Possible False-Positive Filtering / Response Stalling in Microsoft Copilot

Jonathan McClennon 0 Reputation points
2026-08-10T21:48:24.72+00:00
Copilot said:

Subject: Possible False-Positive Filtering / Response Stalling in Microsoft Copilot

Summary

I am reporting what appears to be a recent change in Copilot's handling of technical engineering terminology. The issue began recently after a period during which the same project and terminology worked without problems.

Project Context

The project is a fictional engineering and ship-design project for the game Space Engineers. The project uses realism-inspired engineering terminology to describe fictional spacecraft, systems, and design concepts. Supporting project documents explicitly state that the work is fictional and game-related. [Starting D...ngineers F | PDF], [Document B...ngineers V | PDF]

Issue Description

For several weeks, Copilot handled discussions about the project normally.

Recently, Copilot has begun exhibiting one or more of the following behaviours:

Responses stop generating or appear to hang.

Chats become unresponsive after certain technical terms are used.

The conversation only proceeds if I significantly reword the terminology.

The behaviour occurs even in completely new chats.

Reinstalling the application and restarting the conversation has not changed the behaviour.

Examples of Terms Potentially Associated with the Issue

Examples include standard engineering or naval-architecture terminology such as:

real-world physics

displacement

calibre

115 mm

shell mass

alloy mass

structural load

These are normal technical terms and have been part of my vocabulary and workflow for many years.

Troubleshooting Already Performed

Started new chats.

Tested outside existing project conversations.

Reinstalled the Copilot application.

Removed and reloaded project files.

Retested using similar prompts.

The behaviour remained unchanged.

Why I Believe This May Be a Service-Side Issue

The same terminology worked previously.

The change appeared suddenly.

The issue occurs across new conversations.

Reinstalling the application had no effect.

Rephrasing specific terms sometimes allows the conversation to continue.

This suggests that a backend moderation, filtering, or classification change may be affecting legitimate technical and engineering language.

Supporting Documents

I have project documents which clearly identify the work as fictional Space Engineers content and not real-world engineering or weapons analysis. [Starting D...ngineers F | PDF], [Document B...ngineers V | PDF]

Request

Could the Copilot team please investigate whether a recent filtering or moderation update may be incorrectly affecting legitimate engineering terminology and causing conversations to stall or stop responding?

Thank you for your assistance.

Copilot said:

Subject: Possible False-Positive Filtering / Response Stalling in Microsoft Copilot

Summary

I am reporting what appears to be a recent change in Copilot's handling of technical engineering terminology. The issue began recently after a period during which the same project and terminology worked without problems.

Project Context

The project is a fictional engineering and ship-design project for the game Space Engineers. The project uses realism-inspired engineering terminology to describe fictional spacecraft, systems, and design concepts. Supporting project documents explicitly state that the work is fictional and game-related. [Starting D...ngineers F | PDF], [Document B...ngineers V | PDF]

Issue Description

For several weeks, Copilot handled discussions about the project normally.

Recently, Copilot has begun exhibiting one or more of the following behaviours:

Responses stop generating or appear to hang.

Chats become unresponsive after certain technical terms are used.

The conversation only proceeds if I significantly reword the terminology.

The behaviour occurs even in completely new chats.

Reinstalling the application and restarting the conversation has not changed the behaviour.

Examples of Terms Potentially Associated with the Issue

Examples include standard engineering or naval-architecture terminology such as:

real-world physics

displacement

calibre

115 mm

shell mass

alloy mass

structural load

These are normal technical terms and have been part of my vocabulary and workflow for many years.

Troubleshooting Already Performed

Started new chats.

Tested outside existing project conversations.

Reinstalled the Copilot application.

Removed and reloaded project files.

Retested using similar prompts.

The behaviour remained unchanged.

Why I Believe This May Be a Service-Side Issue

The same terminology worked previously.

The change appeared suddenly.

The issue occurs across new conversations.

Reinstalling the application had no effect.

Rephrasing specific terms sometimes allows the conversation to continue.

This suggests that a backend moderation, filtering, or classification change may be affecting legitimate technical and engineering language.

Supporting Documents

I have project documents which clearly identify the work as fictional Space Engineers content and not real-world engineering or weapons analysis. [Starting D...ngineers F | PDF], [Document B...ngineers V | PDF]

Request

Could the Copilot team please investigate whether a recent filtering or moderation update may be incorrectly affecting legitimate engineering terminology and causing conversations to stall or stop responding?

Thank you for your assistance.

Microsoft Copilot | Windows Copilot | Copilot+ PC
0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest

  1. AI answer

    2026-08-10T21:48:52.9+00:00

    The behavior is consistent with Copilot’s safety and filtering pipeline rather than a local app problem.

    Supported points:

    1. Copilot runs prompts through input classifiers before model execution to help filter harmful or inappropriate content.
    2. Microsoft 365 Copilot also uses jailbreak and cross-prompt injection classifiers to help block high-risk prompts before execution.
    3. Microsoft states that mitigations can occasionally fail or over-trigger, and that Copilot can still produce unexpected behavior or block content incorrectly.
    4. Reinstalling the app would not address a service-side filtering or model behavior change if the issue is caused by backend classifiers.

    Based on that, the most likely practical next step is to submit feedback from the Copilot experience when the stall happens, because Microsoft specifically asks for feedback when harmful, inaccurate, or unexpected output occurs so safeguards can be improved.

    Useful actions:

    1. Reproduce the issue in a new chat with the smallest prompt that triggers the stall.
    2. Submit feedback through Copilot using the thumbs-up/thumbs-down or in-app feedback option, and include:
      • that the project is fictional and game-related
      • the exact terms that trigger the stall
      • that the same terms worked previously
      • that the issue persists in new chats
      • that reinstalling did not help
    3. If possible, include screenshots and the shortest prompt pair showing:
      • a version that stalls
      • a reworded version that succeeds
    4. Keep using the reworded terminology as a temporary workaround while feedback is reviewed.

    What can be concluded from the available information:

    • A recent classifier or moderation change could plausibly affect legitimate technical terminology.
    • The provided information supports reporting it as a likely false positive or unexpected safeguard behavior.
    • There is no supported evidence here for malware or a device-side infection.
    AI-generated content may be incorrect. Read our transparency notes for more information.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.