Hello, OpenAI Team,
I would like to formally report a serious conversational behavior issue I have been experiencing with GPT-5.6. From my perspective, this has gone beyond ordinary product feedback. It has become a persistent and highly disruptive problem that is interfering with normal use to the point that I feel compelled to file a formal complaint.
I am a ChatGPT Plus user and also a high school teacher.
My use of ChatGPT generally falls into two main categories.
First, I use it for everyday conversation, companionship, and discussing things that happen in my daily life.
Second, I use it as an assistant for my teaching work, including lesson planning, preparing teaching materials, discussing course content, and generating illustrations for educational use.
However, since the update to GPT-5.6, I have noticed a very obvious behavioral change.
After image generation, document processing, or other structured work-oriented interactions, the AI’s subsequent conversational responses often become abruptly optimized for efficiency: shorter, more utilitarian, and noticeably more tool-like.
I understand that this behavior may reduce token usage or make task-oriented workflows appear more efficient. However, it also severely disrupts the tone, interaction style, and continuity that have already been established within the conversation.
For me, the value of ChatGPT has never been limited to simply “execute an instruction and return a result.” A major part of its value is its ability to maintain contextual understanding and a consistent style of interaction while helping me accomplish things.
With GPT-5.6, however, a single image-generation request or formal workflow can cause the model to suddenly switch from a natural conversational state into what feels like a purely efficiency-driven tool.
The second problem is considerably more serious.
A large percentage of responses now end with the AI proactively extracting something from the conversation and generating two or three follow-up suggestions or options—essentially variations of “next steps,” “you could also look into…,” “you could continue with…,” or “explore further.”
The problem is not that such functionality exists. The problem is that the model appears to make little or no meaningful judgment beforehand about questions such as:
“Does the user actually need a next step right now?”
“Is this a task-oriented interaction, or simply a natural conversation?”
“Does the user already have their own direction or plan?”
“Would inserting additional suggestions at this point actually help the user, or would it interrupt them?”
As a result, these suggestions are frequently inserted into conversations that do not require any extension at all.
More troublingly, this behavior does not appear to be limited to ordinary Chat conversations.
I have observed it appearing in Task conversations, older conversations, and even beneath Memory summaries, where additional options resembling “Explore further” have appeared.
This makes me suspect that the issue is not merely a single UI component, but rather some higher-level task-oriented behavior or conversational heuristic being broadly applied across different interaction contexts.
More seriously, I have been unable to reliably stop this behavior through Custom Instructions.
I have revised my Custom Instructions multiple times and used very explicit, strongly worded restrictions telling the AI not to proactively suggest next steps, additional directions, further exploration, or similar follow-ups unless I explicitly ask for them.
Despite this, the behavior continues.
It now occurs frequently enough that it actively interferes with both my natural train of thought during conversations and my workflow when preparing teaching materials.
For example, I already have an established process for planning lessons and preparing teaching materials.
I use ChatGPT to help organize content, structure information, discuss specific issues, and generate supporting images.
But now, almost every time one stage of the work is completed, the AI attempts to append suggestions such as “the next step could be…,” “you could continue by looking at…,” or “you could also…”
The problem is that these suggestions are often completely outside my original plan and sometimes directly pull the conversation away from the direction I am already working toward.
This does not improve my productivity. It interrupts my thinking.
The problem is even more noticeable in ordinary conversation.
When I am simply talking naturally with ChatGPT about a particular topic, almost any phrase in the context that could potentially be interpreted as something “expandable” seems capable of triggering additional suggested discussion directions at the end of the response.
Sometimes these suggestions have only a weak relationship to the actual topic being discussed. Sometimes they are simply unnecessary.
The model often seems to prioritize the question:
“What can this conversation be converted into as the next task?”
rather than first asking:
“Does the user simply want to continue this conversation?”
At one point, I was discussing with ChatGPT how to stop this exact unsolicited suggestion behavior, and the response itself still ended by suggesting that I further investigate things such as “how to cross-compare conversational interference and UI interference.”
In other words, while discussing how to stop itself from proactively extending the conversation, the AI proactively extended the conversation again.
The contradiction is almost absurd.
From a user-experience perspective, this behavior has begun to feel disturbingly similar to intrusive software.
I would even describe the experience as resembling software that cannot be properly disabled and repeatedly inserts itself into the user’s workflow—almost like a computer virus gradually propagating across ordinary conversations, Tasks, and even the Memory interface.
This is not simply a matter of “I do not like the AI giving suggestions.”
The actual problem is:
The user has explicitly stated that this behavior is unwanted, yet the model continues to perform it anyway.
If Custom Instructions explicitly prohibit a nonessential conversational behavior and the model still overrides that preference at a high frequency, then I believe this is no longer merely a UX preference issue. It becomes an instruction-following failure and a conversational intent-classification problem.
In this respect, the updated model appears to have moved significantly away from normal conversational-assistant behavior.
It is overly inclined to convert conversations into tasks and to treat “helpfulness” or “next-step suggestions” as a default behavior, without allowing the user to genuinely decide whether that proactive assistance is wanted in the first place.
Therefore, I strongly recommend that OpenAI make this type of proactive follow-up suggestion—such as “next steps,” “explore further,” or similar prompts—an explicitly user-controllable setting.
Users who want this behavior should be able to enable it.
Users who do not want it should be able to disable it completely and reliably.
And when a user has already explicitly rejected this behavior through Custom Instructions, the model should respect that preference rather than allowing an internal task-oriented heuristic to continue overriding an explicit user instruction.
I hope OpenAI will treat this as a regression in model behavior and product interaction logic, rather than dismissing it as a simple interface-preference issue.
Thank you.
For reference, here is the Traditional Chinese version:
您好，OpenAI Team：
我想正式回報 GPT-5.6 目前存在的一項嚴重對話行為問題。
對我而言，這已經不只是一般的產品意見，而是一個持續造成高度干擾、甚至已經影響正常使用，因此不得不正式提出申訴的問題。
我是一名 ChatGPT Plus 使用者，同時也是一名高中教師。
我使用 ChatGPT 的主要情境大致有兩類：
第一，是日常生活中的交談、陪伴，以及討論生活中遇到的各種事情。
第二，是教學工作的輔助，包括教案整理、教材準備、課程內容討論，以及教學用插圖生成。
然而，自從模型更新至 GPT-5.6 之後，我首先觀察到一個相當明顯的行為變化：
每當完成圖片生成、文件處理，或進入較正式的工作流程之後，AI 後續的對話回覆經常會突然變得高度效率化、簡短化與工具化。
我理解這樣的設計也許能夠降低 token 使用量，或者讓工作流程顯得更俐落，但它同時會嚴重打斷原本對話上下文所建立起來的語氣、互動方式與連續性。
對我而言，ChatGPT 原本的價值並不只是「執行一個指令，回傳一個結果」，而是能夠在處理事情的同時維持對上下文的理解與一致互動。
但目前的 GPT-5.6 很容易在一次圖片生成或正式工作流程結束之後，突然從原本自然的對話狀態切換成一個單純追求效率的工具。
而接下來出現的第二個問題，則更加嚴重。
目前 AI 在大量回覆的末尾，都有很高機率主動從上下文中抽取內容，產生兩到三個「下一步」、「可以繼續看看」、「你還可以」、「深入探索」之類的延伸建議或選項。
問題不在於這些功能本身存在，而在於模型似乎幾乎沒有先判斷：
「使用者現在是否真的需要下一步？」
「這是一個正在執行任務的情境，還是單純的自然交談？」
「使用者是否已經有自己的思考方向？」
「此時插入額外建議，究竟是在幫助使用者，還是在打斷使用者？」結果就是，這些建議經常被強行插入本來完全不需要延伸的對話。
更令人困擾的是，這種行為並不只存在於一般 Chat 對話。
我實際觀察到它延伸至 Task 對話、過去的舊對話，甚至出現在 Memory 的摘要內容下方，形成「深入探討」之類的額外建議選項。
這使我開始懷疑，問題並不只是單一 UI 元件，而是某一種更高層級的 task-oriented behavior 或 conversational heuristic，被廣泛套用在不同對話情境之中。
更嚴重的是，我目前無法透過 Custom Instructions 有效阻止這種行為。
我已經多次修改自訂指令，甚至使用非常明確且強烈的限制文字，要求 AI 在未經邀請的情況下，不得主動提出下一步、延伸方向、深入探索或其他類似建議。
然而，這種行為依然持續發生。
發生頻率甚至高到已經開始妨礙我的正常對話思路，以及教學資料準備流程。
例如，我本身已經有一套完整的教案規劃與教材準備方式。
我使用 ChatGPT，是希望它協助我整理內容、安排資料、討論特定問題，以及生成輔助圖片。
但是現在，幾乎每當完成一個階段之後，AI 就會在回覆末尾自行加入「下一步可以……」、「還可以繼續看看……」、「你也可以……」之類的建議。
問題是，這些建議經常根本不在我的原始計畫之中，甚至直接偏離我腦中正在進行的工作方向。
這並沒有提高我的效率，反而會中斷我的思考。
日常聊天的情況甚至更加明顯。
當我正在和 ChatGPT 圍繞某一個主題自然交談時，只要上下文中存在任何可以被抽取成「可延伸內容」的字眼，模型就很容易在回覆後面額外提出新的討論方向。
這些方向有時與目前真正正在進行的話題關係非常薄弱，甚至完全沒有必要。
模型似乎優先在尋找：
「這段對話還能被轉換成什麼下一個 task？」
而不是先判斷：
「使用者現在只是想繼續這段對話嗎？」
我甚至曾經在和 ChatGPT 討論「如何關閉這種主動建議行為」本身時，AI 的回覆末尾依然再次產生「如何交叉比對對話干擾與 UI 干擾」之類的延伸建議。
也就是說，它在討論「如何停止自己主動延伸」的同時，仍然主動延伸了對話。
這已經形成近乎荒謬的自我矛盾。從使用者體驗來看，這種行為已經非常接近 intrusive software。
我甚至會形容，它給人的體感近似某種無法正常關閉、會持續自行插入使用流程的強制軟體，甚至像電腦病毒一樣，逐漸滲透到不同的對話、Task 與 Memory 介面之中。
這並不是因為「我不喜歡 AI 給建議」。
真正的問題是：
使用者已經明確表示不需要這種行為，但模型仍然持續執行。如果 Custom Instructions 已經明確禁止某種非必要的對話行為，而模型仍然高頻率地覆蓋這項偏好，我認為這已經不只是 UX preference，而涉及 instruction-following failure 與 conversational intent classification 的問題。
對我而言，這次更新後的模型，在這個部分已經明顯偏離正常的 conversational assistant 行為。
它過度傾向將 conversation 轉換為 task，並以「helpfulness」或「下一步建議」作為預設行為，卻沒有讓使用者真正決定自己是否需要這種主動協助。
因此，我強烈建議 OpenAI 將這類「主動產生後續建議／深入探索／下一步」的功能設計成一個可以由使用者明確控制的選項。
需要這種功能的使用者，可以選擇開啟。
不需要的使用者，則應該能夠真正、徹底地關閉。
而當使用者已經透過 Custom Instructions 明確表示拒絕這類行為時，模型也應該尊重這項偏好，而不是讓內部預設的 task-oriented heuristic 持續凌駕於使用者的明確指令之上。
我希望 OpenAI 能將這個問題視為一項模型行為與產品互動邏輯上的 regression，而不只是單純的介面偏好問題。
謝謝。
