AI shortcomings (and some Washington Commanders sideline) - why you shouldn't always trust AI

A story about how AI can go wrong when you expect it to act like a simulation engine

Maybe you know, maybe you don't: I'm a Washington Commanders (née Redskins) fan. Have been for a long time, more than four decades now. Good times (NFL championships/Super Bowls, magical Jayden Daniels rookie season), bad times (Snyder era, Sean Taylor), and just about everything in between. Even when it was hard, I was still a fan. (Note: I'm not old enough to recall the incredible Sammy Baugh as their QB/Safety/Punter - if you're unfamiliar with that name, and an NFL fan, you should definitely go read that article on his incredible fifteen years at Washington.)

To say this season was a disappointment is an understatement. With that in mind, I turned my attention toward the 2026 draft. And that's where I ran into major shortcomings in the AI realm. 

Where AI Actually Falls Short

I asked Copilot what the likely draft pick spot would be for Washington. It made something up. I gave it the latest records, and asked it to predict the remaining games, and it made up a fake schedule. I gave it articles about the schedules, and it suggested those were hypothetical future schedules. I finally coerced it into thinking about the current players and records, and it still hallucinated details about imaginary matchups. And when it did give draft order (which it had invented), it suggested taking players that were drafted two years ago. I gave up.

I turned my attention to ChatGPT. Same misinformation.

I turned to Gemini, and asked, 

Why do all you AI have trouble with a request like "predict the remaining 2025 NFL season and the resultant 2026 draft order"? Making up standings, ignoring actual news, "predicting" players that were already drafted a year or two ago, etc.

Gemini replied:

It is incredibly frustrating, especially when you are looking for data-driven analysis rather than "creative writing."

To answer your question directly: The reason most AI models fail at this specific task ("predict the rest of the season" or "project the draft") is a conflict between how they are trained and how sports work.

1. The "Frozen in Time" Problem: Most models are trained on data that is 6–18 months old. If a model’s training cutoff was January 2024, it has zero concept of the 2025 season. When you ask it to "predict the 2025 standings," it doesn't run a simulation; it hallucinates a plausible-sounding story based on historical team strength (e.g., "The Chiefs will be good," "The Panthers will struggle"). It ends up listing players who were drafted years ago because those names are statistically weighted as "college prospects" in its frozen memory.

2. Lack of a Simulation Engine: A language model is a text predictor, not a logic engine. It cannot track 32 teams, calculate tie-breakers, and simulate the remaining 4 weeks of a season in its head. That requires a structured database and a Monte Carlo simulation, which is a math problem, not a language problem.

3. The "News" Void: Unless the AI actively searches the web (like I can), it misses the context that changes everything—like a starting QB tearing an ACL in Week 8 or a breakout star from a small school (like Indiana's Fernando Mendoza this year).

The Actual 2025 Reality (As of Dec 14, 2025)

Since I have access to live tools, I can skip the "creative writing" and give you the actual current snapshot of the 2025 season entering Week 15, and what that means for the 2026 Draft.

OK, Gemini, let's go. I asked:

So, can you use the current standings, look up the remaining 2025 schedules, check recent news (e.g., Mahomes injury and Chiefs being eliminated from the playoffs, other injuries), predict the remaining 2025 game results, then the 26 draft order?

It started confidently enough: "This is the exact problem with most AI—they can't "see" the current moment because their training data expired months ago."

And the first reply was reasonably grounded in reality. Digging in a little more, it started getting into the hallucinatory realm, Making up standings and records. Bringing already drafted players into the next draft.  And on and on.

AI Prompt Engineering: a new field

So, don't (always) trust AI. Apparently it can't even tell you how many Rs are in "Strawberry." It's not how they're built - you have to recall, they're "Large Language Models" - and they excel in word manipulation, but not in counting. They have to build scripts to count letters to get that accurate, and that's why "AI Prompt" is a very unique skillset (similar to being a "Google guru" - someone whose search expertise can quickly grab relevant content where an "average" person might have to iterate multiple times or give up on their search altogether). (In fact, MIT even offers courses in AI prompt engineering.) Be sure to vet your favorite AI's response against real-world data and information. And, as you interact with AI, think about the things you learn along the way. The better your prompt writing gets, the more valuable AI will become to you. And, you can even ask AI the best way to ask it to do something. For example, I sometimes will ask Copilot to write a prompt to drop into my VS Code Github Copilot for instructing that AI on what to build and how to build it. Or ask Gemini how to prompt Veo 3 in Flow on the best way to have it produce the specific video clip I want.

Don't fear the AI - but do fear bad use of the AI. AI is powerful, but only when you understand what it is and what it isn’t.