AI and the National Security State

What does NSA’s LLM know?

Long before Edward Snowden’s leaks about the NSA’s surveillance activities, Bill Binney had already revealed to the world that NSA had blanket surveillance capabilities.  Essentially the entire internet backbone is mirrored, collected, stored in their Salt Lake City facility, and indexed. This is not merely metadata (who communicated with whom and when) but also the content—not only text and email, but voice data is recorded and transcribed.

We also learned that the FBI has a back door into this system and apparently maintains their own illegal copy of it, circumventing the legal requirement for a warrant, and inviting its abuse for political purposes. Before his inauguration, Trump was made aware of this espionage used against him by Admiral Mike Rogers, former NSA director. 

We all know that big tech, Silicon Valley, telecoms, and freemail services like gmail “have it all.” Everyone has taken for granted voice dictation on their smartphone and voicemail transcription, these are rather poor in comparison to AI-based services like Trint, but we know this works.

In addition to this, Social Media has your social graph and a model of your behaviors, opinions and activities. They have your private communications conducted over their platforms. They really do have it all, and this includes communications of ‘bad actors’ in the political echelons and in the private sector.

Worse than that, they have your purchases and credit card information. Telecoms and banks are constantly guilty of this too. Much of this data is breached and leaked, so the notion that Large Language Models have not been exposed to this information is very naive. Much of the difficulty in training AI like ChatGPT has to involve making sure personal data does not leak out in its answers, either by pre-training or post-training methods. 

Not considering myself much of a threat to the Deep State, I personally did not worry about being a target for domestic surveillance abuse. It would take a lot to get myself on a watch list for which a secret FISA warrant would be obtained. This is not necessarily the case for whistleblowers and investigative journalists. This capability has been used against such people and it has a chilling effect on all of society.

Now, are we to suppose that the NSA does not also have its own LLMs inside its firewall? Let’s not forget that the DoD, NSA included, have technology that is light years (usually by decades) ahead of what is commercially available or even speculated to be possible. I had witnessed this fact with my own eyes in the late 80s when I was in their employ. 

Think about what that means if it is the case. Just as with search engines, in NSA’s databases the problem was always how to query it—how to ask for a specific unknown needle in the haystack. We’re all familiar with the classic poor results obtained with simple indexing of keywords used by search engines: often there is not sufficient semantic interpretation to separate out what you are really looking for from all the irrelevant garbage.

This all changed with AI services like ChatGPT and Perplexity.AI. Now you can immediately get right to what you’re after thanks to intelligent semantic parsing of the question and available data linked to it in the language model. 

If NSA or some other intelligence agency had a LLM that was trained on ALL THE DATA, it becomes a simple matter to ask for very specific communications about specific topics. One would not even need to know the names, email addresses, or phone numbers of the persons or entities involved—those would be discovered in the query results. Private corporations, government employees, politicians, non-government organizations, international criminal organizations, and think tanks would be laid bare in the digital dragnet. Hidden “deep” organizational structures would be graphed just the same way that social media has a social graph of all its users. 

Imagine the power! Should we be afraid of it? It depends on who has it and whether they are using it for good, legitimate reasons. For better or for worse, such power certainly exists.

That’s why I’m not particularly concerned that General Paul Nakasone, also a former director of the NSA, has recently been embedded into OpenAI’s board of directors.  Of course it had to be done. Of course it’s a matter of National Security. Of course it is essential that ChatGPT is not accidentally or intentionally misused or that the super-power it represents does not fall into the wrong hands. It is especially crucial that Big Tech does not have a monopoly over this power.

What would this capability mean to the Deep State? It means they are unmasked and vulnerable and that someone “has it all” and they can be prosecuted with air-tight indictments.

The cat’s out of the bag. Not only is the great equalizer unleashed into the wild, but now we know what this technology portends for organized political and crony corruption. They are certainly panicking.