Controlling Human Perception of Basic User Traits

Daniel Preotiuc-Pietro, Sharath Chandra Guntuku, Lyle H. Ungar

Abstract

Recent advances in Natural Language Processing (NLP) have enabled highly accurate prediction of user traits such as age and gender from text. However, the ability to systematically control how humans perceive these traits remains less explored. This paper investigates a text selection approach to manipulate human perception of an author's gender and age. By using automated classifiers to filter a user's tweets, researchers from the University of Pennsylvania demonstrate that human gender-prediction accuracy can be reduced to near-random chance or significantly amplified. Similarly, the perceived age of users can be shifted up or down through targeted content selection.

Introduction

As AI becomes increasingly integrated into daily life through virtual assistants and chatbots, the ability to adapt a "persona" is becoming a critical requirement for effective interaction. Beyond personalization, controlling trait perception is vital for privacy, allowing users to obfuscate sensitive characteristics from both algorithms and human observers. This study addresses the feasibility of controlling human perceptions of gender and age using automated text selection methods.

Methodology

The researchers framed the challenge as a text selection task. Rather than generating new text, which often suffers from issues of fluency and coherence, they employed a selection strategy based on automated scoring:

  1. Automated Classification: A regularized Linear SVM classifier was trained to score thousands of tweets for gender and age.
  2. Selection Criteria: For each user, 20 tweets were selected based on four conditions:
  3. Random: A baseline selection of tweets.
  4. Same/Opposite: Tweets predicted most strongly as the user's actual gender versus the opposite gender.
  5. Youngest/Oldest: Tweets predicted as belonging to the youngest or oldest age brackets.
  6. Human Evaluation: The selected tweet sets were presented to human raters on Amazon Mechanical Turk (MTurk), who guessed the author's traits and provided confidence ratings.

The study utilized verified Twitter datasets for gender (200 users) and age (200 users, ages 15–34), with preprocessing to remove metadata that could leak identity.

Key Results

Gender Manipulation

The results demonstrate a significant capacity to influence human perception of gender. In the baseline Random condition, human accuracy was approximately 76.7%. However, when presented with tweets selected for the Opposite gender, accuracy plummeted to 55.7%, effectively reaching near-random chance. Conversely, the Same condition amplified perceived traits, raising accuracy to 91.3%.

Human gender prediction accuracy across different selection conditions.

Age Manipulation

Age perception proved more resistant to manipulation than gender, yet significant shifts were still achieved. The model successfully moved the mean predicted age: in the Youngest condition, the mean predicted age was 22.0, compared to 26.4 in the Oldest condition for the same set of users.

Real vs. Predicted age for Younger, Random, and Older selection conditions.

Discussion and Insights

The study revealed an asymmetry in gender perception; males were more easily mistaken for females than vice versa. Interestingly, human confidence was often misplaced; in the "Opposite" condition, raters who incorrectly identified a male as female were more confident than those who were correct. This suggests that specific linguistic "anchors" (e.g., "dress") strongly bias human intuition regardless of accuracy.

Age perception appears "stickier," relying on a broader range of stylistic cues that are harder to mask through simple selection. While topical markers (e.g., mentioning family) influence age perception, stylistic markers like "literally" or "so" remain potent indicators of youth.

Limitations and Ethics

While text selection is an effective proof-of-concept, future work should focus on text transformation—rewriting sentences to change perceived traits while preserving meaning. Additionally, the study noted a "topic confound," where selection often prioritized stereotypical subjects.

Ethically, the authors emphasize the need for transparency and user control. Systems that manipulate persona should inform users and provide the option to disable such features to prevent deceptive or unwanted profiling.

Summary updated

BibTeX
@inproceedings{preotiucpietro2017controlling,
  author = {Daniel Preotiuc-Pietro and Sharath Chandra Guntuku and Lyle H. Ungar},
  title = {{Controlling Human Perception of Basic User Traits}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2017},
  doi = {10.18653/v1/d17-1248},
}
Figure