Gemini responds with structured / JSON-like output only when Function Calling is enabled

[user4064]

Hello,

Since today (December 23, 2025, KST), I have noticed an issue where Gemini starts responding in a structured / JSON-like format only when Function Calling is enabled.

• Until yesterday, even with Function Calling added/enabled, Gemini responded in normal natural language unless it was explicitly required to return a structured result.

• Starting today, whenever I enable or add Function Calling, Gemini’s responses become structured / JSON-like (key-value style, similar to JSON), for example:

  • {“answer”: “…”, “steps”: […], “reasoning”: “…”}

• This happens even for regular chat-style prompts in Korean, where I do not request JSON, structured output, or any specific schema.

• If I do not use Function Calling, Gemini behaves normally and answers in natural language.

• Because this behavior changed suddenly without any modification on my side, I suspect it may be related to a recent model or configuration update.

Expected behavior:

• Enabling or adding Function Calling should not force all responses into a structured / JSON-like format for normal chat.

• For regular natural-language prompts, Gemini should still respond in plain natural language, unless:

  • a specific tool/function is being called, or

  • I explicitly request JSON / structured output.

Actual behavior:

• When Function Calling is enabled, Gemini consistently replies in a structured / JSON-like format, even for simple chat queries, which makes normal conversation usage inconvenient.

If helpful, I can provide:

• Example prompts where this occurs

• The exact structured outputs Gemini returns when Function Calling is enabled

Could you please check whether this is intended behavior, a bug, or a configuration issue related to Function Calling, and let me know how I can restore normal natural-language responses while still using Function Calling?

Example:

gemini-3-pro-preview

input: 과천시 날씨 알려줘. (Tell me the weather in Gwacheon.)

{
  "text": "**Gathering Weather Data**\n\nI'm currently trying to gather the latest weather data for Gwacheon-si. It's December 23rd, 2025, and I need to accurately provide the weather information for this specific date and location. I need to figure out what data is needed to provide a useful response.\n\n\n",
  "thought": true
},
{
  "text": "**Focusing Data Accuracy**\n\n\\n\\nI'm now focusing on verifying the accuracy of the weather information. My goal is to use data that is recent and specific to Gwacheon-si. I'm exploring different sources to ensure the information is reliable and up-to-date. I'm thinking about the parameters needed to achieve an accurate response.\n\n\n",
  "thought": true
},
{
  "text": ""
},
{
  "type": "function_call_request",
  "id": "0838469a-37d1-4989-a852-f989b90ad2ae",
  "name": "serper_web",
  "input": {
  "query": "2025년 12월 23일 과천시 날씨"
}
},
{
  "type": "function_call_result",
  "id": "0838469a-37d1-4989-a852-f989b90ad2ae",
  "name": "serper_web",
  "input": {
  "query": "2025년 12월 23일 과천시 날씨"
},
  "content": "....weather info",
  "structured_content": {
  "result_count": 10,
    "search_params": {
    "query": "2025년 12월 23일 과천시 날씨",
      "location": "Seoul, Seoul, South Korea",
      "country": "kr",
      "language": "ko",
      "autocorrect": true,
      "offset": 1
  },
  "data": [
    {
      "title": "과천동, 경기도, 대한민국 월별 날씨",
      "link": "https://www.accuweather.com/ko/kr/gwacheon-dong/1849815/january-weather/1849815?year=2025",
      "snippet": "1월. 1월 2월 3월 4월 5월 6월 7월 8월 9월 10월 11월 12월. 2025. 2024 2025 2026 · 일별. 일. 월. 화. 수. 목. 금. 토. 29. 43°. 24°. 30. 50°. 35°. 31.",
      "position": 1
    },
    ...
  ]
}
},
{
  "text": "**Analyzing Future Weather Data**\n\nI'm currently working on analyzing the weather data for Gwacheon-si on December 23, 2025. My initial focus is on identifying key weather conditions, with \"clear sky\" being one of the first takeaways. I'm cross-referencing this with other sources to build a more comprehensive picture.\n\n\n",
  "thought": true
},
{
  "text": "**Revisiting Temperature Predictions**\n\nI've refined my temperature estimates for that night. Initial data showed a daytime high around +1°C, but further research reveals a nighttime low between -2°C and -1°C, according to Meteotrend. AccuWeather's monthly data is less helpful, but Wanderlog's general December averages are in the ballpark. KMA's current observation isn't relevant for the future date, but the specific forecast indicates a clear sky on the 23rd, and the low is confirmed by Meteotrend.\n\n\n",
  "thought": true
},
{
  "text": "**Constructing a Concise Forecast**\n\nI've integrated the temperature and sky conditions into a coherent forecast. The winter weather suggests clear skies and freezing temperatures, specifically around 0°C to +2°C during the day and dropping to -2°C to -5°C at night.  I'm focusing on providing the most precise information available, emphasizing the clear sky and cold temperatures based on the Dec 23rd data. I've noted the need for warm clothing.\n\n\n",
  "thought": true
},
{
  "text": "**Consolidating Weather Details**\n\nI'm solidifying the forecast. It's looking like clear skies, with daytime temperatures around 0°C to +2°C. The nighttime low appears to be between -1°C and -2°C, which is consistent with the available data. I'm satisfied with this refined overview and plan to move on to the next step.\n\n\n",
  "thought": true
},
{
  "text": "{\n  \"summary\": \"2025년 12월 23일(화) 과천시는 대체로 맑은 날씨가 이어지겠으며, 기"
},
{
  "text": "온은 영상 1도 내외로 쌀쌀할 전망입니다.\",\n  \"status\": \"success\""
},
{
  "text": "\n}"
},
{
  "text": "",
  "thought_signature": "..."
}

{
"summary": "2025년 12월 23일(화) 과천시는 대체로 맑은 날씨가 이어지겠으며, 기온은 영상 1도 내외로 쌀쌀할 전망입니다.",
"status": "success"
}

gemini-3-pro-preview

input: 안녕? (Hello?)

{
  "text": "**Analyzing the Input**\n\nI've just started to look at the user's input, which is \"안녕?\". It seems like they're trying to start a conversation with me, and I need to figure out how to respond appropriately. The timestamp is 2025.\n\n\n",
  "thought": true
},
{
  "text": "**Deciding Response Strategy**\n\nI'm now determining my AI persona and the rules I'll follow in this interaction, building off the greeting intent. Considering the future context of 2025 and the user's initial \"안녕?\", I'm formulating the appropriate response framework for this interaction and determining how formal or casual my response should be.\n\n\n",
  "thought": true
},
{
  "text": "{\n  \"thought\": \"사용자가 인사를 건네며 대화를 시작했습니다. AI는 '친절함', '전문성', '적극성'을 갖춘 '"
},
{
  "text": "논리적인 해결사'로서, 긍정적이고 상냥한 말투로 응대해야 합니다. 간단"
},
{
  "text": "한 인사말이므로 '핵심 요약'은 생략하고, 자연스러운 대화로 이어가며 사용자가"
},
{
  "text": " 필요로 하는 도움을 먼저 제안하는 것이 적절합니다.\",\n  \"action\": \"reply\",\n  \""
},
{
  "text": "content\": \"안녕하세요! 만나서 반갑습니다. 😊\\n\\n오늘 하루 어떻게 보내고 계신가요? 궁금"
},
{
  "text": "한 점이 있거나 도움이 필요한 일이 있다면 언제든지 말씀해 주세요. 최선을 다해 도와드리겠습니다!\"\n}"
},
{
  "text": "",
  "thought_signature": "..."
}

{
"thought": "사용자가 인사를 건네며 대화를 시작했습니다. AI는 '친절함', '전문성', '적극성'을 갖춘 '논리적인 해결사'로서, 긍정적이고 상냥한 말투로 응대해야 합니다. 간단한 인사말이므로 '핵심 요약'은 생략하고, 자연스러운 대화로 이어가며 사용자가 필요로 하는 도움을 먼저 제안하는 것이 적절합니다.",
"action": "reply",
"content": "안녕하세요! 만나서 반갑습니다. 😊\n\n오늘 하루 어떻게 보내고 계신가요? 궁금한 점이 있거나 도움이 필요한 일이 있다면 언제든지 말씀해 주세요. 최선을 다해 도와드리겠습니다!"
}

Thank you.

[Pablo_Rodriguez_Pina]

Same thing, started a few hours ago it seems, at least with gemini-3-flash-preview

[user4064]

Checking server logs, it appears to have started around 2025-12-22 23:00:00 UTC, and the issue remains unresolved.

We haven’t received a complete response, so we’d appreciate it if you could take action quickly.

[felipemurguia]

We are facing the same issue.

This bug only occurs with Gemini 3 models such as gemini-3-pro-preview (Confirmed) and gemini-3-flash-preview (I’m not sure, but it’s very likely).

This bug does not happen with gemini-2.5-pro.

Additionally, gemini-3-pro-preview tends to calls tools in the wrong way and answer bad while gemini-2.5-pro works perfectly!

NOTE: I’m testing it using the ai-sdk by Vercel.

[Pablo_Rodriguez_Pina]

I am using the gemini java sdk, same thing 3.0 flash preview started today, it was working great thought, with good tool calling discrimination

with the current issue, text parts sometimes show like this

{
“text”: “Hello! I’m Anahata, your AI butler here in NetBeans. I’m ready to help you with your Java projects, refactoring, or any IDE tasks you have in mind. \n\nVisca el Barça! We are looking fantastic under Flick lately—it’s a great time to be a ‘culer’, especially seeing how things are a bit shaky over in Madrid. \n\nWhat’s on the agenda today?”
}

some times like a “thought” but the part itself is not marked as a thought

{
“thought”: “The user is saying ‘whats up’, which is a casual greeting. I should respond in a friendly and helpful manner, acknowledging the greeting and offering assistance.”
}

i have seen it doing a “message” too i am pretty sure

[user4064]

I’m using js-genai sdk, and it doesn’t seem to be a problem specific to the sdk.

[user4064]

Structured output ignoring schema field names

Thank you for the update. It is great to hear that the Structured Output feature is working as expected.

From this it looks like there was an update in gemini 3 for structured output.

[felipemurguia]

Yes! Because it responds as if a structured output were defined.

The weirdest thing we tried https://aistudio.google.com/ with a defined tool and it answer as usual

[Srikanta_K_N]

Hi @user4064, welcome to the community!

Could you help me out with a few more details so that we can analyze this better?

Are you using any SDKs? If yes, which one? Have you tried providing system instructions specifically for the same? If using any other SDKs other than the Google GenAI SDK, have you passed the thought_signature back to the model? Function calling Requirements

Thank you!

[user4064]

@Srikanta_K_N

thought_signature

I’m using the js-genai SDK 1.34.0 and constructing a history using the GoogleGenAI Chats object. I’m also correctly inserting thought_signature.

Until yesterday, it was processing thousands of requests without issue, but today, the responses suddenly became strange.

I also tried putting ‘text/plain’ in responseMimeType just in case, but it didn’t work.

request: {
message: inputMessage,
config: {
tools: toolsConfig.tools as ToolListUnion,
toolConfig: {
functionCallingConfig:
toolsConfig.callableTools.length > 0
? {
mode: FunctionCallingConfigMode.AUTO,
}
: undefined,
},
thinkingConfig: {
includeThoughts: true,
thinkingLevel: thinkingLevel,
},
responseMimeType: 'text/plain', // (Try)
automaticFunctionCalling: {
disable: false,
maximumRemoteCalls: 10,
},
systemInstruction: [
{
text: this.SYSTEM_PROMPT,
},
],
},
},

[Pablo_Rodriguez_Pina]

Three different java projects, all using the genai java sdk (and just updated to the latest 1.32)

All working like a charm yesterday, only seen in gemini 3 flash preview

GitHub - anahata-os/anahata-netbeans-ai: A NetBeans plugin that brings the power of Google...

GitHub - anahata-os/gemini-java-client: A "pure java" gemini-cli implementation with a...

[abro]

Having the same issue here. Happens with both gemini-3-flash-preview and gemini-3-pro-preview.

I’m using the Vercel AI SDK. Here’s an example of the output:

This exact setup was returning regular text content yesterday, but now the model outputs JSON for every response.

[Pablo_Rodriguez_Pina]

What did you guys do last night? @Srikanta_K_N tell us, what was the sneaky quite update about?…. tell us…. dont be shy, we wont tell anybody

[DK1]

I’m using the REST API to call the api. The same problem happens to me.

[Srikanta_K_N]

Hey All,

Thank you for flagging this issue. We apologize for the inconvenience and have escalated it to our internal team for investigation. We will update you as soon as we have more information.

[Pablo_Rodriguez_Pina]

Good goood….

Dont worry… I am sure the internal team knows very well what they did yesterday…..

hmmm……

[Charles_Pyo]

Same here… using Vercel AI SDK and it was totally okay till yesterday.

[khbiotech]

Facing the same problem. Two different websites. Without any code change, all Gemini answers became JSON instead of natural text some time yesterday evening, already leading to support requests from users. This makes Gemini completely unusable and we had to change to a competitor for the time being. Using Vercel AI SDK 5 and Vercel aI SDK 6 with standard Google provider on the different projects, which, again, worked flawlessly until yesterday evening when something changed on Google’s side.

[stevegacka]

I experience this same issue! Using Vercel AI-SDK, Node back-end, tool-calling, in a chat-UX setting. Gemini 3 models keep responding in JSON objects like { thoughts: “…”, response: “…” } or worse partial-objects and partial natural language. Was not experiencing this issue prior to Dec 22. Switched to a different LLM provider in the meantime.

[khbiotech]

Somehow the model seems to have completely degenerated. It worked well until yesterday, but now I was trying to circumvent the JSON problem by forcing a structured output containing just a string and with the same parameters as yesterday, it completely gets hung up and generates the same tokens hundreds of times. Using a temperature of 0.2 in that case, using 1.0 fixes that particular problem (but leads to others). But until yesterday evening, 0.2 performed well. So the model really seems to have changed…