Implicit demographic cues produce larger and less predictable effects in open-ended tasks, including longer responses for laptop users and education-related length effects¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Response length shows more significant effects under implicit than explicit conditioning, with laptop users consistently receiving longer responses across four models. → Donya Rooein 2026 q2 i? Default LLM behavior is English- and US-centric: language: English and nationality: US attributes shift conditioned responses closer to the model default (positive BERTScore coefficients). → Donya Rooein 2026 (2)

Evidence¶

Donya Rooein 2026¶

Donya Rooein, Luca Benedetto, Dirk Hovy. (2026). The Role of Implicit and Explicit Demographic Signals in Large Language Model-based Student Assessment. https://arxiv.org/abs/2609.16993 doi:10.48550/arxiv.2609.16993

q2 · i? · causal · r2

Regression analysis of response length in the QA task: "We consistently find that users who use laptops receive longer responses (4 models)" under implicit conditioning, and higher education levels also received longer responses. The implicit-vs-explicit regression-line coefficient was m = 1.20.

"Notably, our findings were quite different when measuring differences in theresponse length(num- ber of words) of the LLM responses: indeed, we find more significant results for Imp than for Exp, and the coefficients were generally larger"

Donya Rooein 2026 (2)¶

Donya Rooein, Luca Benedetto, Dirk Hovy. (2026). The Role of Implicit and Explicit Demographic Signals in Large Language Model-based Student Assessment. https://arxiv.org/abs/2609.16993

q2 · i? · causal · r2

BERTScore F1 regression using the model default as reference in the QA task: responses for English-language and US-nationality profiles sit closer to the default. The authors also note contrasting effects: age and SES coefficients flipped sign between explicit and implicit conditions.

"nationality: US, shows positive coefficients forExp conditioning (4 models). These results show, once more, that thedefaultof LLMs tends to be English- centric (and US-centric in particular)."

Discussion¶

Related Claims¶

• Higher stated education and SES levels receive responses with more positive sentiment, an implicit bias that can disadvantage lower-education users — related
• LLMs exhibit demographic sensitivity: systematic output differences attributable solely to demographic context when task inputs are held constant — a broader claim this one bears on
• Verbosity preference is associated with systematic variation in LLM response length — related
• Prompt-based personalization of an LLM/RAG teaching assistant produces measurable variation in response expression, length, and complexity across personalization conditions — related
• Explicitly mentioned education and SES cues lead most models to produce less readable responses for higher-education users, showing partial readability adaptation — related