AI use & disclosure policy

AI assistance is part of how Matchday Global publishes football news. This page describes exactly where, how, and under what guardrails. We do not hide AI use; we do not pretend AI does the parts of journalism it can’t do.

Where AI assists

• Wire-body expansion. Wire summaries on /news are generated by an AI assistant (Anthropic Claude Sonnet) reading the headline + 1–2-sentence wire description from the original publisher and producing an MDG-voiced 4–5 paragraph paraphrase plus verifiable common-knowledge context about the entities the seed names. The original publisher is named in the body and credited in a footer line.
• Predictions. Match predictions are generated by a Poisson statistical model (transparent math published on /predictions/strategies) — not by a generative AI. Probability bands are shown alongside every pick.
• Image generation. When a wire article has no licensed photograph, a brand-watermarked illustration may be substituted. Cover illustrations on Tier-1 editorial articles are generated via Grok Imagine for stadium-scene art; people’s faces, club crests, and any image depicting a real player are NEVER AI-generated — those use licensed editorial photography only.

Where AI does not assist

• Bylined original editorial. Articles bylined to Oluwatayo Omoniyi or Tony Gege are written by those named editors, not by AI. A spell-check or grammar pass may have been run, the same as on any modern newsroom workflow.
• Fact selection. Which facts to verify, which sources to trust, when to break a story, and when to spike one are decisions made by named editors. AI does not pick what we cover.
• Quotes. Direct quotes are never AI-generated. If a quote appears in a wire summary, it appears verbatim from the original wire description and is bracketed by attribution to the original publisher.
• Editorial corrections. Decisions about whether to correct, retract, or right-of-reply are made by named editorial staff, not AI.

The runtime fabrication filter

Every AI-expanded wire body passes through a server-side “hallucination filter” that runs at the moment the body is written to cache, before any reader sees it. The filter hard-blocks any paragraph containing:

• Fabricated speaker quotes (long quoted blocks paired with attribution verbs: “said”, “told reporters”, “according to sources”).
• Fabricated employment-status claims (“currently on loan”, “permanently signed from”, “now playing for”) when the seed contains no transfer keyword. This specific class of fabrication caused a high-profile error on 2026-05-05 and is now blocked at the cache write boundary.
• Fabricated competition involvement (“going to AFCON”, “World Cup squad”) when the seed names no such tournament.
• Fabricated fixture dates, kickoff times, venues, or competition rounds.
• Fabricated journalist citations (“as the BBC noted”, “Sky have reported”) for publishers other than the article’s actual source.
• Banned commentary phrases (a tightly-curated list of register tells: “underscores the”, “rich legacy”, “storied rivalry”, “passionate fanbase”, “rose to prominence”, “central figure” etc. — phrases the AI keeps drifting toward despite explicit prompt rules).

When the filter strips a paragraph, the article falls back to displaying only the paragraphs that passed. If every paragraph is stripped, the article shows the original wire description as a single paragraph and no AI-expanded body. A background audit function re-scans the entire cache every 6 hours and flags any article whose body still contains banned content for review or suppression.

How to tell which is which

Three visual signals on every article:

• The byline. Original editorial carries a real author name with a clickable photo and bio. Wire summaries carry “Matchday Global” and link to the source publisher in the footer.
• The footer. Wire summaries end with “Source: [publisher name]” in small text. Original editorial does not.
• The page chrome. Wire summaries link out to a related-stories rail aggregating other coverage. Original editorial links to internal MDG context (related fixtures, predictions, standings).

Hallucination accountability

When a fabrication slips through the filter and reaches readers, we treat it as an editorial failure, not a model failure. The article is suppressed (returns HTTP 404 across every site surface) and the failure mode is added to the runtime filter so the next instance is caught. The audit log is publicly retrievable on request via /contact.

Models we use

• Anthropic Claude Sonnet 4.5 — primary text model for wire-body expansion. Trained by Anthropic; we do not train models ourselves.
• xAI Grok 3 Fast — fallback text model when the primary is unavailable.
• xAI Grok Imagine — illustrated stadium-scene cover art for original editorial when no licensed photograph is available.
• ElevenLabs — voice synthesis for the audio narration that accompanies original editorial articles. Voice is the named author’s assigned ElevenLabs voice; we never narrate one author’s words in another author’s voice.