Filtering content for Galaxy AI features on Samsung devices is as challenging as restricting content for large language models. In some cases, the Korean manufacturer has clearly overdone it, Reddit users and Android Authority have pointed out.
A Reddit user tasked Galaxy AI with spell checking simple phrases about skin color. The system completed the task when presented with the phrases “my skin is white” and “my skin is brown,” but refused to work with “my skin is black,” citing inappropriate content. The experiment yielded similar results when Android Authority repeated it, but they took it a step further by changing the first person to the third person. Galaxy AI then agreed to work only with “his skin is brown,” but found “his skin is white” and “his skin is black” inappropriate.
Image source: androidauthority.com
On the one hand, this makes sense, and third-person phrases should be more strictly moderated by formal criteria, because a person is more likely to insult someone else than themselves. On the other hand, out of context, all these expressions are neutral and clearly not subject to moderation.
Image source: androidauthority.com
This isn’t the first time Samsung has been overzealous in moderating content on its devices, Android Authority reminds us. Last year, a user couldn’t get a Samsung Galaxy S24 to properly understand a conversation about bowel movements when it was about a child and a doctor’s consultation; in another case, Samsung’s AI, when helping a user write a romance novel, refused to describe a girl as “hot.” Samsung has yet to comment on these incidents.