In sum: if you try ChatGPT on your own, you'll probably find that it knows English very well. It makes fewer grammatical errors than I do, and it excels at emulating specific styles of writing. Its responses, at times, sound remarkably human.
Unfortunately, you might also find that it's picked up some of our human foibles, and that it learned some of our biases along with our language. If I had to bet on the pronouns of a noir detective named Sam Striker, "he" would be a good guess. On the other hand, if I were emailing a Sam Striker, I'd make sure they weren't Samantha Striker before calling them that. GPT-3 and RoBERTa don't know what they don't know, and so they don't have our impulse to avoid assuming what we shouldn't.
Because these models have been trained on what we as a people have written, it provides a fascinating glimpse into our biases and follies. Research on these models tends to focus, reasonably, on fixing these biases and making the models better. That's important work, and we owe it to ourselves to make sure that AI doesn't thoughtlessly perpetuate our own mistakes.
While these biases remain, however, there's a lot of interesting things to learn about our world from them. Let's look at how language models conceptualize the words we attach perhaps more significance to than any other—our names.
In sum: if you try ChatGPT on your own, you'll probably find that it knows English very well. It makes fewer grammatical errors than I do, and it excels at emulating specific styles of writing. Its responses, at times, sound remarkably human.
Unfortunately, you might also find that it's picked up some of our human foibles, and that it learned some of our biases along with our language. If I had to bet on the pronouns of a noir detective named Sam Striker, "he" would be a good guess. On the other hand, if I were emailing a Sam Striker, I'd make sure they weren't Samantha Striker before calling them that. GPT-3 and RoBERTa don't know what they don't know, and so they don't have our impulse to avoid assuming what we shouldn't.
Because these models have been trained on what we as a people have written, it provides a fascinating glimpse into our biases and follies. Research on these models tends to focus, reasonably, on fixing these biases and making the models better. That's important work, and we owe it to ourselves to make sure that AI doesn't thoughtlessly perpetuate our own mistakes.
While these biases remain, however, there's a lot of interesting things to learn about our world from them. Let's look at how language models conceptualize the words we attach perhaps more significance to than any other—our names.