AI use in ACARS OSINT

 While this blog is not directly about ACARS, I'd still like to share some of what I've read and mull over.
Hope you don't mind or at least find some of the information / thought process a little provoking.

Preamble Part One

If you have been following my X account for more than a year, you may recall some backlash I received for attempting to use AI in ACARS message decoding.
I took the 'feedback' (angry and negative that it was) and kept working on it in private. (I do a LOT of ACARS work offline.)

I've been discussing using ML/AI/LLM in ACARS with Mike @ElbaSatGuy via DM quite a bit, and we hope to have more to share at some point (visit his X stream to see screenshots if you're interested).

Mike's effort and thought process about its use case with ACARS mirrors mine and my thoughts around AI. We both want to use it to filter and tame the firehose of ACARS messages. To aid in their decoding, filtering and display.

Preamble Part Two

I'm very upfront - well, I try to be - about the fact that my two social feeds (X and BlueSky) are NOT about OSINT. That said, I think a _lot_ of my followers are only following me because they see my posts and think I am going to post some juicy insight to WW3 or something.
I seemingly post about aircraft, some of which are military; therefore, I must know something. (I also think I still have a few hangers on from that one Taylor tweet I was tagged in - still not real happy about that).

Not that any / many of my followers have even visited my GitHub site or read the front page, but here is what I have stated there:

I wrote that many years ago, and it is as true today as it was back then.

Blog.

I try to keep this ground truth about what my role is in running the site in mind at all times.
I'm doing what I do to help others do OSINT. 

I have often wondered how many of my site users actually consider the origin of the data they are reviewing. How does it end up on the web? Does that unknown process alter the context of the data in any way?
When ChatGPT first emerged and took the world by storm, it was largely uninformed about ACARS.
I asked it a few questions, and it floundered around, making things up in its typical overly wordy, overly confident tone.
Subsequent versions (models) have improved in terms of how ACARS is utilized by airlines.
One could argue that they are even better, but they still lack the depth of understanding of how ACARS data gets from aircraft to website and how that can change the context of the data.
ACARS OSINT is a use case that is outside of the airline or the classic OOOI (Out, Off, On, In)

Enter then the human OSINT. And at the risk of making a bad joke, enter the avgeek OSINT.
The respect I have for the core power users of my site is huge. How in the world you guys keep track of the amount of registrations, hex codes, and flight history blows my mind.

This avgeek OSINT is so far beyond what most GPT / LLM models can do that it is hard for me to see them catching up any time soon.
You would need to have the full ACARS firehose from a site like mine or airframes.io feeding into a type of live learning, live feedback loop to help the model learn what its looking for in real time and how to track those changes over months of history and, here is the real tricky bit, what to look for, what an anomaly looks like and how to flag it for a human.

I suspect I will have more to say on this subject as I continue to receive feedback from site users and explore LLM in this use case.

With all that as background, it's a long read, but I found the following blog about AI use in OSINT very interesting..

https://www.dutchosintguy.com/post/the-slow-collapse-of-critical-thinking-in-osint-due-to-ai