Ellipsis research finds ChatGPT salary estimates are wrong by 23.3%
Ellipsis and Figures tested 10,000 European salary predictions. ChatGPT missed total compensation by 23.3% on average and C-level compensation by 61.1%.
Ellipsis research conducted with Figures found that ChatGPT’s total-compensation estimates were wrong by 23.3% on average across 10,000 European salary predictions.
How inaccurate were ChatGPT’s salary estimates?
ChatGPT’s base-salary predictions had an average error of 16.8%. Its total-compensation predictions had an average error of 23.3%.
The error increased with seniority. Total-compensation predictions for C-level roles were wrong by 61.1% on average. One Paris Chief Marketing Officer estimate missed the verified figure by 62.5%.
Sixty percent of predictions were too high and 37% were too low. Across the dataset, GPT-5.1 overestimated European salaries by about 9%.
How did Ellipsis and Figures test ChatGPT?
Ellipsis ran 10,000 prompts covering 200 roles in Paris, London and Munich. The study used 50 prompt variations per role and compared the answers with Figures’ verified European salary data.
Prompt style changed error rates by only 2.3 percentage points across the ten prompt categories tested. The evidence points to a knowledge problem rather than a prompt-writing problem.
What are the limits of the salary study?
The comparison data comes from European companies participating in Figures’ dataset. It does not represent every employer or compensation package. The research selected the most reliable available data points and measured prediction error against that defined sample.