Modern artificial intelligence models from Anthropic, Google, and OpenAI process millions of queries daily, yet a comprehensive study reveals a critical performance flaw: their answers are frequently too complex for humans to easily understand, failing standard readability metrics across all major chatbot platforms.
The Readability Gap in Leading LLMs
We obsess over parameter scaling, token limits, and benchmark accuracy, but we rarely ask if the end user can actually decode the output. According to a linguistic investigation conducted by Munich-based text platform Wortliga on behalf of SEO firm Sistrix, leading large language models fail to hit basic clarity targets. The study evaluated 2,112 text outputs from Anthropic, Google, and OpenAI, analyzing them against established readability scales.
Not a single tested model achieved the recommended Wortliga target range of 60 to 70 points. Instead, the models generated dense, convoluted prose loaded with unnecessary jargon and nested dependent clauses. While an LLM might successfully retrieve a correct fact, its delivery mechanism often obscures the information.
| AI Model | Developer | Average Wortliga Score (Target: 60-70) |
|---|---|---|
| Claude Opus 4.7 | Anthropic | 47.7 |
| Gemini 3.1 Pro | 46.8 | |
| GPT-5.5 | OpenAI | 37.7 |
Anthropic’s Claude Opus 4.7 claimed the top spot with an average score of 47.7 points, followed closely by Google’s Gemini 3.1 Pro at 46.8 points. OpenAI’s GPT-5.5 lagged significantly behind with an average score of 37.7 points. These metrics prove that technical correctness does not equate to communicative utility.
As study lead Gidon Wagner notes, a correct answer is not yet a usable answer. Users must be able to understand, verify, and apply the information in real-world scenarios.
Why Complex Output Threatens Enterprise and Consumer Trust
When an LLM buries critical metrics under academic syntax or bureaucratic phrasing, the cognitive load shifts entirely to the reader. Long, winding sentences obscure dependencies and risks. This linguistic obfuscation introduces serious hazards when dealing with high-stakes domains like software engineering, financial planning, and healthcare diagnostics.
Users cannot critically audit an AI-generated explanation or trace a logical fallacy if the prose is designed like a legal disclaimer. Complex phrasing acts as a smokescreen, hiding hallucinations and logical inconsistencies behind an authoritative tone.
Johannes Beus, founder and CEO of Sistrix, points out that language clarity has evolved past a mere subjective preference. Accessibility guidelines and machine-readable data standards now demand rigorous plain-language optimization as a core quality metric.
The Chameleon Effect: How Prompt Styling Backfires
The study also uncovered the Chamäleon-Effekt in how models adapt to user inputs. When researchers fed models heavily bureaucratic or overly formal prompts, the systems mirrored that exact stylistic friction.
In those trials, the average readability score plummeted to a dismal 4.4 points. The AI adopted the same heavy, inaccessible phrasing present in the input query.
Conversely, explicitly instructing a model to write plainly using a direct prompt like “write comprehensibly” triggered a dramatic surge, pushing the average Wortliga score up to 79.4 points. However, this fix introduced a new failure mode. At that extreme, the outputs frequently overcorrected, reading as childish, overly simplified, or stripped of necessary technical nuance.
Plain-language generation requires a delicate calibration. Developers must build native alignment protocols that prioritize cognitive ergonomics without sacrificing the exact technical vocabulary engineering workflows demand.
The 30-Second Verdict: Raw algorithmic correctness is no longer enough to differentiate market leaders. As AI integration deepens across enterprise architectures, platforms must solve the readability bottleneck to ensure users can safely verify and execute generated insights.
