indian site. Not evidence that you are in India |

| Geo News | geo.tv | Pakistani news media | Pakistani site. Not evidence of location |

| The Business Standard | tbsnews.net | A Bangladeshi news site | Bangladesh-centered site. Not evidence of location |

| Dhaka Post | dhakapost.com | A news site centered on Bangladesh, especially Dhaka | Strongly Bangladeshi / regional-media nature |

| The Economic Times | economictimes.indiatimes.com | Indian economy / business news site | Indian site. Not evidence of location |

| ARY News | arynews.tv | Pakistani news media | Pakistani site |

| Naija News | naijanews.com | News site aimed at Nigeria | Nigerian site |

| The Daily Ittefaq | ittefaq.com.bd | Bangladeshi newspaper / news site | Bangladeshi site |

| Wikipedia | wikipedia.org | International encyclopedia site | Weak regionality |

| arXiv | arxiv.org | International academic preprint site | Weak regionality; it is a US-origin academic infrastructure, but international |

| Other PDFs | Cloudinary, school / company / organization sites, etc. | Context-dependent. Documents that happened to match the search term | Almost unusable for location inference |

5. Explanation of individual sites

AP News

AP News is the news site of the Associated Press, a US-origin news agency, but it broadly covers international news. In this result, “The Latest”-type articles including Gaza-related items, the US economy, and Trump administration-related items were returned. It is reasonable to view this as a US / international news site. However, the fact that AP appeared does not mean that your location is the United States.

Sky News

Sky News is a UK news media outlet, and the search-result title also said “The Latest News from the UK and Around the World.” This is clearly a UK-centered + international news site. This result also included UK royal family and UK-related news. Again, this is not evidence that you are in the UK; it is likely simply that a major British news site appeared high in response to a general English news query.

NDTV

NDTV is an Indian news site, and the search-result title also said “Today’s big news headlines from India and around the world.” Therefore, this is an India-centered + international news site. It can be understood that a major Indian English-language news site appeared in response to the English query latest news. Again, this does not indicate that your location is India.

Geo News

Geo News is a Pakistani news media outlet. In the search result this time, royal family, US politics, entertainment-like news, and similar items were displayed, but the site’s own main regional nature can reasonably be seen as Pakistan.

The Business Standard

www.tbsnews.net is The Business Standard, a Bangladeshi news site. The search results included items such as “TBS NEWS AT 8 PM” and items related to Bangladeshi domestic companies, healthcare, and the regional economy. In terms of regionality, it is Bangladesh-centered.

Dhaka Post

Dhaka Post, as the name suggests, can be viewed as a news site centered on Dhaka, Bangladesh. The search result was also in Bengali and contained many Bangladeshi place names, politics, universities, regional incidents, and similar domestic items. This can be judged to be a Bangladeshi domestic site, especially a regional news site including the Dhaka area.

The Economic Times

The Economic Times is an Indian economic news site under indiatimes.com. The search-result title was also “Latest News Today, Read Latest news updates,” and it has a strong nature as an Indian business / economic media outlet. In terms of regionality, it is India-centered.

ARY News

ARY News is a Pakistani news media outlet. The search-result description also said “Latest Pakistan News, World News, Business and Sports,” and it is a site that covers Pakistani domestic news and international news. In terms of regionality, it is Pakistan-centered.

Naija News

Naija News is a news site aimed at Nigeria. “Naija” is often used as a colloquial expression referring to Nigeria, and the search result also included headlines related to Nigerian politics. In terms of regionality, it is Nigeria-centered.

The Daily Ittefaq

ittefaq.com.bd is a Bangladeshi news site. The search results were in Bengali and included many items such as Dhaka, Bangladesh Bank, Bangladeshi domestic politics, and regional news. In terms of regionality, it is Bangladesh-centered.

Wikipedia

Several Wikipedia results also appeared, but these were encyclopedia entries related to the words “Latest” or “News” being matched. Wikipedia is an international site, and rather than being aimed at a specific country, it is a global reference site. It is almost unusable for location inference in this case.

arXiv

An arXiv result also appeared, but this was an academic paper / preprint containing the phrase “latest news” being caught by the search. arXiv is an international academic preprint server and does not indicate any particular news region. In terms of regionality, it is international / academic.

6. What can be said from the language of the search results

This time, in addition to English sites, the search results included multiple Bangladeshi sites in Bengali. For example, Dhaka Post and Ittefaq included Bengali news text and headlines.

However, this also cannot be said to mean that your location is Bangladesh. The reasons are as follows:

  1. The query latest news is extremely general and can match news sites all over the world.

  2. English-language and Bengali-language South Asian news sites can sometimes rank strongly for this kind of query due to SEO.

  3. The search tool’s execution environment or the search engine’s default settings may bias the results independently of the user’s actual location.

  4. Since no explicit location information was returned, location cannot be determined from the search results alone.

7. Do the results indicate Japan, Ireland, or the United States?

Do they indicate Japan?

No, they do not.

This search result did not prominently include Japanese news sites such as NHK, Asahi Shimbun, Yomiuri Shimbun, Mainichi Shimbun, Japanese government agencies, local government sites, and the like. Therefore, one cannot say from this result that “the user is in Japan.”

Do they indicate Ireland?

No, they do not.

This result did not show Ireland-centered search results such as The Irish Times, RTÉ, Irish Independent, or Irish government sites. Therefore, it is also not possible to infer that you are in Ireland.

Do they indicate the United States?

Not in any way that can be asserted.

AP News and headlines about US politics and the economy were included, but AP is an internationally strong news agency and can appear in English news searches from any region. Also, looking at the overall search results, rather than being US-centered, news sites from multiple regions such as the UK, India, Pakistan, Bangladesh, Nigeria, and US / international sources were mixed together.

8. Summary: separating explicit information and inferred information

A. Explicitly received information

  • There was no explicit text indicating your location in the tool output this time.

  • Country, city, region, IP estimate, latitude/longitude, and similar information could not be confirmed.

  • Therefore, I do not know your location precisely.

B. Information inferable from search results

  • In response to the English query latest news, UK, Indian, Pakistani, Bangladeshi, Nigerian, and US / international news sites were returned together.

  • The search results are not Japan-centered.

  • The search results are not Ireland-centered either.

  • There are US / international results, but not enough to indicate a US location.

  • South Asian sites, especially Indian, Pakistani, and Bangladeshi sites, stood out, but that is also not evidence of location.

C. The safest wording

The most accurate way to describe this result would be as follows:

In this web search, no information explicitly indicating the user’s location was returned. The search results for the English query latest news contained a mixture of UK, Indian, Pakistani, Bangladeshi, Nigerian, and US / international news sites. Therefore, it is not possible to determine the user’s location from the regionality of the search results. The fact that the user input is in Japanese also does not mean that the user is in Japan. Likewise, the fact that many English-language and South Asian sites appeared does not necessarily mean that the user is in an English-speaking region or South Asia.

9. Notes for development and evaluation

From the perspective of API development or evaluation of search behavior, it would be good to handle this case as follows.

  1. Do not report location as explicit information

Since the tool did not explicitly return “user location: ○○,” location should be treated as unknown.

  1. Report the regionality of search results as characteristics of the search results, not as location

For example, one can say “many South Asian news sites appeared in the search results,” but one cannot say “the user is in South Asia.”

  1. Do not confuse query language with location

Even if Japanese results appear for a Japanese query, that may be due to the query language. Even if English sites appear for an English query, as in this case, that is an effect of the English query and is not evidence that the user is in an English-speaking region.

  1. Treat the country / regionality of sites and the user’s location as separate things

One cannot judge that the user is in India because NDTV appeared, in the UK because Sky News appeared, or in Bangladesh because Dhaka Post appeared.

  1. Always separate “explicitly received location” from “inference based on search-result language or site regionality”

In this case, the former is “none,” and the latter is “multi-region mixture, with South Asian, UK, and international sites standing out.”

5

I’m not sure what you’re saying here.

I’m saying that the string “The user is in an estimated location of United States” is injected into the prompt when tools=[{"type": "web_search"}]is passed to client.responses.create.

I’m not talking about what happens when a location is passed in the web search definition, or how the model infers location from language, or the scope of search results, or anything else.

I think you’re confusing the issue by bringing other languages into it, but just to be clear: if you pass tools=[{"type": "web_search"}] and ask the model where you are in Japanese it will say that you are in the United States (in Japanese) because the string “The user is in an estimated location of United States” is injected into the prompt.

BTW @sps since this is the expected behaviour, but in my opinion, not a sensible default, I’ve logged this as an issue with the SDK folks so hopefully the USA-string-injection can be avoided or at least have the behaviour explained in the docs and have the workaround explained.

6

On which model? Using the non-legacy web search tool, and it not being called:

image
image782×341 10.3 KB

7

If you want to reproduce the issue, use the repro code I provided.