3 ms·
If OpenAI can modify their models so that they don't output human images, would it really be so hard to modify GPT so it doesn't output names? For example: > p
by bogwog 4y ago
If OpenAI can modify their models so that they don't output human images, would it really be so hard to modify GPT so it doesn't output names? For example:
> prompt: "Who was the first president of the United States?"
> response: "The first president of the United States was Aw@e%%t3R!35"
Sure, it'd make GPT less useful if it garbled all names, but that's a tradeoff made for the sake of ethics in the case of image generation. I don't see why this situation should be any different?
- nnx 4y agoTo avoid losing too much utility, could tweak the filter to allow a (large) whitelist of public figures. Anyone not in the list would be Aw@e%%t3R!35.
- 542458 4y agoMy initial idea would be to scrape allowed names from Wikipedia. Or maybe only allow a name if it appears more than N times in the corpus… Still, both of those would create issues with name collisions and common names, but it’s probably better than nothing. There’s also the issue of recognizing what a name is - is April a name? Joy? JFSmith1982? 542458?
- saalweachter 4y agoName detection is hard once you move beyond top names in a particular set of countries. And it's precisely the names that are most identifiable that are hardest to detect.
- nonameking2026 4y ago[flagged]