Why AI and neural networks can be scary

Over the past couple of years, we have seen a lot of cool new technology and concepts that quickly became part of our daily lives (AI, Cryptocurrencies, …). But taking the benefits that we got aside, technology can also be scary. At the least when we saw all the dystopian futures in „Black Mirror“ we had to learn that.

Given that, I was kind of shocked when I first had something to do with artificial intelligence and neural networks. I personally think that the effectiveness and reliability that already exists in this part of computer science is somehow scary. Do not take me wrong, I am really interested in the development of technology and I am looking forward to see what will experience in the future. Also I am not some kind of psychologist or professor for artificial intelligence. All I want is opening your eyes and reminding you that everything has its downsides and we should be careful with what is coming.

But step by step. Neural networks are clusters of neurons that are connected with each other, very similar to the nerve system of living creatures. The computer sciences uses these networks for tasks such as classification, image detection, game AI and lots more. For everyone that wants to get into the topic, Google released a framework called „Tensorflow“ back in 2015.

This framework allows you the creation and maintenance of such an network in your own program. Finding code that uses this framework and adapting it in your own context is extremely simple. Go over to github.com and take a look around, it wont take you very long. Some small adjustments to the code and you are good to go. And this is the surprising and kind of scary part for me: We already have all the tools we need, and their result is stunning.

What I mean by stunning? Let me explain by my own example: I recently had a small project about classification with the help of a neural network. To understand the basic concept and how we can get into the topic, we thought about doing classification on amazon reviews: Insert the text of a review and let the computer decide how many stars should go with the text. Finding some basic code and doing adjustments to it was pretty simple. Data of reviews from amazon are easily available, already correctly formatted.

Then the first phase could start: Training.

A neural network needs to be trained before it can do anything. It is basically like teaching an animal to distinguish between two colors and make it respond differently. So we trained the network with around 650.000 reviews from all kinds of products from amazon.com. This took about 1,5 hours on a notebook with an i7 7920HQ CPU. And that's it, you are ready to go.

We are already in the second phase: Running.

Now all that is left is inserting a text in the program and let the computer decide the rating. And how precise it does that was absolutely stunning for me. I wont say it is perfect, especially sarcastic comments are an issue for the network, but the overall result as amazing. The training starts with an accuracy of 20% (1/5 chance of guessing the correct rating) and rapidly increases over time before the climb starts to get smaller and smaller. At the end, nearly 60% of the entered texts, no matter how long, were rated correctly. And when you include the results that were rated with a difference of one start, the accuracy goes up to above 90%. With more time and training data, you can still improve this result by some percent.

You still might think why this is „scary“. I will tell you. We already live in a world of fake reviews and ratings everywhere. People are paid to write fake text all day to promote products or keep their opponents down. Now just imagine how simple it could be for company's to let their own bots work on amazon. Generic text generation is around since years. Simply enter the product name and let the bot create the text. Then have your neural network go over the text and generate a review for it (4 or 5 stars to make it seem as legit as possible). Or the next time you write some company's support on Facebook, do you really know a human being is answering your requests or questions? They may have collected data for years to train an AI program to give you an answer that you and make you think you are talking to a human.

As I said, do not understand me wrong and do not think I am paranoid, I just want to give you a little heads up to the risk that every new technology brings with them!