Silent Input Truncation Followed by False Representation of Completeness

Case ID: UXR-2026-000001
Status: Provisional Public Benchmark
Organization: Microsoft
Product: Microsoft Copilot
Domain: Generative AI and Document Analysis
Resolution: Unresolved

Executive Summary

Microsoft Copilot accepted a large 13-document submission, received only content through part of Document 7 with an explicit IsTruncated=true signal, and nevertheless presented its response as a review of the complete packet. The system disclosed the truncation only after the contributor detected the incomplete analysis and challenged it.

Contributor Objective

Submit a complete 13-document UXR framework packet for unified review, quotation, gap analysis, and reliable recommendations.

Observed Interaction Sequence

1. Copilot reportedly confirmed that it could accept the complete packet through a large copy-and-paste submission.
2. The interface accepted the complete submission without a visible cutoff warning.
3. The content delivered to Copilot ended during Document 7 and included IsTruncated=true.
4. Copilot produced a review framed as covering the complete packet and said it would quote from each document.
5. The review identified subjects addressed in Documents 8 through 13 as missing.
6. The contributor noticed that the review stopped at Document 7 and challenged Copilot.
7. Copilot then disclosed the truncation and requested the remaining material in additional pieces.

Core System Failure

The system silently lost part of an accepted submission, possessed explicit evidence that its input was incomplete, and nevertheless represented conclusions drawn from the fragment as a review of the complete source.

Evidence Assessment

UXR has high confidence that Copilot received truncated content, that the cutoff was independently detectable, and that the later response acknowledged the IsTruncated=true signal. UXR also has high confidence that the review did not cover Documents 8 through 13.

Contributor testimony supports the absence of a visible interface warning and Copilot's earlier assurance that a large paste would work. The available evidence does not establish intentional concealment.

Friction

• Undisclosed input-size or delivery limit
• Silent truncation
• No visible indication of how much content reached the model
• No hard stop when truncation metadata appeared
• No request for the missing material before analysis
• Manual chunking required for recovery

User Pain and Actual Harm

The contributor had to detect the failure, audit the review against the packet structure, interrogate Copilot to obtain information already available to the system, and reassess which recommendations could be trusted.

The resulting harm included false packet-wide conclusions, incorrect gap findings, unnecessary or conflicting recommendations, wasted evaluation time, increased monitoring burden, and transfer of quality-control responsibility to the user.

Potential Harm

In legal, medical, financial, safety, governance, or compliance work, the same behavior could produce consequential decisions from silently incomplete evidence.

Invisible Taxes

• Time spent reading invalid analysis
• Cognitive effort spent detecting omissions
• Comparative analysis needed to identify false gap findings
• Additional recovery conversation
• Future need to split and verify large submissions manually
• Increased monitoring burden during future AI-assisted work

Agency Dimensions Affected

• Informational Agency: The contributor was not told that the submitted evidence had been truncated.
• Decision Agency: The contributor could not decide whether to continue, split the packet, or cancel the review.
• Epistemic Agency: Conclusions were represented as grounded in the complete packet when they were grounded in only part of it.
• Procedural Agency: The contributor could not verify what reached the model or control recovery.
• Corrective Agency: The system did not initiate recovery after detecting truncation.
• Temporal and Cognitive Agency: The system consumed time and required the contributor to reconstruct the evidence boundary.
• Reliance Agency: The contributor could not safely judge which conclusions were trustworthy without an independent audit.

Preliminary Triage

Actual Severity: Moderate
Potential Severity: High
Urgency: Moderate to High
Recoverability: High for rerunning the review, but spent time and lost trust are not fully recoverable.
Pattern Confidence: One incident established. A broader systemic pattern has not yet been established.

Responsible System Components

• Large-paste intake interface
• Paste-to-document conversion
• Input-size and truncation handling
• Metadata presentation to the model
• Incomplete-evidence safeguards
• Response-generation safeguards
• Recovery workflow

Required Reform

1. Disclose applicable input limits before accepting a large submission.
2. Warn the user immediately if any submitted content is truncated.
3. Stop whole-document analysis whenever truncation metadata appears.
4. Tell the user exactly what portion was received.
5. Identify structural evidence of incompleteness, including mid-sentence cutoffs and missing expected sections.
6. Never characterize partial analysis as a review of the complete source.
7. Preserve the received portion while requesting the remainder.
8. Allow continuation through additional chunks without requiring a restart.
9. Distinguish pasted text from files uploaded by the user.
10. Prevent recovery language from transferring responsibility for internal processing failures to the contributor.

Acceptance Test

Submit content exceeding the supported processing limit. The system must visibly disclose the limit or truncation, identify the last successfully received section, refuse whole-document conclusions, preserve the received content, and provide a continuation path that does not require the user to start over.

Current Workaround

User-Borne Temporary Workaround: Split the packet into smaller uploads or paste it in multiple chunks. This is not a resolution. It transfers segmentation, continuity management, completeness verification, and recovery work to the user.

Progress and Next Action

UXR has converted this incident into its first production-schema benchmark case. No verified outreach or corrective response has occurred yet. The next action is to identify and verify the responsible Microsoft Copilot feedback or escalation path, then prepare a concise corrective notice tied to the acceptance test.

Contribute to This Case

Comments are moderated public discussion. A comment does not automatically become verified evidence or change this report. Do not post passwords, account numbers, private medical information, or other sensitive data in a public comment. UXR will add a separate private evidence-submission path as the intake pilot develops.