My first "real world" AI use

I’m late jumping on the AI bandwagon. I wouldn’t even say I’m on the bandwagon. Maybe I’m just sitting on the side and watching it and wondering if it is merely a passing curiosity. Every now and then I walk over and take a look to see what’s going on. When Stable Diffusion 1.4 was released that’s what I did. I immediately signed up to hugging face to download the model to try it out with my kids. The results were… mixed. Image models have come a long way in 3 years though and so have language models and I’ve finally found a practical use: objective reviewer.

I’m a stay at home parent at the moment but I do some volunteer work to give back to my community. As part of that work, there are sometimes disagreements and or even tension during discussions. One reason for tension is that a lot of discussion happens over email so we lose the nuance that facial expressions and tone of voice provide. One advantage of email is that it’s a written record and therefore easy to reread and think about.

Problem

In one particular instance last spring, one member of a committee had written a letter and I did not feel it was ready to send to the hundreds of people on our mailing list. I didn’t mind the idea of the letter and many of the ideas in the letter were okay but the tone, especially in the opening lines, was defensive and not something I could put my name on. Perhaps I have a high bar for business writing. As a general principle though. the larger the audience then the higher the quality bar we should have.

As a software engineer, I come from a really collaborative environment where I review code and have others review my code daily. I did the same thing here. For this letter, I wrote a handful of edits with some rationale and expected to collaborate on a second draft. It didn’t work that way. Most other committee members were giving the okay to send it. It felt lonely. Did no one else see what I saw?

The author responded to me privately by dismissing my comments as stylistic and sent a second draft, almost the same as the first, to the committee for approval. I’ll be honest, it felt a little disrespectful.

This situation left me with a few problems actually:

  1. What to do when someone disrespects your time and work?
  2. How can I get my name off this letter?
  3. Am I crazy? Are my edits and suggestions reasonable?

People Problems

The first two problems are classic people problems and happen all the time in volunteer committees and the workplace. I’ve said before I’m not a confident person but I do recognize that it’s important to speak up when something isn’t right. The author said they had addressed my concerns and that was misleading, intentional or not, so I felt a need to clarify:

To be clear, we did not reach an agreement. I would say that [the author]‘s private reply did not address my concerns but was dismissive of my concerns. I’d like to remind everyone that constructive feedback is a gift and, if we are going to succeed then, we need to respect, listen to, and value different viewpoints and life experience.

That may look harsh but the author’s private reply felt dismissive and even attacked me for making comments. Something needed to be said. I didn’t mention the attacks but rather I wanted to focus on the kind of culture and environment we should have. One that respects, listen to, and value others.

For the second problem, I needed to get someone else on my side. I was afraid the letter would be sent if my voice was alone saying “don’t send it”. I reached out to chat with other people on the committee face-to-face and managed to talk to one. We chatted about various things but they agreed it’s probably better not to send the letter. This felt like a win but there was still that third problem.

Personal Problem

The third problem, the “Am I crazy? Are my edits and suggestions reasonable?” problem, is actually all in my head. No one else knows I’m thinking this but how do I fix it? I wish I had someone objective to read through this. I talked it over with my wife and, while she agreed with me, I’m not sure she was completely objective. She loves me and supports me and could probably see my stress. Then I thought about an LLM. They would be objective, right?

It was easy to paste the letter into an LLM and ask for feedback:

%> What is the tone of this letter: [contents of letter]

The tone of the letter is measured and mildly defensive, aiming to [redacted]. The authors present themselves as transparent, responsible, and community-focused, but the writing carries a subtle undercurrent of frustration at having to address what they see as unfounded concerns.

The letter is a generally clear, well-meaning attempt to explain [redacted]. While it’s informative and transparent, the tone carries slight defensiveness and the writing could benefit from editing for clarity and conciseness.

That was a breath of fresh air. I pasted the email exchanges with my suggestions and the response I got from the author.

%> Was the author’s response dismissive?

Yes, [the author]‘s email was dismissive and displayed a patronizing tone towards Lloyd. While it started with an attempt at collegiality, the rest of the message largely brushed aside Lloyd’s detailed feedback as unnecessary nit-picking.

%> Are Lloyd’s edits reasonable?

Lloyd’s edits are reasonable and well-justified, primarily because they focus on principles that are widely accepted as best practices in professional and public communication.

Lloyd’s edits aren’t just about style; they are about substance and strategy. They aim to make the letter more effective, trustworthy, and less vulnerable to criticism, aligning it with best practices for public communication.

Vindication. An objective reviewer who has, presumably, been trained on good writing says my edits are reasonable and the author is not acting in good faith toward me.

Moving On

The situation was months ago. The author had some personal issues at the time so I can forgive them not being completely professional but there hasn’t been any follow up, either from the committee or the individual, on what happened. Maybe people want to move on and hope it doesn’t happen again? I’m doubtful but I don’t need justice. The LLM’s feedback was enough for me.

I’ve actually used an LLM several times since then to get help on how to respond to other discussions. I’ve also used it to pick up subtle nuance in email discussions and my own emails. I recall one discussion where I was surprised by a negative response to one of my emails. I pasted in the email exchange and asked some questions:

%> How would [person] feel about Lloyd’s reply:

It’s tricky to definitively say… but we can analyze the reply and predict potential reactions:

Potentially Positive/Neutral Reactions:

  • Appreciation for Research: […]
  • Logical Arguments: […]
  • Agreement on Community Need: […]
  • Recognition of Complexity: […]

Potentially Negative/Frustrated Reactions:

  • Feeling Dismissed/Pressured: […]
  • Misunderstanding the Core Concern: […]
  • Lack of Collaborative Spirit (Potentially): […]
  • Information Overload/Too Much Detail: […]

The LLM picked out a few key phrases of what I wrote and anticipated a bunch of positive and negative responses. I ended up getting a negative one which surprised me but after rereading my email with this additional information I recognize it was my own fault.

Caution

The AI isn’t always right so I’ve developed a few best practices to get a better outcome:

  1. Add in context. I’ve seen the AI focus on the wrong part of the text and take the discussion in the wrong direction. As a human, you have more context so share that context and see if the AI corrects itself.
  2. Try to be neutral. For example, a prompt like “Is this dismissive?” will hint at what you want while a prompt “What is the tone?” should give more objective feedback. Or try both and compare!
  3. Don’t take the first answer. Start a new session with the same or similar information and compare. You may get a different outcome or get new insights. Try a different model. Gemini, ChatGPT, Copilot, Claude, Perplexity and so on. These LLMs are all easy to access and may give different responses to the same input which is related to the next point…
  4. Use your head. You decide to take or leave the AI feedback. You are responsible for the emails you send. An LLM is a tool and can be helpful but it can also mislead:
Play