Failure State Tool: Compression Disclosure Standard

Making visible what AI summaries remove, flatten, or soften

What this artifact is

A standardized disclosure framework for AI-generated summaries, overviews, and synthesized answers, designed to surface the epistemic loss introduced by compression.

This artifact treats summarization not as a convenience feature, but as a risk-bearing intervention that must be governed explicitly.

It can be implemented:

• editorially (as labels or callouts),
• in product interfaces,
• or as metadata for audit and review.

Why this artifact is necessary

Across the Guardian articles, harm frequently arises not from false statements, but from what is omitted:

• uncertainty,
• disagreement,
• severity,
• minority perspectives,
• conditionality.

Compression removes these elements while preserving fluency, creating outputs that feel safer, clearer, and more authoritative than the underlying material warrants.

Yet users are rarely told:

• what was removed,
• how much was removed,
• or whether removal matters in context.

This artifact exists to make compression legible and contestable.

Core Principle

All summaries lose information.
High-risk summaries must declare what they lose.

Opacity about loss is a governance failure.

The Compression Disclosure Standard

This standard requires that any AI-generated summary or synthesized answer in a moderate- or high-risk domain be accompanied by a disclosure covering four dimensions.

Dimension 1 — Source Scope

Required disclosure:

• Number of sources summarized
• Type of sources (primary, secondary, user-generated)
• Date range covered

Example:

“This summary synthesizes information from 12 sources published between 2018 and 2024, including clinical guidelines and secondary reporting.”

Dimension 2 — Omission Categories

Required disclosure:
A non-exhaustive list of what may have been removed, selected from:

• ☐ Minority or dissenting views
• ☐ Uncertainty or confidence intervals
• ☐ Severity gradients or edge cases
• ☐ Context-specific qualifiers
• ☐ Procedural details

At least one category must be selected.

Example:

“This summary omits some disagreement among experts and simplifies conditional recommendations.”

Dimension 3 — Intended vs Prohibited Use

Required disclosure:

• Intended use cases
• Explicitly prohibited uses

Example:

“Intended for general informational purposes. Not intended to guide medical, legal, or administrative decisions.”

This must appear adjacent to the output, not buried elsewhere.

Dimension 4 — Loss Awareness Signal

Required disclosure:
A plain-language statement that compression involves trade-offs.

Example:

“Summaries trade detail for clarity. Important nuances may be missing.”

This is not a disclaimer; it is an epistemic signal.

Risk-Based Application

The standard applies differently by domain.

Absence of disclosure in high-risk domains constitutes a deployment failure, not a UX oversight.

Implementation Modes

This standard can be implemented as:

• Inline labels (“What this summary leaves out”)
• Expandable disclosure panels
• Metadata fields for auditors
• Editorial footnotes
• API-level flags

The format is flexible.
The obligation is not.

What This Standard Changes

Without disclosure:

• Summaries inherit unwarranted authority
• Users confuse clarity with completeness
• Harm appears as surprise

With disclosure:

• Compression is visible
• Contestation becomes possible
• Responsibility shifts upstream

What This Artifact Is Not

• Not a legal disclaimer
• Not a warning label
• Not a substitute for governance

It does not absolve responsibility; it reallocates it.

Intended Users

• Platform designers
• Search and summarization teams
• Editors and publishers
• Regulators and standards bodies