Limitations
Rehan N -
Hey guys! So with all the data collection and testing put behind me, I’ve finally had a change of pace and began to transition from the hustle of gathering results to more of the analysis with the write-up portion for my research paper.
A large part of this involves putting phase 1 (quantifying the extent of the biases) in comparison with phase 2 (adapting the detectors to mitigate bias). Watching the results in tandem has been pretty exciting since it’s clear that we saw improvements with the changed training data. That’s exactly what I hoped to see!
Of course, no research project is without its constraints though. First of all, although I did get a variety of text samples, there is still a world of dialects out there that I obviously couldn’t feasible capture. Second, although the open source text classifier I used can be retrained, the process might be different for commercial detectors that might be more complex and updated. Finally, generative AI in general is improving so what might be relevant today might not be as relevant tomorrow.
So now that I have a thorough grasp of the data, I’m beginning to tie my results back into the broader questions that were raised at the beginning of my project.
And while it’s gratifying to see the light at the end of the tunnel, there’s still plenty to do. That’s all for now, but I’m anticipating polishing up everything and sharing a comprehensive overview soon! As always, thanks for reading and if you are curious about anything, feel free to reach out.