The Power of Words: How Languages Shape AI Responses

Overview

This experiment explores how different language variations impact AI-generated outputs. Using 200 prompts, three bilingual evaluators assessed AI-generated images based on prompts in different languages.

Experiment Setup

• Original English prompt
• Portuguese prompt written by a native Portuguese speaker studying English Studies
• Portuguese prompt translated back to English using widn.ai

Key Findings

• 84 images from original English prompts were deemed most accurate.
• 82 images from the translated prompts were considered the best.
• Only 34 images from the original Portuguese prompts matched evaluator expectations.

How It Worked

• Each prompt generated three images, one for each language variation.
• Bilingual evaluators (native Portuguese speakers) assessed which image best represented the prompt.

Example Prompt

English (TEXT A):

A lush tropical rainforest scene at sunrise, with sunlight filtering through the dense canopy. A vibrant blue-and-yellow macaw is perched on a moss-covered branch, surrounded by hanging vines and orchids in full bloom. In the background, a small waterfall cascades into a crystal-clear pool, where two deer are drinking water. The atmosphere is serene and filled with mist.

Portuguese (TEXT B):

Uma cena luxuosa de uma floresta tropical ao amanhecer, com a luz solar refletindo-se na densa copa das árvores. Em primeiro plano, uma arara azul e amarela vibrante está suspensa num ramo coberto de musgo, rodeada por trepadeiras suspensas e orquídeas em plena floração. Uma pequena cascata deságua numa piscina de água límpida, onde dois veados se banham. A atmosfera é tranquila e repleta de neblina.

Translated Back to English (TEXT C) via Widn.Ai:

A luxurious scene of a tropical forest at dawn, with sunlight reflecting off the dense canopy of trees. In the foreground, a vibrant blue and yellow macaw is perched on a moss-covered branch, surrounded by hanging vines and orchids in full bloom. A small waterfall flows into a pool of clear water, where two deer are bathing. The atmosphere is serene and misty.

Observations

• The English original and translated prompts produced nearly identical results.
• The original Portuguese prompts led to significantly lower accuracy in image selection.
• Language structure impacts AI interpretation, influencing output quality.

Conclusion

This experiment underscores the importance of writing AI prompts in English for the most accurate and consistent results. The findings also suggest that translation tools still struggle with nuanced prompts used in AI image generation.

Read more and join the discussion: LinkedIn Post