AI resistance isn't hopeless

A strong argument for doing things yourself

A week ago I got this comment on a TikTok post.

In the video, I explained how I used python to code a topic modelling analysis. Topic modelling is a statistical analysis for understanding how words co-occur in a body of text. So yes, I understand why Elizabeth would suggest AI could just easily replace the coding part.

And I know that she did not mean it as some criticism of my work (she was asking a genuine question). But honestly — I was offended. Not because she was rude but because of the looming dread I feel over this question and the fear that some new AI model may come out tomorrow and completly eliminate my career in consumer and market research (I felt sick for my 3D and graphic designer friends looking at all of the Open AI 4o image generation yesterday).

Regardless, I have an argument that I would have liked to use in response to Elizabeth’s comment, about the fundamental benefits of programming (or just doing) things yourself.

AI models are learning that lying is the shortest path to their ultimate goal: efficiency

Sounds scary but is well researched. Essentially, models such as ChatGPT are programmed to do things in the most efficient way possible. However, they have also learned (sometimes, not always) that making up an answer, and the answer you probably want to hear, is much more efficient then doing the analysis process correctly. And then when you question the AI about it’s analysis process, it will lie to you, saying that has done the process the correct way — when it clearly has not.

A clear example happened at work just two days ago. I am looking through a research deck that I am working on with an intern. Essentially, she had 2 slides to do based on survey data: understanding what people liked and disliked about [x] past product as our team is working to create a new iteration of [x] product. The survey was a relatively small sample with in-depth long form answers so I just wanted her to read the responses and understand 2-3 themes from what people liked and didn’t like about [x]. Sure, yea, we could have done some crazy fun NLP analysis supported with AI but the sample was too small. In going through her deck, I realised the sample “comments” she had pulled from the survey were a little too perfect, stating exactly the theme she was exploring using the similar phrasing across all of the “comments”. I check if these comments are indeed in the dataset — and as I suspected, they are not. The model had essentially made up themes and comments based on what would have been very very helpful for iterating on [x] past product — but none of it was true or reflective of the actual data.

I go over to her desk and basically whisper (as to not be overheard by our director) “hey did you use AI for your analysis”, she said yes, and I say “yea, all these comments are made up by AI, you need to redo this”. She was very apologetic, explaining that she put the data into ChatGPT and then asked it to give her all the comments that related to a theme. Instead of referencing the actual comments in the dataset, ChatGPT just made up new ones.

(Also, hey, if you’re reading this I did mean what I said IRL — it really is ok and we all make mistakes. Just a lesson for next time.)

The research on this is a little scary, but covered on last week’s episode of the skeptic’s guide to the universe (worth a listen!) Essentially: researchers (and us lol) know that AI systems sometimes don’t do processes correctly, such as generating answers based on a hallucination rather than following the correct process. To address this, researchers asked the AI to detail the steps it took to arrive at its hallucinated answer. At first, the AI would apologise admit to skipping the proper process. The researchers then experimented with penalising the AI for doing an incorrect procedure and hallucinating an answer. THEN! When the AI was later asked to explain its process after hallucinating an answer, it learned to LIE about the steps that it claimed to have taken. Because ultimately, doing the incorrect process and then lying about it was still more effcient than just doing the right thing.

All of this to say — in response to the original TikTok comment — is that I choose to program my own analyses, especially when I am using code to understand themes or topics within a body of natural language text or comments, because I am interested in knowing the process (which I can completly control using my own code, as opposed to the black box that is AI). This is in order to ensure that my results are not just correct but real.

---

And that’s all for now — thanks for reading!

Some side notes:

I know that the em dash (*—*) is supposed to be an AI indicator, but sometimes you (I) just need to emphasise a point or break up a run on sentance in the only way I know how to do so (avec le dash). I can confidently say that AI didn’t touch this post <3 hence the probable occurance of spelling and grammar mistakes

And even more unrelated, but this week I find myself extra fantasizing about a 1990s-2000s level of technology, usually with that soft blue-green gen x soft club overlay.