The 4 biggest disadvantages of artificial intelligence

Artificial intelligence, also known as AI, is announced as THE great revolution of the 21st century.

The government has even decided to inject 2 billion euros into the training and recruitment of artificial intelligence specialists.

Because its potential is substantial and inspiring.

While the media, films and video games have taken up some futuristic applications, others are very real:

• Autonomous cars
• Cancer detection algorithms
• Image analysis
• 3D protein prediction, enabling much faster iterations in biology and vaccine research

A company that hears about artificial intelligence will conclude that it’s the solution to every problem.

When it comes to taking action, it’s a different kettle of fish.

Resources, technical skills, acculturation… I tell you all about the obstacles to adopting artificial intelligence in this article.

Spoiler alert: it’s still accessible to everyone, with just a little knowledge. We tell you all about it in this other article.

Disadvantage 1 – The technical challenges of artificial intelligence

It may seem simple to use artificial intelligence without understanding it. Unfortunately, when it comes to debugging, things can get a lot more complicated.

N.B. for the beginners among us: debugging means eliminating malfunctions from a program.

Although I said earlier that artificial intelligence is accessible to everyone, let’s be honest: it requires a wide range of skills, including computer science, engineering and maths.

So it’s essential to try and understand optimisation issues and the different models involved.

The first challenge is to understand the mathematical models.

Basically, if you understand the principle of optimisation, machine learning algorithms are nothing more than particular functions to be optimised. Nothing more, nothing less.

The second challenge is technical skills, and development in particular. While modelling is essential, orchestrating training and putting it into practice are also key steps in :

• Collecting input data
• Retrieving a model that is already learned or is starting from scratch
• Making it learn
• Verifying results
• Making it available for predictions

To use an analogy, once you’ve designed the mechanism of your toy, the next step is to create the plastic shell you’ll use to dress it up.

Speaking of development, Python is the language of choice for artificial intelligence. Take our free course here to learn the basics.

Some examples of tools to master:

• Flask and Django for packaging your code in Python
• Mlflow for development
• Airflow for orchestration
• Docker and Kubernetes for deployment

You should also bear in mind that every company already has its own IS network. So integrating new tools can be a real challenge.

Disadvantage 2 – Resources to train an artificial intelligence algorithm

Beyond technical skills, artificial intelligence algorithms are also very resource-intensive.

Take GPT 3, for example, an artificial intelligence developed by OpenAI, the AI research company co-founded by Elon Musk. It is capable of creating written content with a language structure worthy of a text written by a human. This invention is one of the most significant advances in AI in recent years, as the algorithm has not been trained for a specific task, but has a truly global “understanding” of language.

Training an algorithm like this requires almost 45 TB of textual data. If you take a simple text note on your computer, it weighs just a few Kb, which represents almost 45 billion Kb.

If you wanted to host GPT 3 in random access memory (RAM) on your computer, you’d need 175 gigabytes of memory. With a powerful PC, you only need 16 gigabytes. So it’s clear that this system is reserved for particularly powerful infrastructures.

It should be known that it’s generally the computing power that is costly.

If you wanted to reproduce a GPT 3 on your own, even if you managed to find the right resources, you’d need 355 years of training for a total of $4.6 million.

Disadvantage 3 – Acculturation to AI

In the corporate world, when you talk about a data project, everyone usually has stars in their eyes. In reality, implementation is usually more painful.

Here’s an exhaustive list of the disadvantages of artificial intelligence for acculturation:

• Fear of the unknown: the workings of AI algorithms may seem obscure to some. This lack of knowledge can be frightening and problematic.
• The relatively long development cycle: in practice, you won’t be able to use trained models as they are. You’ll have to run tests, play with parameters and run more tests. All of this requires a good number of development cycles.
• Governance: you need to make sure you know where the data is and what it corresponds to at all times. Teams need to be made aware of these issues.
• Ethical issues: it is indeed possible to encounter algorithmic biases. Keep in mind: these biases always depend on the data you integrate during the training phase. This is what we call “trash in trash out”: you put biased data in, and it comes out biased.
• Return on investment. This type of project is very expensive: the recruitment of engineers, IT resources… And yet, it’s impossible to know the revenue before the launch. No ROI is therefore guaranteed.
• Sharing results. If your results contain errors, it’s very difficult to analyse them in detail. When you’re working on models containing billions of parameters, you won’t be able to determine the path taken by your algorithm. So you need to make your stakeholders aware of this and work on the “explainability” of your model.

Disadvantage 4 – Artificial intelligence and its algorithmic biases

In my opinion, this is one of the major disadvantages.

Watch any video, such as a replay of The Voice. The algorithm will then suggest other videos from the show, as it has been optimized to keep you on the platform as long as possible. In fact, this dwell-time is one of the indicators most closely monitored by Youtube in particular.

This method is known as collaborative filtering. It’s this system of recommendations that has made Netflix’s name.

So if you’re watching conspiracy videos, the algorithm will keep suggesting more and more conspiracy videos. You’ll end up with a mass of potentially erroneous information on this theme. As you only see this side of the story, you’ll end up believing it.

Some social networks, like Youtube, are trying to combat this bias by deciding what type of content can and cannot be shown. This opens up another ethical debate.

There are also many cases of racism in the algorithm results. As I said earlier, it all depends on the input data.