Plain-language & easy-read simplification (Leichte Sprache)

Millions of people find official notices, contracts, websites and patient leaflets hard to read. They include people with learning disabilities, low literacy, dementia or aphasia, language learners and many older readers. Human translation into Leichte Sprache or plain language is slow and capacity is limited. LLM-based simplification now produces a rule-conformant draft in seconds. Trained translators and reviewers from the target group check and correct the draft before it is published. This turns accessibility duties under BGG and BITV 2.0, and the general push for understandable communication, from occasional projects into routine production.

Problem & value

• Who has the problem: authorities, municipalities and public broadcasters are obliged to communicate understandably. Insurers, utilities, banks, clinics and social organisations lose time and trust when customers cannot understand their letters. Private individuals receive letters they cannot decode.
• Legal drivers (DE): §11 BGG requires federal public bodies to explain notices (Bescheide), general orders, public-law contracts and forms in plain language on request. Where that is not enough, they must explain them in Leichte Sprache, and they should provide more information in Leichte Sprache generally [5]. BITV 2.0 §4 requires public websites to offer start-page explanations in Leichte Sprache and German Sign Language. Anlage 2 sets the Leichte Sprache rules: short sentences, no passive, subjunctive, genitive or negation, hyphenated compounds, at least 1.2 em font size, one sentence per line [4].
• Standards: DIN SPEC 33429 "Empfehlungen für Deutsche Leichte Sprache" (March 2025) covers language, layout, images, the production process and the qualifications of writers and reviewers. It can serve as tender criteria [3]. ISO 24495-1:2023 covers plain language generally. Inclusion Europe's "Information for all" is the European easy-to-read standard [8].
• AI value: a draft in seconds instead of days of translator time, consistent terminology, and coverage of long-tail content (every web page, every letter template) that was never affordable to translate by hand.
• Measured adoption: SUMM AI reports more than 1 million translated texts and more than 200 organisations using its Leichte Sprache product (vendor figures) [1]. No independent studies with time or cost savings were found for this review. Measure your own baseline.

Variants

• Leichte Sprache / easy-read (≈A1–A2): strictest rules. The target group is people with learning disabilities. Review by target-group testers is part of the standard. International equivalents: UK Easy Read, Spanish Lectura Fácil, French FALC, Finnish selkokieli.
• Einfache Sprache / plain language (≈B1–B2): for the general public. Used for notices, insurance letters, terms and conditions, energy bills and news. Example: Deutschlandfunk's nachrichtenleicht publishes news in Einfache Sprache [12].
• CEFR-targeted simplification for language learners and integration courses. This is the setting of the TSAR 2025 shared task [9] and links to EDU-009 (differentiation).
• Domain variants: patient information and discharge letters (HLTH-011, PERS-014), letters from authorities and their answers (PERS-008), technical and safety instructions (LANG-022), legal and contract summaries.
• Multilingual: simplify, then translate (LANG-009), or simplify directly in the target language. SUMM AI and capito both offer foreign-language output [1][2].
• Private use: browser add-ins or chat assistants that "explain this letter simply".
• Out of scope: sign language (LANG-030), PDF/UA tagging and alt text (DOC-026), web code accessibility (CODE-017), plain summaries without the simplification rules (LANG-008).

How it works

1. source text (CMS page, letter template, PDF via DOC-026/DOC-014 parsing)
2. segmentation + protected terms (names, amounts, deadlines, legal references)
3. LLM rewrite with level-specific system prompt + rule set + glossary (few-shot from approved texts)
4. automatic checks: rule linter (sentence length, passive, negation, genitive, compounds), readability score (e.g. Amstad/Flesch-DE, Wiener Sachtextformel, CEFR classifier)
5. meaning check: second LLM pass / NLI compares facts, numbers, dates, obligations; diff of omissions/additions
6. human translator edits + target-group review (Prüfgruppe) for Leichte Sprache
7. layout (one sentence per line, ≥1.2 em, pictograms)
8. publish (CMS, PDF, print), optional TTS audio

• Core model: an instruction-following LLM. Simplification is a style-and-content transformation, not word substitution. Good systems explain terms instead of deleting them, and they restructure the text with the most important information first.
• Checks: rule-based linters plus small classifiers estimate the readability level. A comparison pass flags lost obligations, deadlines or amounts. capito, for example, shows omitted or added information explicitly [2].
• Human-in-the-loop: mandatory for anything legally relevant (notices, contracts, medical information). For certified Leichte Sprache the target-group review step cannot be replaced. SUMM AI also states that final approval stays with the organisation [1].

Data preparation

The source material is mostly finished text: web pages in the CMS, notice and letter templates in the document system or ERP, forms and leaflets as PDF, and contract terms in the DMS. Much of it exists only as print-ready PDF or scan, and the same wording lives in many copies.

Typical gaps are scanned notices without text, outdated templates, legal terms with no approved explanation, and no list of protected terms. Letters generated for individual customers contain personal data and must be handled with care.

Preparation steps: 1. Select the 20 to 50 texts with the highest reach, such as top web pages and standard letters, rather than the whole archive (PREP-002). 2. Extract clean text from PDFs and scans (DOC-014, PREP-006). 3. Build a glossary of protected terms and approved plain explanations (PREP-015). 4. Collect gold-standard examples already approved by reviewers or the Prüfgruppe of people with learning difficulties (PREP-016). 5. Use templates with placeholders so that personal data never enters the model (PREP-022).

Effort: low, about 10 to 15 percent. The real effort is in review, not data.

Implementation paths

Which path for whom

• Public: Path C, because authorities handle citizen data and legally relevant notices, and sovereign open-weight models on German infrastructure meet procurement and accessibility duties.
• SMB: Path A, because specialised easy-read tools and mainstream assistants with review workflow are enough for a few documents per month.
• Enterprise: Path B, because insurers, banks and utilities need the rewrite embedded in their CMS and document factory with rule checks and approval steps.
• Developer: Path B, because cloud APIs plus a rule checker are the fastest way to build accessible-content features.
• Private: Path A, because a chat assistant or browser tool is enough to understand a letter.

Path A — Buy (recommended for most organisations)

• Specialised simplification tools (EU): SUMM AI (DE). Covers Leichte and Einfache Sprache, oriented to DIN SPEC 33429, plus 51 foreign languages. Offers a web app, CMS plugin, API and live website translator. Hosted in DE/EU with an Art. 28 GDPR DPA and an optional training exclusion. References include German municipalities, the Berlin Senate Chancellery and Deutsche Bahn [1]. capito.ai (AT). Three plain-language levels with a level score, word explanations, an omission/addition diff and custom dictionaries. Available as Word, Outlook and browser add-ins, through an API, and on-premises [2].
• General writing tools: DeepL Write (DE) has a simple writing style for German and other languages, available via the API [7]. Microsoft 365 Copilot and Google Gemini in Workspace offer "rewrite simpler" prompts but do not enforce Leichte Sprache rules.
• Human services: Leichte Sprache translation offices and review groups (e.g. members of Netzwerk Leichte Sprache [10]). They are still needed for certification and target-group testing.
• Choose A when: you need standard-conformant output quickly and integration via a CMS plugin or Word add-in is enough.

Path B — Build on cloud APIs

• Models: Mistral Large / Medium (EU provider, EU hosting), Claude, GPT and Gemini via EU data-residency regions (Azure OpenAI EU Data Zone, Vertex AI europe-west, Amazon Bedrock Frankfurt). Frontier models handle German Leichte Sprache rules well when given a precise rule prompt and examples.
• Design: a level-specific system prompt encodes the DIN SPEC and BITV rules. Use few-shot examples from your own approved texts, a glossary injected via retrieval, and structured output (text plus a list of explained terms plus a list of changed facts). Add a second "fact-checker" call that compares source and output.
• Integration: CMS (TYPO3, WordPress, Drupal) workflow states such as draft → review → publish. Letter templates in DMS or customer-communication tools. n8n or Power Automate for batch jobs.
• Choose B when: simplification must sit inside your own portal or letter generation, or you need house-specific rules.

Path C — Self-host (open source)

• Models: Qwen3 (e.g. 14B/32B), Mistral Small 3.x, Gemma 3 (12B/27B) or Llama 3.3 70B (HW3), served via vLLM or Ollama. EU-built options with strong German include EuroLLM and Teuken-7B, which are better suited as a base for fine-tuning than as zero-shot rewriters.
• Tooling: spaCy (German) for the rule linter (passive, negation, sentence length, compounds); textstat or own formulas for German readability; EASSE for SARI evaluation; the DEplain corpora and models on Hugging Face for evaluation and fine-tuning [6][11].
• Hardware: HW2 (one 24–32 GB GPU) runs a quantised 14–32B model for interactive use. HW3 is needed for 70B-class quality or many concurrent users.
• Choose C when: texts contain sensitive personal data (social benefits, health, case files) and must stay on-prem, e.g. municipal IT service providers.

Path D — Custom training

Worthwhile for large public bodies, broadcasters and vendors with thousands of approved standard/simplified text pairs. LoRA fine-tuning of a 7–14B model on in-house pairs teaches house style and terminology and cuts inference cost. Research shows that style-specific pre-training on unaligned Leichte Sprache text reduces the need for scarce parallel data [11]. DEplain gives a public starting point of about 13k aligned news sentence pairs and about 750 web document pairs [6]. For most users, prompting plus glossaries is enough.

State of the art (2026)

• From seq2seq to LLMs: early German systems were fine-tuned mBART and mT5 models on DEplain and APA data [6][11]. Instruction-tuned LLMs now dominate for document-level simplification, because the task needs restructuring, explanation and layout, not just sentence shortening.
• Evaluation: SARI and BLEU correlate poorly with human judgement on document-level output. Current practice combines several signals: a level classifier (e.g. CEFR compliance), meaning preservation measured as semantic similarity, and human ratings. The TSAR 2025 shared task scored systems this way on English A1/A2/B1 targets [9].
• Products have added DIN SPEC 33429 orientation, word explanations, omission diffs and CMS live translation [1][2].
• Failure modes:
• Meaning loss: deadlines, conditions, legal consequences or exceptions get dropped. This is the biggest risk.
• Hallucinated explanations of legal or medical terms.
• Over-simplification that turns into a patronising tone.
• Inconsistent terms across documents.
• Rule violations: passive voice and negation creep back in.
• Weaker quality for languages other than German and English.
• Not a substitute for target-group testing. The Leichte Sprache community insists on review by people with learning disabilities [10].

Shared resources & combinations

• Same LLM endpoint or GPU (llm-general, HW2/HW3): LANG-011 proofreading and brand voice (the same rewrite pipeline with a different rule set), LANG-008 summarisation, LANG-009 translation (glossary and terminology services shared too), LANG-022 technical documentation.
• Accessibility bundle: DOC-026 (accessible PDF), CODE-017 (web accessibility), LANG-030 (sign language), SPEECH-004 (captions) and TTS read-aloud. One "barrier-free publishing" pipeline in the CMS serves BITV 2.0 and the European Accessibility Act together.
• Citizen and customer communication: CONV-011 citizen chatbot (answer in plain language), AGENT-026 case processing (understandable notices), PERS-008 and PERS-014 (explaining letters and medical reports to individuals), HLTH-011 patient-friendly discharge letters.
• Shared building blocks: a glossary or termbase, a rule-linter service (spaCy), a readability classifier, a review workflow in the CMS (workflow-automation), and an optional pictogram library or image generation for illustrations.

Effort, cost & KPIs

• Path A: days. SaaS priced per user, per volume or per website plugin. Budget separately for human review and target-group testing, which remain the main cost driver for Leichte Sprache.
• Path B: 2–6 weeks for a CMS-integrated pipeline with rule prompt, linter and fact check. Token cost is negligible compared with translator time.
• Path C: 1–2 months including model selection, evaluation on your own texts and GPU operations.
• Path D: 2–4 months and requires a curated parallel corpus.
• KPIs:
• turnaround time per text
• share of content available in Leichte or Einfache Sprache
• reviewer edit distance (post-editing effort)
• meaning-error rate found in review
• readability or level score
• comprehension test results with target readers
• fewer follow-up calls or objections on notices

People & change

Involve management and the department lead (communication, citizen service, product) to set the target level and which texts qualify. Legal decides which texts may be simplified at all, because a simplified notice must stay legally correct and may only supplement the binding original. The DPO is needed when letters with personal data are processed. Involve customers or citizens early: test readers with learning disabilities, ideally via a Prüfgruppe, check understanding, and their approval is part of the quality.

Works council: depends. The tool mainly processes texts, not employee behaviour, so §87(1) no. 6 BetrVG applies only if the platform logs staff activity. Inform the council under §90(1) no. 3 when editors' work processes change. In authorities, involve the Personalrat.

AI literacy: train editors on typical failure modes (dropped negations, changed meaning, invented explanations) as a support measure under Art. 4 AI Act (GUIDE-006). Communicate that AI produces the draft and trained reviewers approve every text (GUIDE-008). Start with a pilot group of two editors and one reviewer team, measure review time and comprehension, and keep the human release step mandatory. Add a note on the page that the text was created with AI support where Art. 50 applies (GUIDE-014).

Risks, regulation & compliance

• EU AI Act: text simplification is not an Annex III use and not an Art. 5 practice. Art. 50(4) (applies from 2 August 2026) requires deployers to disclose AI-generated text published to inform the public on matters of public interest. This does not apply where the text has undergone human review or editorial control and someone holds editorial responsibility [13]. A reviewed workflow therefore normally avoids labelling duties, but label unreviewed live website translation. AI literacy under Art. 4 applies. If simplification is embedded in a high-risk system (e.g. benefit decisions), it inherits that system's obligations.
• Legal accuracy: the simplified version is usually explanatory. The original notice remains legally binding, so state this and link to the original. Never let AI change amounts, deadlines or legal remedies (Rechtsbehelfsbelehrung).
• GDPR: generic web content contains no personal data. Individual letters, case files and health data do (Art. 9). Use EU hosting, an Art. 28 DPA, no training on customer data, or Path C. SUMM AI offers DE/EU processing and a training exclusion [1].
• Sector rules: for patient information, medical-device labelling under MDR or medicinal product leaflets (AMG) remain the responsibility of the regulatory function. AI drafts need regulatory sign-off. For insurance and finance, mandatory pre-contract information must not be shortened.
• Quality and dignity: involve self-advocacy groups. Avoid patronising tone. Use the Inclusion Europe logo only if its standards are met [8].

Getting started

1. Inventory your obligations (BGG §11, BITV 2.0 §4, EAA) and the top 20 texts by reach and complaint volume: start pages, common notices, letter templates.
2. Test 2–3 tools (e.g. SUMM AI, capito.ai, DeepL Write simple, a frontier LLM with a DIN SPEC 33429 rule prompt) on 10 of your own texts. Have a professional translator rate meaning loss and rule conformity.
3. Build a glossary of protected terms and approved explanations. Define which text types need full Leichte Sprache review and which need only plain-language editing.
4. Set up the workflow: AI draft → linter and fact check → editor → target-group review → publish. Record who holds editorial responsibility (Art. 50(4)).
5. Track turnaround, edit effort and reader feedback. Feed approved pairs back as few-shot examples, or later as fine-tuning data.

Sources

1. SUMM AI — product and data-protection information, reference customers and usage figures — https://summ.ai/
2. capito — capito.ai digital tool (levels, add-ins, API, on-premises) — https://www.capito.eu/digital/
3. DIN Media — DIN SPEC 33429:2025-03 "Empfehlungen für Deutsche Leichte Sprache" — https://www.dinmedia.de/de/technische-regel/din-spec-33429/387728031
4. Barrierefreie-Informationstechnik-Verordnung (BITV 2.0), §4 and Anlage 2 — https://www.gesetze-im-internet.de/bitv_2_0/BJNR184300011.html
5. Behindertengleichstellungsgesetz (BGG) §11 Verständlichkeit und Leichte Sprache — https://www.gesetze-im-internet.de/bgg/__11.html
6. Stodden, Momen, Kallmeyer (ACL 2023) — DEplain: A German Parallel Corpus with Intralingual Translations into Plain Language — https://aclanthology.org/2023.acl-long.908/
7. DeepL — Write API writing styles (simple, business, …) — https://developers.deepl.com/docs/translate/controlling-writing-style-and-tone
8. Inclusion Europe — European easy-to-read standards ("Information for all") — https://www.inclusion-europe.eu/easy-to-read-standards-guidelines/
9. TSAR 2025 Shared Task on Readability-Controlled Text Simplification — https://tsar-workshop.github.io/shared-task/
10. Netzwerk Leichte Sprache e.V. — https://www.netzwerk-leichte-sprache.de/ls/
11. Anschütz et al. (Findings ACL 2023) — Language Models for German Text Simplification: Style-specific Pre-training — https://aclanthology.org/2023.findings-acl.74/ ; DEplain models/datasets — https://huggingface.co/DEplain
12. Deutschlandfunk — nachrichtenleicht (news in Einfache Sprache) — https://www.nachrichtenleicht.de/
13. EU AI Act Art. 50 (transparency obligations) — https://artificialintelligenceact.eu/article/50/
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj