QH WK 06: Untrack Me

So my keylogger has been talking hourly to IBM’s tone analyzer API for a couple weeks now, and I’ve been noticing strange responses to the chunks of time I’ve spent coding:

When I first noticed this behavior, I thought it amusing yet unforgivably flawed; I was trying to do some Serious Sentiment Analyses here, and clearly I was not experiencing such extreme mood swings while programming (albeit, the examples above illustrate my experience pretty well).

But in the context of this assignment, it provides an easy opportunity for manipulation— I just needed to figure out what exactly was triggering these false positives. So I gathered all the erratic predictions and fed them “word” by word into IBM’s demo, adding and deleting until I found precisely which characters the model was responding to. For the foregoing examples:

A few other absurd samples:

I also found that several individual words triggered extremely confident predictions:


So what can be used to consistently hack IBM’s predictions?

For starters, any built-in function (that’s also a complete word) will get picked up for displaying confidence:

Even when followed by text that’s obviously not-so-confident:

Even when in nonsensical function salad:


To sound analytical, one must simply add the word “if” anywhere in a sentence:

Regardless of whether or not the surrounding words are analytical:


To seem instantly and dramatically happier, just add {}:

Curly brackets are so joyful that they even neutralize fear and sadness:


Lastly, the singular most effective way to express anger is with this emoji: =P

Its rage is so complete that it literally sucks the joy out of life:

2 thoughts on “QH WK 06: Untrack Me

  1. Excellent exploration into this piece of technology. Quite scary and surprising to see the serious gaps in logic in the IBM Tone Analyzer Particularly upsetting and terrifying is how they seem to sell it as “revolutionizing call centers” etc. Really happy to see your interrogation into the ways in which the tech is failing.

    Thanks for your effort! I’m curious to see how this research be used to produce perhaps some more statements about the “state of the art” in sentiment analysis and what strategies we might adopt to mask our Trackers’ ability to read us.

Leave a Reply

Your email address will not be published. Required fields are marked *