Recently, I used LLMs to draft an intermediate report on my progress. Alongside some details on the current limitations, I also provided comments from the revie

Recently, I used LLMs to draft an intermediate report on my progress. Alongside some details on the current limitations, I also provided comments from the review team (e.g. “it’s fine to be incomplete”, “I expect certain sections” etc) to provide context. Surprisingly LLMs, all of them, suggested me to lie about the limitation and re-frame them as deliberate features, even though I emphasized that the limitations are totally okay. I suspect that mentioning some high profile people in the review team pressured LLMs to derail in the hopes of saving me. While I’ve never been a big fan of LLMs, I definitely lost a big chunk of my remaining faith in this technology.