AI-tells and filters

Provenance. Curated from the personal writing-style-profile (TAKS) and the decisions in planning/voice-filters-plan.md. Pattern definitions and examples are transcribed and re-scoped from existing material, not newly authored brand voice. Provenance is marked per pattern. Nothing here is brand-voice prose; it is detection method.

What this page is

A second-pass review layer for any text produced with or by a language model before it reaches a reader. It answers one narrow question: does this text carry tells of machine generation? It does not decide whether the text sounds like Aleris. That judgment stays with Voice (F4). A text can pass every filter here and still fail the voice test.

The filters exist because every content-generating process at Aleris — prototyping, presentations, documents, the brand assistant — produces draft text that an LLM has shaped. The same set of patterns should catch the same tells across all of them, so the floor is consistent regardless of who or what generated the draft.

The three layers

The patterns have three different scopes. They are kept separate because they have different owners and change at different rates.

Layer 1 — Generic AI-tells. Patterns that mark text as machine-generated regardless of language: dramatic sentences carrying no information, mirror structures, announcement framings, inflated vocabulary, self-praise, em-dash as default punctuation, empty emphatic tails, rhetorical section endings. These apply to English, Swedish, Norwegian, and Danish alike. This page owns them. Examples are given per market where the pattern shows differently.

Layer 2 — Language quality (per market). Patterns that mark text as non-native rather than machine-generated: calqued syntax, anglicised compounds, untranslated tech jargon, English verbs given local inflection. These live on sibling pages, one per market. Swedish is mature: language-quality-sv. Norwegian and Danish are scaffolded but empty — they cannot be written without a native speaker, so they carry status open rather than a borrowed Swedish ruleset.

Layer 3 — Brand voice. Voice (F4), already decided: den nära experten, the five principles, the test. The filters sit under F4, not beside it. F4 remains the judge of what Aleris sounds like.

This is also the distinction from patterns/anti-patterns.md. Those are behavioural guardrails — consent, referral, medical advice. They govern what the brand may say. The filters here govern how generated text betrays its origin. Different families; do not merge them.

Severity

Two levels. Severity can differ per pattern per genre, and the rules file carries the per-genre overrides.

hard — always fix. In a runtime surface with no human pass before the reader (brand assistant replies, generated UI copy), a hard hit is rewritten automatically. In a working surface (a document or presentation draft) the same fix produces a cleaner draft that the human still owns before it ships.

soft — flag for human judgment. A soft hit may be the right choice in some genres. A metaphor in an internal presentation can land; the same metaphor in a patient guide does not. The filter marks it and a person decides.

Scope — prose and headings

Every pattern runs on prose unless its line says otherwise. A pattern marked headings and prose runs on both; a pattern marked headings only runs on h1–h6 and on markdown heading lines, and is not evaluated against body text.

Scope is stated because leaving it unstated is what let headings go unchecked. This page was written against paragraphs in June 2026. Nobody decided headings were out of scope — the scope was simply never written down, so the omission was invisible until a heading set was read as a list. The rules file carries the same field, so a consumer can run the heading patterns against a document's heading tree rather than against its body.

Headings are where a model's defaults are most visible and least noticed. They are also read far more often out of context than in place — in a nav, in search results, in a screen reader's heading list, in a chunk retrieved by a language model. That is why the heading test below is a review method rather than a pattern: its first question is a property of the set of headings, which no per-heading rule can see.

The patterns (Layer 1)

Named in Swedish because these are the established working terms. Use the term, look for the pattern. Each carries an example (the tell) and a counter-example (the plain version), plus severity. Provenance in brackets.

The last two are the exception and say so in their own provenance: they were coined here in August 2026 rather than inherited from the writing-style-profile, because the profile has no vocabulary for headings. Coining was the lesser cost against breaking the page's naming in the middle of its own list, but it is a coinage and the reader is owed that. Because they were invented here, both carry a The word line unpacking the compound — the other twelve are established terms a reader can look up, and these two are not.

Effektmakeri — hard · headings and prose

Sentences that sound dramatic or insightful but carry no new information. [writing-style-profile]

• Tell: "You don't read your way to it, you work your way to it."
• Tell (heading): "Rules that can be checked, are." The claim is real; the ellipsis makes the reader assemble it.
• Plain: state the thing once and stop.

Fläsk — emphatic tails — hard

Small emphatic additions to a claim that stands without them. [writing-style-profile]

• Tell: "Three sentences of context gives a dramatically better answer. Every time." The "Every time" is fläsk.
• Plain: "Three sentences of context gives a better answer."

Kiaster — mirror structures — hard

ABBA mirror structures and their antithesis siblings (not X but Y, less about X than about Y, it's not X — it's Y). Almost always effektmakeri. [writing-style-profile]

• Tell: "It's about designing an organisation that develops people while people develop the organisation."
• Tell: "It is not a change of style. It is a change of order."
• Plain: if the substance is already in an earlier sentence, cut the antithesis — it only repeats with rhythm. If the contrast must be said, make it a plain clause: "The first paragraph leads with the situation rather than the procedure."

Onödiga inramningar — announcement framings — hard

Framings that announce a point instead of making it. [writing-style-profile]

• Tell: "The solution is simple:" before stating the solution.
• Plain: state the solution.

Uppblåsta ord — inflated vocabulary — hard

Vocabulary chosen to sound weighty rather than to be accurate. Aligns with the F4 rule against empty intensifiers. [writing-style-profile]

• Tell: "non-negotiable", decorative "leverage", strategy nouns used as filler.
• Plain: say what actually changes. If the abstraction cannot be replaced by a description, it probably carries nothing.

Töntigt — soft

Over-familiar warm-ups that try to heat the text and deflate it. English drift includes "worth sitting with", "let's explore", "this is interesting". [writing-style-profile]

• Tell: "Reach out when you hit something. I've probably hit it already."
• Plain: drop the warm-up; begin with the content.

Effektmakeri-avslutningar — punch endings — hard

Section endings that land on a rhetorical punch instead of just stopping. [writing-style-profile]

• Tell: "That difference is the whole guide." / "Use AI to enhance your voice, not replace it."
• Plain: end on the last substantive sentence.

Dygdesignalering — self-praise — hard · headings and prose

Text that claims a virtue about itself instead of demonstrating it. [writing-style-profile; a second sentence was cut 2026-08-31 — see the note under the pattern list]

• Tell: "the intellectually honest position is", "the real question is not X but Y", "this report is deliberately designed to challenge".
• Tell (heading): "Five things, described plainly." The heading praises the writing below it.
• Plain: cut the label. Any sentence calling itself "honest", "uncomfortable", "real", or "interesting" loses the label.

Em-dash som standardpaus — soft

Em-dash used as default punctuation instead of comma or period. A strong machine signature. [writing-style-profile]

• Method: comma for a short pause, period for a new thought. Em-dash only for a genuine break or interpolation. After drafting, count em-dashes per paragraph; if most paragraphs have one or more, the text is over-dashed.

Formellabelloop — repeated labels — soft · headings and prose

The same label shape reused across sections, building mechanical structure. [writing-style-profile; heading scope added 2026-08-31]

• Tell: "Insight for X:" or "The real question:" appearing as a repeating section closer.
• Tell (heading): three or more sibling headings sharing a grammatical template — "The writing / The values / The checks / The package / The surfaces". Read as a list they say nothing, so the reader has to enter each body to learn anything at all.
• Tell (heading): a run of headings opening on the same interrogative — What it is… / What happens when… / What is there… / What would be…. Wh-word openers are the shape a language model reaches for by default when naming a section; the run is a stronger tell than any single one.
• Plain: vary the label, drop it, or replace it with a heading that does its work. The threshold is three consecutive siblings, not two.

Smarta metaforer — clever metaphors — soft

Metaphors reaching to be clever. Borderline personal taste, but F4 favours plain explanation, so it transfers — as soft, because some genres allow it. [writing-style-profile; severity confirmed 2026-06-11]

• Tell: "Like furnishing a room instead of carrying chairs to every meeting."
• Plain: describe the thing directly.

Tvåmeningsrytm — two-sentence AI-poetic rhythm — soft

Short, paired two-sentence cadence that reads as machine-poetic even when each sentence is fine on its own. [writing-style-profile; included as soft 2026-06-11]

• Method: flag for human judgment, not auto-fix. If many paragraphs fall into the same two-beat rhythm, vary the sentence lengths.

Formrubrik — form instead of content — hard · headings only

The word. Form + rubrik, heading: a heading about the text's form. Its subject is how the section is built — how long it is, how many parts it has, what kind of writing it is — rather than what the section says. The name is descriptive, not a term of art; it was coined here because no established one existed.

A heading that names the form of the text below it instead of its content. Length, count and manner are already visible on the page, so a heading reporting them adds nothing. [coined 2026-08-31, not inherited; examples read off the Brand OS Explained heading set]

• Tell: "What it is, in one paragraph" — the reader can see it is one paragraph.
• Tell: "In short", "Overview", "At a glance", "The basics", "What is there, and what is not".
• Method: strip the heading and read on. If nothing is lost, the heading was carrying nothing.
• Plain: say what the section says. "The published site is three months out of date" over "What is there, and what is not".

The matchable list in the rules file holds the vocabulary a check can match. It is narrower than the pattern: "What is there, and what is not" is a clear instance that no list will ever hold, and it is caught by the method above or not at all.

Etikettrubrik — label heading — soft · headings only

The word. Etikett, label, + rubrik, heading: a heading that labels rather than tells. It names what the section is about and stops, the way a label states what a thing is without saying anything about it. Etikett is the word Digg's guideline 61 uses for a label, which is the one anchor either of these two coinages has.

A heading that names a topic without saying anything about it. A noun phrase with no verb and no claim. [coined 2026-08-31, not inherited; the term is journalism's "label headline", and Swedish klarspråk rules out the same thing when it names "Introduktion" as a heading that does not help]

• Tell: "The menu", "The constituent parts", "The writing".
• Plain: give it a verb, or give it a fact. "The published site is three months out of date" is a heading; "The status" is a label for one.
• soft because nav items, glossary entries, table columns and a template's structural sections are legitimately labels. The tell is a label in running prose, and a run of them is worse than one.

A note on this list's own wording, 2026-08-31

Two pattern definitions on this page and its Swedish sibling were themselves kiaster, and had been since the patterns were transcribed in June 2026.

• Dygdesignalering carried "The text is honest by being clear, not by saying it is honest."
• Anglicismdrift, on language-quality-sv, carried "Kodväxling är inte djup, det är drift."

Both were cut rather than rewritten, which is what the Kiaster entry above prescribes: if the substance is already in an earlier sentence, cut the antithesis — it only repeats with rhythm. In both cases it was. Claim-versus-demonstrate is already in dygdesignalering's first sentence; drift is already described in anglicismdrift's.

Four live copies were corrected together — both pages here, the counter_example in the rules file, and both source entries in the writing-style-profile in TAKS, which is where they were transcribed from and where they would otherwise have been re-imported. workspace/status/board.md card 108 keeps the sentences quoted, because a record has to be able to name what it retired.

What was deliberately left standing. "The split is by surface, not by content" and "It is the floor, not the gate" have the same shape but state a distinction that appears nowhere else, and the Kiaster entry's second clause allows exactly that: if the contrast must be said, make it a plain clause. Both already are. The rule targets an antithesis that repeats, not every contrast.

Not made into a check. A scan for these shapes was run across both pages, the rules file and the profile; of eleven matches, four were real, one was a deliberate quotation of the pattern being taught, and six were false positives on ordinary prose. A regex cannot tell a repeating antithesis from a load-bearing one, and that judgment is the whole rule. This stays a review method.

Review methods

The patterns are a list; the methods are what make them usable inside a process.

Second pass, not prevention. The reliable mechanism is re-reading the draft against the pattern list before it leaves the session, not trying to suppress the tells while generating. The Swedish calque check states this explicitly and it has held in practice: the generating pass misses what a dedicated re-reading pass catches.

Progressive enhancement. Write the plainest version that carries the content and stop there. Rhetoric, rhythm, metaphor, and emphasis are enhancements added only when something is missing — not inserted by default. This is the inverse of how a model drafts, which is why the second pass is needed.

The translation layer. Human prep before generation, human ownership after. The filters reduce the tells in a draft; they do not make the draft shareable. See the boundary below.

The heading test

Three questions, run on the heading set rather than on any single heading. Added 2026-08-31. Each question traces to an external source rather than to preference, and the first is the one no pattern above can catch.

1. Read the headings alone, in order, with the body hidden. Do they summarise the page? A page's headings should work as its summary. This is the criterion in Swedish klarspråk guidance, and it is the manual procedure Digg's guideline 61 prescribes for WCAG 2.4.6: list the headings and check each is comprehensible without its surroundings.
2. Does each heading say something the reader cannot already see? This is what kills the form-describing heading. Length, count, and manner are visible on the page; a heading that reports them has spent a line saying nothing.
3. Does it survive being lifted out of the page? Into a nav, a search result, a screen reader's heading list, a chunk retrieved by a language model. Headings are read out of context more often than in place.

Where a heading fails, the repair is usually the same: state the fact or name the action. Front-load the words that carry it — attention concentrates on the first few words of anything read as a list — and keep it inside roughly ten words.

This is an accessibility floor, not only a voice preference. baseline/BASELINE.md commits to WCAG 2.1 AA, and 2.4.6 Headings and Labels is a AA criterion: headings and labels describe topic or purpose. A heading that names its own container does not describe the section's topic or purpose. That criterion is currently cited nowhere in the corpus while 1.4.1, 1.4.11 and 2.5.8 are cited by number — see workspace/evaluations/headlines-2026-08-31.md for the gap and the open question of whether it enters constitutional/accessibility-is-foundational.md.

The boundary

This tooling makes AI-assisted output carry fewer tells. It does not change the content creation boundary. Filtered output is still generated output, and the human translation layer — prep before, ownership after — remains the quality control for anything that reaches a stakeholder. Tooling that reduces visible tells creates exactly the temptation to skip the human pass, so the rule is stated plainly here: for stakeholder-facing text the filters are the floor, F4 and a human are the gate.

The auto-rewrite of hard hits does not weaken this. The split is by surface, not by content. In runtime surfaces (brand assistant, generated UI copy) the filter is the last line before a reader, so hard hits are fixed automatically. In working surfaces the auto-rewrite produces a cleaner draft, and the translation layer still applies before anything ships.

Relationship to F5 and the rules file

F5 Writing Conventions will own sentence-level production mechanics: du-tilltal, active voice, instruction plus reason, terminology. This page owns detection of generation artefacts. They cross-reference but stay separate: F5 says how to write, the filters say how to catch what should not have been written.

The machine-readable rendering of every pattern on this page and the sibling language pages lives at voice-filter-rules. That file is generated from these pages and never edited directly, the same discipline as tokens. It is what the brand assistant, prototyping prompts, and any future linting consume.