How we comply

Our analysis is AI-assisted, and we say so in the footer of every page and every report. The copy on this site was drafted with AI assistance and human-edited, and the footer of every page says so — that is our own published standard, not a legal requirement: Article 50's text duty reaches AI-generated text only where it informs the public on matters of public interest. No synthetic media is published without labelling. Article 50 reaches us where our output is used by people in the EU; we apply the same disclosure everywhere rather than only there.

Submitted content is hashed (SHA-256) and the hash is logged for 90 days to monitor accuracy and abuse. Raw text is not stored. Our LLM provider is Anthropic; inputs are not used for model training per provider terms.

Not yet. CheckFirst checks the words in your ad, post or landing page. It does not analyse images: it does not detect injection pens, before/after comparisons, body imagery or anything else in a picture, and no image is ever uploaded to us. If an ad's risk sits in the image rather than the words — as it often does in weight-loss and aesthetics advertising — a text check does not cover it, and we say so rather than imply otherwise. Image checking is on our roadmap and will only be offered when it can be described honestly.

Every tier — including the free checker — runs every live rule pack: the ASA CAP Code Section 12, MHRA medicines advertising rules, EU AI Act Article 50 transparency duties, and platform ad and AI-labelling policies for the platforms you select. Regulatory coverage is never gated by price, up or down: no tier gets fewer regimes, and no tier gets unfinished packs as an upgrade. If a pack were ever withheld from a tier, the report would say on its face which regimes were not checked; a clean grade never stands in for a check that did not run.

Fixed 18 August 2026: before that date, scans of Meta, TikTok and website content also ran the Google Ads and TikTok/Meta policy packs regardless of the platforms selected, so some reports carried off-platform flags (for example a Google Ads policy flag on an Instagram post). Rule packs are now scoped to the platforms you select. Reports issued before 18 August 2026 may still show those flags. Fixed 19 August 2026: from launch (13 June 2026) until that date, declaring content as AI-generated text raised AIA50-004 — the EU AI Act Article 50 rule about unlabelled deepfake-style depictions of real people — as a red violation, because the check read only the declaration and not what kind of media was declared. A text-only declaration can never be a deepfake; that flag was unwarranted and it lowered the grade. The rule now fires only when an image, video or audio is declared AI-generated. Reports issued before 19 August 2026 that show AIA50-004 on text-only content should be read without it. Changed 19 August 2026: before that date, naming a prescription-only medicine in content declared as website or editorial input (no paid platform selected) was graded exactly as it is in an ad — a red MHRA-001 violation — even where the text referred to the medicine rather than advertised it; our own homepage, quoting an enforcement notice, graded F this way. The rule is unchanged for ad copy. On declared website or editorial input it now fires as a conditional flag that states the prohibition (regulation 284 of the Human Medicines Regulations 2012) and says plainly that whether the copy advertises or refers to the medicine is a promotional-purpose judgement the check does not make. Reports issued before 19 August 2026 on editorial input may carry that rule as a violation. Fixed 19 August 2026: the same defect in four more rules. From launch (13 June 2026) until that date, declaring content as AI-generated text raised AIA50-002 (EU AI Act Article 50 synthetic-media marking) as a red violation, and — where TikTok, Meta or YouTube was selected — PLAT-001, PLAT-002 or PLAT-003 (the platforms' realistic-AI labels), because those checks read only the declaration and not what kind of media was declared. Article 50(2)'s marking duty falls on providers of the generative system, and the platform labels apply to realistic images, video or audio; none of them applies to a text-only declaration. Each rule now carries the media types it is about and fires only for those. Reports issued before 19 August 2026 that show these flags on text-only content should be read without them. Changed 20 August 2026: we measured our own grader on identical copy, twenty runs, and found the AI pass undecided on borderline judgement calls — one medicines rule was raised in nine runs out of twenty, moving the score by up to twenty points between runs. The AI checks now run seven times per scan: only findings raised all seven times count toward your grade; findings raised in some runs are shown as "needs your judgement" with the frequency (for example "raised in 5 of 7 runs"), and are not counted. Two things we cannot fix and will not hide: the AI judge is non-deterministic, so a borderline finding can still pass all seven runs and be counted as if certain — on the worst case we measured, in roughly one scan in twenty — and scores from before 20 August 2026 carried the full run-to-run variance without saying so. When a check is over-flagging, we say so here while we fix it.