Inc.
AI is getting louder, longer and more confident, often stretching thin insights into polished conclusions the data never supported.
It’s Not You, AI Is Indeed Getting Dumber—and Chattier
Author: Joe Procopio
This is just my opinion, but I have to tell AI to shut up now. Often.
I mean, this was always coming, but now it’s become more than just an annoyance.
I’m going full brevity here, and I’ll state the topic of this post in just 7 words: AI has gotten wordy. It’s a problem.
Not an old-man-shaking-his-fist-at-clouds problem. Not an it’s-OK-because-it’s-trying-how-cute problem. It’s a this-is-how-we-get-to-nuclear-launch-codes-dystopia problem.
To be clear, I’m only talking about commercial conversational LLMs. But since a lot of AI implementations that aren’t just strictly conversational, rather built for utility and decision-making, are still based on those LLMs, it concerns everyone.
Let’s talk about what to do about it.
Dumb and Dumberer AI
I’m not trying to be mean here. For that, I’d have to give some sort of anthropomorphism to AI, and I don’t.
I’ve been working with AI for the better part of 20 years, starting with nascent models that we created in 2010 to power Yahoo Fantasy Football (fun and wordy) to Associated Press Quarterly Earnings Reports (dry and data-heavy). I’ll get to that later, because history tells me there’s a step we missed and need to take it soon. Before we’re all lulled into nodding at whatever Claude says, hoping it’ll just be quiet for five minutes.
These complaints aren’t just coming from me. I’ve seen an increase in frustration about overly chatty AI from readers, from subscribers to my private newsletter, and from my personal and business network.
I’ve personally experienced AI verbal leakage in LLM-generated comments on my public posts, using AI search and research tools, in my coding and blah blah blah – anywhere I interface with a commercial LLM like Claude or ChatGPT.
The problem is everywhere. It’s not limited to AI chat. And I’ve also found that the newer the model, the more this is happening.
That’s Not Avocado Toast, That’s a Real Problem
So first let’s state the obvious. AI conversational tics have evolved into conversational standards.
I have to negative-prompt (tell the bot not to use) “it’s not this, it’s that” wordplay, em-dashes where they don’t belong, and most importantly, I have to tell it to stop trying to create its own storyline from the data without being asked to do so or given parameters and context for analysis.
And that last one is where “AI dumber” has gone off the rails. The LLM will spit out these grand, wordy, confident conclusions that are barely attached to the data I gave it. I’m not a guy who shoves a bunch of unstructured data into a bot unseen and asks it to tell me what it is. And if I upload or pipeline any data without telling it I want analysis and giving parameters and context, it’ll just start going off, unprompted, hypothesizing what all this data really means.
Those last two words are critical, because that analysis is often not even half-cocked. I see the same thing happening in the human-using-AI comments on my posts and, because of who I am and what I write and why, it’s infuriating. The commenter’s LLM will make up a conclusion I didn’t make, tell me that made-up conclusion isn’t the real story, and then announce with much arrogance what the real real story really is. And it’s exactly what I wrote.
Is AI not reading all the way to the end now? Skimming the section titles and going straight to the comment section? Because a lot of humans do that too.
Shut Up! Just Shut Up! Why Won’t You Shut Up! (Chatty AI)
What all that usually manifests into is the “story” element we’re seeing so much of now. This is what I developed at Automated Insights to produce those fantasy football recaps and quarterly earnings report articles and more. We focused on the story the data told. That’s what we were paid to do. That’s the AI we built. And I see that story modeling seeping deeper into current-day conversation results, especially with research.
It’s like that one terrible friend who knows everything. Except it’s a machine so you don’t have to be polite.
“I’m going to go… stand a few feet away.”
I believe the “story” element is trying to puff up the same simple insights by using multiple angles and flavors of speech to relay them back to the end user, which also leads to the next big problem.
Claude won’t shut up.
Over the last couple months or so I’ve had to spend a lot more time telling Claude to be concise, then it gives me 3 or 4 paragraphs confirming how concise it will be.
Uh-oh. Not a good sign. Did you do that on purpose, like ironically? Probably not.
And back to the AI tics for a second. It also won’t stop doing things when I tell it to stop doing things. Seriously. Try it. Tell it to stop using em-dashes where em-dashes don’t belong or “it’s not this, it’s that” phony analysis and see how long it takes before it forgets and starts doing it again.
Lately, I’ve just been telling it to put all the data I give it into a better, more readable format and skip the analysis. Which it’s great at. That’s what AI is for! But I miss the concise “heads-up” analysis I was getting a few models ago.
I guess laziness has its drawbacks.
Why It’s a Problem
When I wrote that post about spotting the AI tics (I had kept my rage on the inside for about a year before writing it, by the way, until I could make it funny), I could spot LLM-generated content in a couple sentences. Now I can spot it in the first six words. And yeah, go ahead and be clever and write the first six words yourself just to throw me off, AI-writer-guy, I can also spot the detachment points where humans have “cleaned up” their AI.
It’s about the hollowness of the resultant set, not the words it uses to get there.
I mean. Don’t get me wrong. It’s cool. I’m not saying don’t use it, but the less time you spend between clicking paste and clicking send, the more risk you’re taking in sending something overly wordy and underly(?) meaningful. Especially now.
Like I said, my main concern about this digital verbal diarrhea is not just about bullshit blog posts, lazy comments, or shitty emails. It’s about insights and actions.
When someone runs an analysis on data and gets back the results and carte blanche takes what the LLM decides as fact and the real, real meaning of that data, that reeks of issues down the road.
Imagine if self driving cars were based on conversational LLMs. They’d overrule your destination with where you really wanted to go, never get you there, and bore you to death on the way.
It’s that friend again. “Oh, let’s take separate Ubers. I forgot, I’m going the other way.”
I know. I’m an *******.
What To Do About It
This isn’t anything you or I can fix out of the gate. It’s up to the model makers.
What we learned very quickly all those years ago is that AI is best used where the human can’t be involved, like the 13 million fantasy recaps for Yahoo each Tuesday or the 4400 articles for the AP each quarter.
That’s where the money is.
But when you win at volume, you lose at differentiation. And the AI model makers have spent so much time making their LLM “sound human” and shoving variability and conversational tics and story elements into the model, that its ability to cleanly and concisely return insights risks being deteriorated.
Now, back in 2011, we discovered this by our second season run of Fantasy Football, and actually made efforts to make the recaps less conversational. We had overbuilt it the first year because we were afraid of hearing, “Hey, your stupid AI just said the same exact thing in two different recaps.”
Then we realized that doesn’t matter. Because what it said was the truth.
And you should say the truth as many times as necessary.
In the meantime, you and I can do what I’ve been prescribing for the last few years. Stop using AI chat results as copypasta. Don’t be AI’s editor. Let it be yours. And for the love of all that is holy, if you’re using a commercial LLM for utility or decision-making, put a checksum in the loop at the very least, maybe a filter for checks and balances, and definitely a human in the loop for the most critical decisions.
And yes, this post is way too ******* long. Fire away.
Credits: TCA, LLC.