The hidden cost of AI is someone else’s time

Sign in<br>Subscribe

I have a love-hate relationship with Large Language Models. At first, I'm like a kid assisting a magic show. I remember how amazed I was when I saw the first tech demo for ChatGPT Omni and its cough definitely-not-Scarlet-Johansson's voice. The movie “Her” was getting real. Yet, besides some very specific use case like using LLMs as the new Google for everyday troubleshooting, I'm frustrated by AI more often than I'm delighted with it. This is true when I use it, but more importantly when it's thrown at my face, unsolicited.<br>The genie is out of the bottle and GenAI is here to stay. What we can influence is the etiquette around its usage, especially in close social circles and at work.<br>Below is my compendium of AI hot takes and rules of thumb I wish everyone I interact with would keep in mind.<br>If you feel the need to specify your answer is AI based, work it more<br>I see this becoming a trend in the corporate world, you send a request to a colleague, and they answer with a gigantic wall of text, word document, or tabular content. As a preamble, you have a one-liner mentioning this was done “with the help of AI”, as if this would absolve themselves of their responsibility if the content was bad. More often that not, they’ve barely read, let alone reviewed the produced content.<br>This breaks the trust relationship as I can’t tell if I’m reading your judgment or the model's. If I asked you in the first place, it was to get YOUR opinion or approval, based on YOUR expertise. I can prompt GenAI tools just fine on my own.<br>Feel free to leverage LLMs during your work, but you need to own your work fully. Guide the models with your expertise, proofread their answers, amend where needed until you reach a point where you’re proud enough of the result to put your name under it.<br>If you are incapable of assessing the quality of the model's output in a specific domain, do not rely on it<br>LLMs make for incredible semantic search engines and summary generators, but they can also be blatantly wrong while writing in a very confident tone. We call them hallucinations and even the best models available still do this regularly. This is a problem that might never get solved.<br>Hallucination is actually one of the main reasons I’m often deeply disappointed by LLM in real use case. I would ask for advice on how some construction work, only to realise later it recommended me to put on walls a product that is only designed to be used on floors, etc. It’s destabilising as most of the computer science world, its governing algorithms, used to be about deterministic behaviour. You would ask a program something, and it would always give you the exact same answer. LLMs however, are stochastic beasts. Their output can vary wildly, yet they mimic human behaviour and answer with confidence.<br>This is a dramatic difference with human-based interactions. When people have doubt, don’t recall exactly, or don’t know, they will usually mention it outright or refrain from giving any feedback. They won’t give you a made up answer only to say “Yes, you’re absolutely right!” when you confront them about their bullshit. People behaving like this would quickly lose credibility and get ignored or ostracized. Paradoxically we have an extremely high tolerance for this behaviour with LLMs.<br>If you use LLMs on areas you’re not comfortable with, be overly sceptic of everything it says, even when it confirms your natural bias. Cross-check whatever you can with other sources or at least double-check the LLM’s own source when it’s the result of a web-search and the website are referenced.<br>If you didn’t put effort in writing your piece, I won’t put effort in reviewing it<br>For the first time in history, producing content can be faster than consuming it.<br>LLM tend to be overly verbose by default. They will produce gigantic walls of text, neatly organised with titles, bullet points, bold highlights and emojis. This leads to a gradual inflation of the average corporate email, ticket, PR, or specification document. I’ve seen people use this as a weapon, bombing you with a 10 pages long AI generated pamphlet pushing against something you’re trying to push and asking you to argue back with the same level of details. Those pieces of content tend to be of very low entropy, abusing rhetorical constructions and repeating the same points over and over; two paragraphs’ worth of ideas diluted in 20 paragraphs of text soup. This creates a strong asymmetry in the relationship where the receiver has to spend a considerable amount of time reviewing what only took a fraction of this time to produce. Alternatively, they can give up, shove it to their own LLM of choice, and ask to generate a summary or directly an answer. At this point, you’re basically doing an GenAI Ouroboros, a total waste of time.<br>The only sane way to fight back is to plainly refuse to take part in such madness and ask for a more refined document, going straight...