“The alchemist knew the legend of Narcissus, a youth who knelt daily beside a lake to contemplate his own beauty. He was so fascinated by himself that, one morning, he fell into the lake and drowned.” — Paulo Coelho, The Alchemist

The Algorithmic Mirror of Narcissus (part1)

A Path Toward Automated Social Network Lie Detection Using Transfer Learning with Google’s T5 Unified Text-to-Text Transformer Model

Mike Casale
8 min readMar 26, 2020

Narcissus drowns while he contemplates the algorithmic reflection of his data. This is because in mathematical terms, a social network is less a connection with others than it is an algorithmic reflection of ourselves. These algorithmic mirrors reflect only the content we are most likely to see, to “Like”, and to “click” — and in this sense, the threat of Facebook is not so much a new terminator-like artificial intelligence, but rather it is the same old threat warned of in ancient myths like Narcissus.

Looking through the lens of a machine learning developer, the real problem with fake news is that it appears where there exists a high probability of it being seen.

This behavior of social network content feeds is analogous to electrons in the famous double-slit experiment of quantum mechanics — where even passive observation of the experiment can influence its observed outcome. In the context of modern social networks, fake news articles appear whenever fake content reflects what we are most likely to engage with and see. From this it follows that an unintended feedback loop exists in machine learning recommender algorithms, one which tends to reinforce our belief systems, and rarely, if ever, challenge them. Realizing this, it’s plausible to think that social network recommender algorithms have inadvertently exploded society’s common set of facts and truth by habitually exposing the public to content which IS NOT AT ALL concerned with fact or truth, and IS concerned with personal-bias i.e. my repulsions, my ambitions, my appetites, my passions, my insecurities, my values, my etc…

With the aim of overcoming this challenge, let’s explore an idea for automated lie detection via transfer learning of Google’s T5 Unified Text-to-Text Transformer model. This idea was originally proposed by Professor Robert Thibadeau — one of the founding Directors of Carnegie Mellon University’s Robotics Institute — in his book and ongoing project, the Internet Court of Lies.

Some Difficulties Involved in Lie detection



Mike Casale

machine learning developer