Blog - Jürgen Kanz

Logo Jürgen Kanz
Go to content

Tell me who wrote this text?

Jürgen Kanz
Published by in Neural Networks · 21 November 2019
Tags: AIArtificialIntelligenceNLPNaturalLanguageProcessingGPT2FakeDeepLearningNeuralNetwork
Please read the following text and give an answer to the question at the end.

"... Jürgen Kanz is the man in charge of the research project, a physicist at the University of Wuppertal, where he is director of the Institute of Nuclear and Particle Physics.

The idea came from Kanz's own research, he said, as he studied the history of the world's most famous meteorite and how it got to Europe.

Scientists believe the meteorite struck the Earth approximately 60,000 years ago, breaking into two pieces.

While there's not much certainty about what happened after the two pieces crashed together, the study does offer a possibility.

The pieces of meteorite, which is known as the Aldebaran, were originally found in Morocco in the 1990s.

Then, in 2003, researchers from the Institute for Geosciences of Wuppertal and the Institute for Nuclear and Particle Physics of the University of Wuppertal found the fragments while they were studying meteorites from the Pleistocene Epoch, roughly 5.3 to 2.5 million years ago. (The scientists later published their findings in a pair of papers.)

"They found a large piece of the meteorite that had a black diamond-shaped patch on the outer part of its surface, similar to the patches in the moon and the sun," said lead author Dr. Klaus Bischoff, from the Institute for Nuclear and Particle Physics of the University of Wuppertal.

The researchers believe that, during the meteorite's meteorite-making life, it formed a black diamond, which then formed on the inner surface of the meteorite, where it would later be exposed to the atmosphere. ..."

What do you think? Make your choice.



Are you surprised?

A neural network generated the text. The only input I gave was "Jürgen Kanz is" the rest is the outcome of an algorithm. Apart from my name, nothing is TRUE. It is FAKE.

The model behind, a neural network named GPT-2: 1.5B, is developed by OpenAI.com [1]. No question the company has done an excellent job, because they have trained this large-scale unsupervised language model (GPT-2) which generates coherent paragraphs of text.

You can work with the model by yourself without special programming skills. Just visit the website "Talk to Transformer" [2] and enter view English words or a sentence as a starting point. The algorithm will do the rest.

That makes it easy to generate text - fake text. There is a fair chance that other algorithms will follow and every user can generate and publish such a text. How can we protect ourselves against AI generated fake text?

Fortunately, there are already smart people, which developed some tools for common browsers to detect GPT text.

Firefox Browser add-on: GPTtrue or false [3]

Google Chrome add-on: GPTtrue or false [4]


Let us think for a moment
We have a neural network able to generate coherent paragraphs of text. Furthermore, we have first tools able to detect such generated text, because the algorithm can be misused to generate and publish fake text.

I am convinced that such an algorithm can help on the way towards a real artificial general intelligence (AGI), but do we have to create progress by fake generation? Apart from this kind of text generation, we can also observe DeepFakes, so fakes created by a misuse of images and videos. I agree that they are funny sometimes, but they can also destroy the reputation of people. Do we believe we can convince potential users by a misuse of the AI technologies? AI developers have a high social responsibility regarding the outcome of their work.

I think common web users must improve their critical thinking skills more and more. There is not always a tool, which can protect us against fake on the web.


Resources:
[1] OpenAI.com, "GPT-2: 1.5B Release", November 5, 2019, https://openai.com/blog/gpt-2-1-5b-release/, last accessed on November, 21, 2019
[2] Adam King, "Talk to Transformer", https://talktotransformer.com/, last accessed on November, 21, 2019
[3] Giulio Starace, "GPTrue or False", Firefox add-on, https://addons.mozilla.org/de/firefox/addon/gptrue-or-false/, last accessed on November, 21, 2019
[4]  Giulio Starace, "GPTrue or False", Google Chrome add-on, https://chrome.google.com/webstore/search/Gptrue or false, last accessed on November, 21, 2019







Back to content