When Machines Get Ahead Of Us

By Roger Kay
Endpoint Analysis

This pattern may not be interesting, but correlated with other information, it may take on much greater significance (Alasdair Allan and Pete Warden, New York Times)

Two recent articles in the New York Times highlight two different aspects of a single phenomenon: we build powerful machines, their capabilities lead to unintended consequences, and they ensnare us in some unforeseen way.

The first discusses a new technology, an improvement, really, of a technology that’s been under development for a long time: generating meaningful news stories from pure data.

The second concerns the case heading to the Supreme Court pertaining to the matter of state surveillance of individuals using Global Positioning System (GPS) technology and how that relates to the 4^th Amendment, which in theory enshrines personal privacy against unreasonable search and seizure.

In the first case, computers’ increased capabilities are in the process of taking over a sophisticated task formerly performed by people.  In the second, computers are getting ahead of existing laws and social covenants.

In the GPS surveillance situation, we have a number of elements that make the story particularly juicy.  The case on which it is based features an unattractive defendant, a drug dealer, who was tailed by a GPS the police attached to his car without a warrant.  That such a poor specimen has become a cause célèbre for civil rights is one of the quaint and charming features of our judicial system.

What is more interesting, however, is not the disposition of this particular case, which could, on the face of it, decide either way once it reaches the Supremes.  What matters here is that warrantless surveillance of GPS data, if sanctioned, could streamline any given authority’s ability to set up a tap of this sort.  The technology itself scales brilliantly, and technology enthusiasts and investors love scaling.  It’s what leads to really big outcomes, like Google, Facebook, and Twitter.

The issue with quick-and-easy GPS surveillance is that it could easily scale to include a fishing expedition on every citizen in the United States.

Don’t think so?  Let me take you through it.

You may remember the disclosures last April that embarrassed first Apple and then a slow-to-acknowledge Google about how their phones (iOS and Android, respectively) captured, stored, and even transmitted location data.  The original New York Times article featured a very convincing map showing an iPhone traveling along Interstate 95.

Okay, so Apple said it fixed the problem after grousing that it wasn’t really an issue.  But here’s the thing: at any moment, that same data could be collected again with a simple software update, about which phone-wielding citizens might not be totally aware.

So, what would the rest of the mechanism look like?

Many years ago, I was involved with the National Security Agency’s (NSA’s) Project Etymon, a computational linguistics effort designed to help find the needle in the intelligence haystack.  None of the people in our little software firm had a security clearance, and it’s quite possible that some of our programmers might not have been able to get one.  So, Fort Meade threw its requirements over the wall, and we threw our software back.

But the working mechanism of the project could readily be deduced from things we did know.  For example, when our contact asked for script support for Oriya, Telugu, Burmese, and Amharic, we knew that only one was of real interest.  And which one could be inferred pretty quickly by reading the front page of the New York Times, which in the early 1990s was carrying on about President Bill Clinton’s debacle in Somalia, where they happen to share a script set (written language) with the Ethiopians called Amharic (the only native African script, as it happens).

Now, our software did something fairly pedestrian long since taken over by Unicode, which created the standard index for all the world’s written languages and then some (e.g., smiley faces and other random digital art).  It mapped script elements (characters) to code points (index numbers).  In itself, this capability was pretty boring, and we had aspirations to do higher-level linguistic analysis, but none of that came to fruition.

We also knew what some of the other subcontractors were doing because we had to make our stuff work with their stuff.  One particularly interesting partner was Lincoln Labs, which had created a speech-to-text converter that took audio voice and spit out International Phonetic Alphabet (IPA, not the ale, the script), a code set that enumerates every meaningful sound a human being can make when communicating with another human being.  All language sounds from Bushman clicks to the Russian whispering щ (shch) or buzzing ж (zh) have an IPA number and a font element.

We worked on the converter that could match IPA to other script sets.  How could this be interesting?  Let’s say a machine captured a recording of someone speaking as an analog sound file, and another machine took that and generated a digital IPA file.  Our stuff could take the IPA file and match script sets against it until a meaningful language popped out.

What could one do with that?  Well, TRW was another subcontractor, and it had created a hardware hammering engine that could compare text strings against each other very fast.  Of course, such engines are much faster now, but TRW had state of the art at the time.  So, let’s say we’ve got our strings in native text now in whatever language, and we start looking for words like “bomb,” “missile,” “North Korea,” “trigger,” “shipment,” “Iran,” you get the idea, say, within 15 words of each other.

The concept was simple even if the implementation was not.  There were fine points (e.g., How would the word Gorbachev look when spoken in Arabic?).

So, now we have a way of snagging audio streams and making them into files, converting those files into the relevant text, and checking that text for interesting words.

Where could the NSA get all those audio files?  Well, it had another project call Echelon, which basically captured all communications traffic, clear and otherwise, going over the air (e.g., microwave, satellite) in Europe.  Now, that’s a heck of a lot of people to be listening to all at the same time, and the one thing the NSA let slip during a meeting that they probably shouldn’t have was that most of this information went “into the ground unexamined.”   In other words, all they were doing was poking around at the haystack in a cursory fashion and moving on to the next haystack.

That was the problem we were trying to solve.

At the other end of the NSA’s process were language specialists of an exceedingly rare breed.  They had to know tongues like Yoruba (a West African language) or Tamil (a South Asian) down to the slang and metaphor level AND they had to be able to get a security clearance.  The intersection of those two sets is tiny.  Essentially, children of American evangelists who grew up natively in those locations while their fathers tried to convert the local residents to Christianity.

We called these guys the endomorphs (pear-shaped torsos) in contrast to the mesomorphs (the hard guys who place the listening devices in the field).

One of the NSA’s operating principles is to “hit ‘em where they ain’t,” to quote Wee Willie Keeler.  So, rather than going after heavily encrypted messages in the middle of transmission, for example, the spooks would prefer to simply read the clear text off someone’s screen through his window with a high-power telescope.  In the same way, if you don’t know what they can do, you don’t protect against it.  Thus, back then, there was lots of clear voice traffic to analyze.  Different techniques are required now, what with all the digitization, packetization, and encryption.

So, now we have a dragnet that can capture, if only momentarily, a huge quantity of available conversation in any language and have a machine look at it quickly for a minimum score of “interestingness.”  If it reaches that threshold, we flag it and print it out for the endomorphs to take a closer look.  If an endomorph marks it as truly interesting, it can be passed up channels for further analysis.

The second news story mentioned way back above, which describes a software program’s ability to write news stories from data, is rather amusing (if you’re not a journalist) in that this technology threatens the jobs of some journalists who thought that a machine couldn’t replace them.  More importantly, though, it demonstrates how far computational linguistic analysis has come since Project Etymon.  This stuff is rocket science compared to what we could do 20 years ago.

Because the work we did is ancient history, and what we learned wasn’t classified in the first place, I don’t feel as if I’m doing the NSA any particular harm.  However, Projects Etymon and Echelon illustrate what could be done two decades ago in terms of mass surveillance.  Obviously, hardware is more powerful, software more sophisticated, and communications more evolved today.

The bad guys have also improved their techniques.  So, the cat-and-mouse game continues.

But we should be aware that it is feasible — even easy, these days, what with wide data capture, huge data storage (think Amazon cloud services), and powerful data mining — for someone to keep an eye on us all.  Machines look for interesting patterns, and people take a closer look at what the machines find.

Did we really do this to ourselves?

© 2011 Endpoint Technologies Associates, Inc.  All rights reserved.

Twitter: RogerKay

Source-1, Source-2 and Source-3

---

Roger Kay. I founded Endpoint Technologies Associates, Inc., an independent technology market intelligence company, in 2005. Previously, I was vice president of Client Computing at IDC, covering client PCs (desktop and mobile computers). Before that, I ran my own research and analysis firm, directed operations for a developer of multilingual text processing software, ran a technology analysis and publishing practice for a consulting company, managed international accounts for a data communications equipment manufacturer, and did new product development for a computerized trading network. I have published in a variety of forums and been quoted in a number of publications and other media outlets. I snagged a B.F.A. from Bennington College and an MBA from the University of Chicago Graduate School of Business. I am multilingual, world-traveled, and have bicycled over the Alps, but am now a family man.