社團指南

AI 圖像生成器中的偏見和刻板印象

AI 影像生成器常常重現並誇大刻板印象,例如在「執行長」或「工程師」的描述中大多顯示男性,在高階職位中大多顯示膚色較淺的人,並且呈現狹隘的美貌標準,因為它們是從大量偏頗的、附有說明的網路影像集合中學習而來。

聆聽此指南AI 語音敘述 · ~3 min

本頁4 分鐘閱讀
  1. 概述
  2. 深入探討
  3. 現實世界的實施
  4. AI 影像生成器中偏見與刻板印象的未來走向
  5. 付諸實踐
  6. 不斷探索
  7. 常見問題

概述

研究人員透過對中性提示產生大量圖像並比較出現的人物來衡量此問題。這很重要,因為這些圖像現在充斥於廣告、教科書與新聞,而偏見與拙劣的修正方式都可能誤導大眾。

深入探討

Text-to-image models such as Stable Diffusion, Midjourney and DALL-E learn the link between words and pictures from huge datasets of image-caption pairs, largely scraped from the web; Stable Diffusion, for example, was trained on subsets of LAION-5B. The web over-represents some countries, languages, body types and occupations, and captions reflect the assumptions of whoever wrote them. The model then learns the most typical picture for a word and tends to amplify it, because generation favors high-probability, 'average' outputs. A prompt that says nothing about gender or ethnicity still produces a default, and that default is often the stereotype. Researchers measure this in a few ways. A common method is to generate hundreds of images for neutral prompts (occupations, adjectives like 'attractive' or 'poor', nationalities) and label perceived gender, skin tone or age, sometimes compared against labor statistics. Hugging Face researchers' 'Stable Bias' project (2023) used clustering of model outputs rather than assigning labels directly, partly because labeling someone's race from a picture is itself fraught. A 2023 Bloomberg analysis of thousands of Stable Diffusion images found that high-paying jobs skewed toward lighter-skinned men and low-paying jobs toward darker skin tones. Fixes are harder than they look. In 2022 OpenAI said it applied a technique so DALL-E 2 produced more diverse people when prompts did not specify, widely reported to involve adding terms to prompts behind the scenes. In February 2024 Google paused Gemini's generation of images of people after it produced historically inaccurate images, such as racially diverse depictions of groups that were historically homogeneous. The lesson is that bias is not a single dial: a system must be representative where the prompt is open and accurate where the prompt is specific. A common misconception is that bias only comes from the training data; model design, filtering, captioning and post-training choices all shape the result too.

現實世界的實施

一位教師向影像生成器請求「實驗室裡的科學家」二十次,結果幾乎全是穿實驗袍的白人男性,於是重新撰寫提示詞,要求出現特定且多樣化的人物。

一個行銷團隊在產生「護士」與「醫師」的圖像時,發現工具總是把護士畫成女性、醫師畫成男性,於是加入審核步驟,確保圖像在投入活動前先檢查。

一位記者比較「尼日利亞的房子」與「美國的房子」的輸出結果,發現模型對前者預設貧困意象,對後者則預設郊區住宅。

使用者要求一個特定族群的歷史場景,卻收到插入了歷史上不存在的人口統計資料的圖像,顯示多樣性修正如何過度校正。

AI 影像生成器中偏見與刻板印象的未來走向

Expect more standardized bias benchmarks, more disclosure in model cards, and more context-aware mitigation that distinguishes open-ended prompts from historically or personally specific ones. There is genuine disagreement about the right target: matching current real-world statistics can entrench existing inequality, while uniform representation can misrepresent reality. That choice is a values decision, not purely a technical one, and organizations are likely to be asked to explain it. Users will still need to review generated images before publishing them, particularly in education and news.

付諸實踐

為什麼這很重要

  • 風險與安全. 災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
  • 更明確的決策. 民眾和專業素養決定強而有力的安全政策在政治上是否可行。
  • 突破炒作. 清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。

留意

  • 將存在風險視為科幻小說,同時能力複合。
  • 混淆了表面產品安全與高度自治下的對準。
  • 只給非英語和非專業觀眾留下低品質的資源。

後續步驟

  1. 單獨的產品危害、誤用和失控/失調風險。
  2. 詢問哪些證據會改變您對時間表和嚴重性的看法。
  3. 比起行銷主張,更喜歡主要來源和具體評估。
  4. 確定一條行動路徑:職業、政策、資金或技能——而不僅僅是意識。

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

Take the Bias and Stereotypes in AI Image Generators quiz

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

What is Bias and Stereotypes in AI Image Generators?

AI image generators often reproduce and exaggerate stereotypes, for example showing mostly men for 'CEO' or 'engineer', mostly lighter-skinned people for high-status jobs, and narrow beauty standards, because they learn from large, skewed collections of captioned internet images. Researchers measure this by generating many images for neutral prompts and comparing who appears. It matters because these images now fill ads, textbooks and news, and both the bias and clumsy fixes can mislead people.

Why does a prompt that does not mention gender or ethnicity still produce a biased image?

With no specification, the model produces a high-probability default, and that default mirrors and often amplifies stereotypes in its training data.

模型像Stable Diffusion這類會學習什麼樣的資料?

像LAION-5B這類資料集包含數十億個網路抓取的圖像-文字對,這些對承載著網路的不平衡。

研究人員常用什麼方法來衡量圖像生成器的偏差?

研究人員會針對中性提示生成數百張圖像,並分析其感知屬性,有時會與現實世界統計數據做對照。

為什麼Stable Bias 專案選擇聚類而非直接標註種族?

從外觀推斷種族不可靠且具倫理爭議,因此聚類輸出是一種在不加標籤的情況下比較模型的方法。

2023 年 Bloomberg 對 Stable Diffusion 圖像的分析發現了什麼?

分析發現高薪職業偏向膚色較淺的男性,而低薪職業則偏向膚色較深的人。