Scran, Mardy, Twagging: The Regional Slang Britain's Platforms Keep Getting Wrong

There is a specific kind of email a UK creator gets at 11pm on a Sunday. It is three lines long, it is written by nobody in particular, and it says something like: we've removed your video for violating our community guidelines. Sometimes the video is a fight. More often, it's a lass in a kitchen in Bolton describing a chippy tea and calling the bread a barm.

If you spend enough time in British creator group chats, the pattern stops looking like bad luck and starts looking like a map. The flags cluster in the North West, the North East, the Welsh valleys, certain postcodes in Belfast and Glasgow. The content that gets pulled is rarely edgy. It's just local. And the machines doing the pulling appear to have never left the M25.

The training data never went past Watford

Platforms are understandably shy about publishing the composition of their safety classifiers. What creators describe, though, is consistent with a moderation stack tuned on enormous volumes of text that skews towards one variety of English: broadcast-standard, London-weighted, aggressively neutral.

That's fine if you're moderating a press release. It's less fine when the incoming speech includes mardy, ginnel, twagging, clart, boggin, canny and cob — words that carry no menace whatsoever in their home postcodes but sit oddly close to insult, threat or gibberish in a statistical model.

Take canny. In Tyneside it's one of the warmest words in the language. Feed it into a classifier trained elsewhere and you get a mild ambiguity that the system resolves the lazy way: downrank. Take mardy, which in Yorkshire and the East Midlands means sulky or soft, and which an automated sentiment pass can happily file as an insult directed at a person. Take twagging, the Yorkshire word for bunking off school, and watch a safety model read it as an admission of something.

None of these are obscure. Millions of people use them daily. That's precisely the problem: the model isn't failing on rare words, it's failing on common ones.

A short field guide to being misunderstood

Creators have effectively crowdsourced the taxonomy. Roughly three failure modes keep coming up.

The homonym trap. Words that look threatening in standard English but are affectionate in region. Bev, brew, bevvy. Sound as a compliment. Dead as a superlative, as in dead good. Gassed as delighted.

The spelling drift. Regional pronunciation written phonetically — summat, owt, nowt, reyt, thi, wor — trips spam and gibberish detection, which is built to flag text that doesn't look like text.

The banter problem. British humour runs on ritual insult. A comment section full of Northern mates calling each other a melt is a functioning friendship. A toxicity classifier that has never been to a school playground reads it as a pile-on and throttles the whole thread.

Welsh-language posts hit a fourth: vowel-heavy text, consonant mutation and diacritics that a language-identification model sometimes tags as spam, boilerplate or 'unknown'. Scots and Ulster Scots get similar treatment. That's not a dialect question anymore. It's a language one, and it's worse.

The workarounds are now a dialect of their own

Ask a creator how they deal with it and you get a taxonomy of small surrenders.

Some respell. Scran becomes skran, mardy becomes maardy, not because they talk like that but because the algorithm prefers it. Some subtitle everything, effectively publishing a translation layer for a machine that has been asked to police a country it doesn't know. Some move the good stuff to Discord or a private group chat and leave the platform a sanitised, dialect-free husk — which, if you care about British culture online, is a genuinely grim outcome.

And some, inevitably, keep a separate account for the local stuff, having learned that mixing regional voice with reach is a losing trade.

Here's the uncomfortable bit: none of this is new behaviour. Creators have always adapted to moderation. What's new is that the adaptation is now geographic. Being from somewhere has become a performance parameter.

Enter the dialect consultant

Brands have noticed, mostly because their community managers are drowning. The past eighteen months have produced a small, quiet cottage industry of regional voice and dialect consultants — people hired not to write copy but to sit on the other side of the glass and say no, that's not how a threat reads in Hartlepool.

Some of this is cynical. Some of it is just good practice that agencies should have been doing all along: hire moderators from the area, build a slang glossary for your brand's comment section, audit your own auto-responses for London bias. A few bigger names are now baking regional lexicons into their social listening, because a dashboard that thinks mardy is an insult is a dashboard that will tell you your customers are furious when they're actually being warm.

It is, in its way, a very British solution to a very Silicon Valley problem: rather than fix the model, hire a person who can interpret it.

Why this is a trust story, not a tech story

The obvious read is that this is a fairness issue. It is. The less obvious read is that it's a legitimacy issue, and it lands right as UK regulation starts asking platforms to demonstrate that their systems are accurate, proportionate and properly explained.

A moderation regime that can't distinguish canny from canny bad isn't just annoying a few comedians. It is producing a written record of enforcement that is skewed by postcode — and the people on the receiving end know it. Trust doesn't survive that. Reach doesn't either. Ask anyone whose hometown has quietly become a moderation liability.

So here's the question for 2026, and it's not rhetorical: if an algorithm can't understand a country, should it really be allowed to police it? Or should we, as an industry, admit that the most British thing about our platforms is that they've never once been north of Watford on purpose?