Vague attribution and promotional register in LLM prose
Parent: Writing and Documentation · Published reference · snapshot 2026-09-23 · skill writing-expert/references/vague-attribution-and-promotional-register.md
↓ Facts as markdownall context files
Source-anchored research for Vague attribution and promotional register in LLM prose. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Definitions
- Wikipedia's guide itself cautions that signs of AI writing are not proof: LLMs are trained on human writing, and such phrasing appears in editorials and blogs, so treat markers as weak evidence. (low confidence) [source]
- Swapping a vague authority for a named one is not sufficient: RAG chatbots may attach statements to named sources (example: 'Roger Ebert highlighted the lasting influence') regardless of whether those sources say anything close, so each rewrite must be checked against the source text. (low confidence) (verified as of 2026-09-23) [source]
How it works
- Wikipedia's 'Signs of AI writing' guide says LLMs struggle to keep a neutral tone and produce advertisement-like or travel-guide prose even when prompted for an encyclopedic style, both in new text and in rewrites. (low confidence) (verified as of 2026-09-23) [source]
- The same guide says LLMs present views from one or two sources as widely held, mention multiple 'reviewers' while citing only one person, and imply lists are non-exhaustive when the sources do not. (low confidence) (verified as of 2026-09-23) [source]
Measurements and reference values
- The Words to watch Bob Dylan example replaces 'defining figure ... brilliant songwriter' with a Time 100 inclusion, a quoted description, and a fact (hundreds of covers by the mid-1970s): the fix is attributed, checkable evidence, not a milder adjective. (low confidence) [source]
Problems, failure modes and limitations
- The typical failure pattern is on cultural or geographic topics: the text keeps reminding the reader of a subject's importance using flowery description ('nestled within the breathtaking region', 'rich cultural heritage'), which asserts significance without a sourced fact. (low confidence) [source]
- Scope limit: the manual says such phrases are not automatically weasel words. They can stand in a lead or topic sentence when the article body or rest of the paragraph supplies attribution, and can be used for views properly attributed to a reliable source. (medium confidence) [source]
Comparisons and alternatives
- Wikipedia's NPOV policy instructs editors 'Do not editorialize' and to attribute opinions to reliable sources rather than state them as fact, while uncontested facts are stated directly in the encyclopedia's voice. (high confidence) [source]
- Wikipedia's Signs of AI writing says LLM text inflates significance with stock phrases (stands as, testament to, pivotal role, underscores, reflects broader trends), regressing to generic positive language instead of specific facts, so output is less specific and more exaggerated. (medium confidence) (verified as of 2026-09-23) [source]
- Plain-language guidance says to use concrete words and specific nouns over abstract ones, give exact figures instead of vague quantities, and prefer active voice that names the actor ('We recommend' rather than 'It is recommended'). (medium confidence) [source]
- Instruction-tuned Llama models overuse specific words relative to base models (e.g. 'nuanced' +8,342%, 'firstly' +4,794%), and a 400-person rating experiment found a significant preference (52.4% vs 47.6%, p<0.01) for high-overuse variants, linking lexical tics to learning from human feedback; this is Llama-only, English scientific text, and the listed words are academic, not brochure vocabulary. (low confidence) [source]
Facts and statements
- Wikipedia's Words to watch lists editorializing markers: notably, it should be noted, arguably, interestingly, essentially, clearly, of course, without a doubt, fortunately, unfortunately. Highlighting something as significant needs attribution. (high confidence) [source]
- Connectives such as but, despite, however, although, furthermore can create unsupported relationships between statements, implying an editorial judgement on how facts relate. (medium confidence) [source]
- Lexical overrepresentation of LLM-associated words (21 focal words such as delve, intricate, underscore in scientific abstracts) is documented; the study finds model architecture, algorithm and training data do not explain it and RLHF is consistent with playing a role, with ambiguous experimental support. (medium confidence) [source]
- Wikipedia's Words to watch defines puffery as unattributed promotional words (legendary, best, great, acclaimed, iconic, visionary, outstanding, leading, celebrated, award-winning, landmark, cutting-edge, innovative, revolutionary) and prescribes replacing them with facts and attribution. (high confidence) [source]
- Wikipedia's Signs of AI writing lists undue emphasis on significance: LLM text overuses 'stands/serves as', 'is a testament/reminder', 'crucial/pivotal/vital role', 'underscores importance'. (medium confidence) (verified as of 2026-09-23) [source]
- The same guide lists promotional LLM adjectives: 'boasts a', vibrant, rich, profound, showcasing, nestled, natural beauty, groundbreaking. (low confidence) (verified as of 2026-09-23) [source]
- Wikipedia editors attribute the pattern to LLMs regressing to the statistical mean: specific facts get replaced by generic importance claims, so text becomes both less specific and more exaggerated. (low confidence) [source]
- Hype language (hyperbolic or subjective wording used to glamorize or exaggerate) is a recognised phenomenon in human scientific writing; an NLP paper on NIH grant applications shows annotators can reliably label candidate hype adjectives with written guidelines, but the task needs domain knowledge and temporal awareness of facts. (low confidence) [source]
- Promotional wording alone does not prove AI authorship: Wikipedia notes not all promotional writing is AI-generated, humans write press releases and travel guides this way, and newer models are subtler than GPT-4-era ones; the signal is consistent overuse across unrelated topics. (medium confidence) (verified as of 2026-09-23) [source]
- Fixing only the marker words can hide deeper problems: the guide calls the phrases signs of a problem, not the problem itself, and warns that superficial swaps make detection harder without fixing the unsupported claims. (low confidence) [source]
- Automated AI-text detectors are unreliable as a substitute for phrase lists: a multi-tool evaluation found neither accurate nor reliable results, with false positives on human text and false negatives on AI text. (medium confidence) [source]
- Wikipedia guidance fixes vague attribution three ways: attribute to a specific named reliable source, use in-text attribution that accurately reflects the source, or rewrite so no attribution verb is needed. (medium confidence) [source]
- Claims about what people say, think, feel or believe should be clearly attributed; phrases like 'some people say' or 'research has shown' give an impression of support without letting the reader assess the source. (medium confidence) [source]
- In-text attribution is required for opinions and dubiously reliable sources; the form depends on source type (author, agency, title for anonymous works), and the full citation should keep all known information. (low confidence) [source]
- The Wikipedia watch-word list for promotional register includes: boasts a, vibrant, rich (figurative), profound, enhancing, showcasing, exemplifies, commitment to, natural beauty, nestled, in the heart of, groundbreaking, renowned, featuring, diverse array. (low confidence) (verified as of 2026-09-23) [source]
- Word tells are weak evidence: the Wikipedia guide itself says not all promotional writing is AI-generated, and a legal-tech commentator notes they used 'delve' before LLMs existed, so single-word detection is insufficient proof. (medium confidence) [source]
- Wikipedia's guide reports older models such as GPT-4 write more blatantly positive text, while newer models are more subtly positive and avoid obvious superlatives like 'the best', so a superlative-only filter under-detects newer output. (low confidence) (verified as of 2026-09-23) [source]
- Wikipedia's 'Signs of AI writing' has a section on vague attributions and overgeneralization of opinions: chatbots attribute opinions to vague authorities via weasel wording, with phrases like 'Industry reports', 'Observers have cited', 'Experts argue', 'Some critics argue', 'several sources/publications'. (medium confidence) (verified as of 2026-09-23) [source]
- Wikipedia's Manual of Style defines weasel wording as vague attribution that dresses a statement in authority with no substantial basis; its examples include 'some people say', 'many scholars state', 'it is believed', 'experts declare', 'it is often reported'. (high confidence) [source]
- A 2009 Wikipedia study grouped weasel wording into numerically vague expressions, passive voice that avoids accountability ('it is said'), and weakening adverbs ('often', 'probably'). (low confidence) [source]
- An interuniversity style guide (UPC) gives the same fix outside Wikipedia: 'Student registrations have been observed to be increasing' becomes 'Student registrations have been increasing'; a straightforward statement beats a vague attribution. (medium confidence) [source]
Children
- Anonymous authority phrases (frontier)
- Attribution vs voice in neutral prose (frontier)
- Connective words implying unsupported relations (frontier)
- Cultural-heritage significance framing (frontier)
- Dropping attribution for undisputed claims (frontier)
- Editorial commentary in neutral prose
- Overgeneralised consensus from one source (frontier)
- Passive voice hiding the agent (frontier)
- Peacock terms and undue emphasis
- Positive-sentiment drift in preference-tuned models (frontier)
- Promotional adjective lists (frontier)
- Rewriting with named sources and plain claims
- Significance and legacy inflation (frontier)
- Travel-brochure and promotional register
- Vague attribution and weasel wording
- Verifying rewritten attributions against source text (frontier)
Frontier under this node: Anonymous authority phrases, Attribution vs voice in neutral prose, Connective words implying unsupported relations, Cultural-heritage significance framing, Dropping attribution for undisputed claims, Overgeneralised consensus from one source, Passive voice hiding the agent, Positive-sentiment drift in preference-tuned models, Promotional adjective lists, Significance and legacy inflation, Verifying rewritten attributions against source text