Halfway Through: Three Weeks Into My Laidlaw Research
I remember my first meeting with my supervisor three weeks ago. Back then, I was finalizing my research design with him. These six weeks felt like an abstraction then. Now, I am on the other side of week three, and the abstraction has turned into a full Windows desktop with a growing number of folders full of scripts.
Week 1. My project investigates whether AI text-to-image generators skew towards a particular cultural lens when instructed with prompts that don't specify one, and how those same models interpret prompts that do name a specific culture. This may sound simple, but it becomes interesting when you sit down to design a way to actually measure it. I spent most of my first week into turning this idea into something specific enough to execute. I locked down two generation models, four world regions, and five separate analytical frameworks to run the outputs through. My supervisor, Prof. Ahmad, gave me a lot of freedom to make most of the methodological calls myself, which was both flattering and terrifying at the same time. There is a peculiar pressure that comes with someone trusting you to lead something end-to-end, especially on a fixed six-week clock.
Weeks 2-3. The plan met reality fast. I was advised to start with a simple pilot study to test out the analytical pipelines I would be using to assess the outputs. Generating the pilot's 1000 images across those four regions took longer than I'd budgeted. Cleaning the pipeline also meant real debugging. Staring at a script for forty minutes before realizing one filter condition was quietly letting the wrong data through, I found it more brutal than I expected it to be. I caught three major separate bugs like that over these two weeks. By the end of week three, I'd run the full set of statistical tests across all five frameworks, something I genuinely couldn't have explained the logic in week one. I came into this project with no major stats background, and there's a specific satisfaction in watching a concept that felt dense on Monday click by Thursday.
I'd be lying if I said imposter syndrome hasn't shown up alongside all of this. It is definitely hard not to feel it when you are building something that is supposed to hold up analytically, especially when you are on a tight schedule that leaves no room to stall. But three weeks in, and I notice it less than I did in week one. Some of it is my familiarity with my project. Some of it is realizing that the moments of getting stuck haven't actually stopped anything. They have made me recalibrate and assess my actions further.
Looking ahead. The second half scales up meaningfully: six regions, more models, and a set of intervention strategies I am genuinely curious to test rather than just document. There is a real chance some of what I built in these first three weeks won't survive contact with that larger scope unchanged, and I am trying to hold that lightly rather than defensively. For now, I am at the halfway point, with a working pipeline, a much better grip on the tools I am using, and enough momentum to be looking forward to the next three weeks.
Please sign in
If you are a registered user on Laidlaw Scholars Network, please sign in
Hi Jordan, this sounds absolutely fascinating. We were discussing at our recent Board meeting studies on LLM biases. I can't wait to read your final paper. I doubt that with just six weeks you will have time to test this too, but if you do discover any gender skews, I would love to see the data. Thanks and good luck with the final stretch!
Thank you so much! This means a lot! I'm concentrating on cultural bias specifically since it's one of the least studied areas compared to things like gender or race, but I'll try to look into gender skews too if the data allows, since it could also show whether findings from earlier studies still hold.