Signals are not certainty.

The main unit on Fireflies is an event. An event can contain witness uploads, publisher links, official statements, location data, a timeline, corrections, dispute notes, and verification status. A post, person, or publisher does not become the event.

Fireflies may suggest that reports concern the same event based on time, place, visual similarity, language, source links, or moderator review. Reports remain separate evidence items with their own provenance. Clustering is reversible.

Unconfirmed

A report exists, but Fireflies has not found enough independent support to treat its central claim as corroborated.

Developing

A time-sensitive event is changing or evidence is still arriving. Important details may be incomplete or unstable.

Corroborated

Two or more meaningfully independent sources or strong evidence support the event’s central occurrence. Independence is assessed, not assumed from account count.

Verified

Available evidence meets a higher review threshold for the stated core claim, typically with reliable provenance, independent support, and/or authoritative confirmation. It is not a guarantee that every detail is correct.

Disputed

Material evidence or a credible source conflicts with the event’s central claim, location, date, attribution, or interpretation.

Corrected

A material part of the record changed. Fireflies preserves what changed, when, and why where safe and lawful.

Resolved

The active phase appears to have ended. Later evidence, official findings, or corrections may still update the record.

Labels can move in either direction as evidence changes. Fireflies should display the scope of a status—what claim is supported—not imply that every allegation or interpretation on an event page shares that status.

• Provenance: whether the original source, chain of custody, file, and method of acquisition are known.
• Independence: whether apparent sources actually derive from different observations rather than one recycled item or coordinated group.
• Time and place: whether capture time, event time, embedded data, environmental details, and geography are mutually consistent.
• Content integrity: whether material edits, generation, splicing, missing context, or re-encoding affect interpretation.
• Source competence: whether the source is in a position to know the specific fact asserted; official status does not make every assertion conclusive.
• Specificity: whether a claim is narrow and testable rather than vague, sensational, or broader than the evidence.
• Contradictions and corrections: whether credible conflicting evidence exists and whether sources have corrected or withdrawn prior claims.
• Risk: whether publishing or attaching a strong label could cause disproportionate harm if wrong.

Likes, followers, watch time, payment, political alignment, outrage, and creator popularity are not evidence-quality signals. A public Firefly history may help assess patterns and provenance but cannot replace event-specific review.

Fireflies may ingest from allowlisted publisher RSS feeds, public agencies, open-data sources, and licensed providers. It should store and show only what it is permitted to use: typically a headline, publisher, byline if supplied, canonical URL, dates, language, short attributed excerpt, correction status, and event metadata—not a copied full article or unlicensed publisher media.

A high source count can still trace back to one press release, video, wire report, anonymous post, or coordinated campaign. Fireflies should display both the number of items and the assessed number of meaningfully independent origins where possible.

• Syndicated copies do not count as independent confirmation of the copied claim.
• Multiple accounts directed by one person or organization do not count as independent witnesses.
• An official statement supports what the authority said; it does not automatically verify every underlying fact.
• A witness can support what they directly perceived, not every cause, identity, or motive inferred from it.
• Anonymous sources may be necessary, but their proximity, track record, incentives, and corroboration require careful handling.
• Financial, political, employment, family, or campaign relationships that materially affect interpretation should be disclosed or reflected in source context.

• Transcription, captioning, accessibility descriptions, optical character recognition, and search indexing.
• Translation and language detection with access to the original and a warning that nuance may be lost.
• Duplicate and near-duplicate hints, source-link suggestions, file-integrity signals, and location clues for reviewer assessment.
• Malware, graphic-content, privacy-risk, and policy-risk pre-screening used to prioritize or protect human review.
• Summaries or draft timelines grounded in cited source material and reviewed before high-risk publication.
• A clearly disclosed synthetic voice or avatar reading a human-reviewed, source-linked script where no reasonable viewer would mistake it for a witness or event capture.

AI assistance should be proportionate to the risk, documented, tested for material error and bias, and designed so reviewers can inspect the source evidence rather than accept a model output as proof.

• Fully or partly generated event footage, disaster scenes, conflict scenes, documents, or audio presented as an authentic capture.
• A fake witness, face swap, lip-sync, cloned voice, synthetic official, or fabricated statement presented as real.
• Removing, inserting, or altering people, objects, speech, smoke, damage, weapons, crowds, weather, signage, or chronology in a way that changes the apparent evidence.
• Using AI to declare guilt, identify a private suspect, establish identity from a face, or decide that a high-risk allegation is true.
• Automatically publishing sensitive claims, exact dangerous locations, or accusations without meaningful human review.
• Using an AI-confidence score as a public guarantee that content is authentic or synthetic.
• Training a generative model on private or quarantined evidence without a separately disclosed lawful basis, vendor controls, and any required user choice.

Contributors must disclose material AI generation or alteration, including the tool and affected parts when known. Ordinary device processing, compression, stabilization, exposure correction, accessibility captions, or redaction need not carry a prominent AI label unless they could change how a reasonable person interprets the evidence.

AI-assisted

A real underlying source was transcribed, translated, summarized, enhanced, or formatted with AI; the source remains available for comparison.

Synthetic presentation

A disclosed avatar, narration, diagram, or illustrative asset presents information but is not event evidence.

Suspected manipulation

Signals justify review or a caution label, but a final determination has not been made.

Confirmed synthetic evidence

Reliable provenance, creator disclosure, or strong technical and contextual findings establish that material represented as authentic evidence was generated or materially fabricated.

Automated systems may quarantine files, detect malware, identify likely duplicates, suggest event clusters, estimate location or synthetic-media risk, and route content to trained reviewers. Fireflies will document the purpose and safeguards of deployed tools in its final transparency materials.

High-risk removal, publication, verification, allegation, and account decisions should receive meaningful human review. A statement of reasons should say whether an automated tool materially contributed. Appeals must not be decided solely by the same automated signal that produced the challenged outcome.

1. Receive the challenge

   A contributor, source, subject, publisher, or member of the public can identify the exact claim and provide supporting context or evidence.

2. Protect against amplification

   Fireflies may temporarily reduce visibility, add a dispute label, withhold exact location, or pause distribution while assessing serious risk.

3. Assess the record

   Reviewers compare originals, provenance, source updates, independent evidence, event scope, and the potential harm of both action and inaction.

4. Correct transparently

   A material correction should identify what changed, when, and why. Quiet formatting or spelling fixes need not receive a full correction note.

5. Notify and preserve

   Where appropriate, affected contributors or reporters receive the outcome. A proportionate restricted audit trail may remain even when public content is removed.

Factual corrections are distinct from allegations of illegality and moderation appeals. Use [SUPPORT EMAIL] for a correction request, the illegal-content form for a legal notice, or the appeal route for a decision affecting you.

This policy is an operational standard, not a claim that Fireflies is a court, regulator, forensic laboratory, or emergency authority. Applicable law can require a different outcome in a specific country or case.

• Community Guidelines
• Moderation Appeals
• Report Illegal Content