“AI脸 生理性厌恶”登上热搜！什么是恐怖谷效应？

近期，“对AI脸感到不适”相关话题接连冲上社交平台热搜。

AI-generated faces may look flawless, but many people say something about them still feels "off". As the topic continues to trend on social media, the Beijing Science Center points to the "uncanny valley" — a psychological phenomenon describing the discomfort people feel when something appears almost, but not quite, human.

uncanny /ʌnˈkæni/: 神秘的; 可怕的; 怪异的; 危险的

打开短视频平台刷微短剧，连续刷几部后，主角个个是高颅顶、大眼窄鼻、冷白皮，精致得挑不出错，可表情僵硬空洞，看着看着心里一阵发毛，关掉手机却说不清到底哪里不对。

网友评论

这种说不清道不明的“不舒服”到底是什么？为什么会让人不适？

北京科学中心近日发文称，这种“似人非人”的违和感，便是“恐怖谷理论”。

1970年，机器人学家森政弘提出：当仿生物体的拟人程度持续提升时，人类对其的好感度会同步上升；但当拟人度接近真人，却又在某些关键动态特征（如眼神变化、面部肌肉联动）上出现可被感知的偏差时，人类的好感会骤然下跌，产生疏离、排斥甚至不适的负面情绪。这个好感急剧下坠的区段，就是“恐怖谷”（uncanny valley）。

The concept was introduced in 1970 by Japanese robotics professor Masahiro Mori. He proposed that as robots or artificial figures become more humanlike, people's emotional response generally becomes more positive. However, when a figure looks almost identical to a real person but still contains subtle imperfections, people's affinity can suddenly drop, giving way to feelings of unease or even revulsion. Mori described this sharp decline in emotional response as the "uncanny valley".

图源：数字北京科学中心

真实人类的面容天然带有个体差异：肤色有自然色差，皮肤有纹理与小瑕疵，五官有独属于个人的特征。这些细节正是构成面容真实感的重要组成部分。

AI生成模型训练数据多偏向端正、高吸引力的面容，加之制作方倾向于选取符合主流审美的生成结果，最终呈现的面孔很容易趋近一种“审美均值”，个体特征被大幅抹平。

One reason AI-generated faces can trigger this effect is the way many AI image-generation models are trained. These models often rely on datasets dominated by symmetrical, conventionally attractive faces. Combined with users' preference for aesthetically pleasing results, AI-generated portraits tend to converge on an idealized, standardized appearance, with many distinctive features smoothed out.

此外，在现阶段，算法虽能拟合出符合大众审美的静态五官，却难以复现人类视觉系统高度敏感的生物运动信号——例如瞳孔的细微缩放、表情肌的协同收缩，以及光影随动作的自然流动。这些动态线索的缺失或失真，会直接触发大脑的知觉冲突反应。

At the same time, today's AI models are far better at reproducing static facial features than the subtle movements that make a face appear genuinely alive. Small details such as natural eye movements, coordinated facial muscle activity, tiny changes in pupil size, and shifting light and shadow during motion remain difficult to recreate convincingly. When these cues are missing or appear unnatural, the brain detects a mismatch between what it expects and what it sees.

当人的视觉系统捕捉到高度接近真人的脸，却同时识别出多处不符合真实人类生理特征的细节时，便会产生排斥、不适的情绪反应与回避倾向。

As a result, people may feel uncomfortable when looking at faces that appear almost real but still differ from genuine human expressions and movement in subtle ways.

图源：数字北京科学中心

对AI脸感到不适该怎么办？

▌接纳不适感。对AI脸感到别扭不是矫情，是人类知觉识别机制的正常反应，不必怀疑自己的感受。

▌拒绝审美趋同。人类对面孔独特性的敏感具有进化心理学基础——面部差异帮助我们识别他人、判断意图与健康状态。不必被算法定义的“完美长相”裹挟，接纳真实的个体差异，本质上是回归人类正常的社会认知本能。

Researchers say this sensitivity has deep evolutionary roots. Humans rely on unique facial features to recognize individuals, interpret emotions and intentions, and even make quick judgments about health. Rather than responding to algorithmically generated "perfect" faces, our brains are naturally tuned to the subtle variations found in real human faces.

▌多关注“活人感”视频。真正打动人心的内容，往往不只是靠无瑕疵的精致，带有真实细节的角色往往更有生命力。

网友评论

人民日报曾发表评论提醒，AI脸受到厌恶，既应引发反思，也应成为信号：可以利用技术压缩经费与制作周期，但不能用流水线产能替代艺术性和人文性；AI归根结底只是辅助性工具，专业的事还是要多交给专业的“人”；生产一百个“差不多”的产品只能争抢存量，创作一件“被记住”的作品才能创造增量。

真心换真心，观众愿意买单的，是被触动的感觉，它可能来自一句足够贴近生活的台词、一个会心一笑的动作，它的力量叫作真实。AI时代，有血有肉有“破绽”的内容，自有其珍贵。

你经常看AI短剧吗？一起来评论区分享你的看法。

来源：中国青年报综合数字北京科学中心、人民日报客户端、网友评论

跟着China Daily

精读英语新闻

“无痛”学英语，每天20分钟就够！