The Robot Inventors

This is a bit of a departure from our usual fare. Rather than talking about specific patents, I want to rant to you about a fresh vein of madness, a new frontier of dumbassery, and a speculative rot in the foundations worldwide intellectual property protections.

I want to tell you about the guy who wants artificial intelligences to be recognized as inventors.


Let’s start with a quick overview of the state of the art. So-called “artificial intelligence,” otherwise known as “machine learning,” generally refers to a type of computer system where you “train” the system with data inputs, rather than explicitly programming its behavior. This is remarkably effective for tasks that would otherwise be prohibitively complicated to design by hand. One example of a problem where machine learning excels is in image processing, and this has led to an unexpected wealth of apps that let you do unsettling things with pictures of your face.

A neural network pasting Stephen Wolfram’s face onto a painting of Isaac Newton, CC BY-SA 4.0

There are many types of machine learning systems, but the category that gets all the attention these days is neural networks. The general concept of a neural network is to have arrays of relatively simple functions that are organized into layers, with connections between the layers governing how the output of one layer affects the input to the next layer. They are called “neural” networks, because they draw their inspiration from the connections between neurons within the brain.

There is nothing magic about a neural network. For the mathematically inclined, it is just a function that translates an input to an output in a deterministic fashion. For all that the name suggests a brain in a box, neural networks are sorely limited in their ability to process novel inputs.

Now, I don’t want to undersell them. Neural networks are phenomenally powerful for certain types of problems. Anything that involves pattern analysis is ripe machine learning. But the most powerful artificial intelligence systems are still constrained by the human element, because they require human inputs to create their training dataset. A facial recognition system would be useless, for example, without a large set of pictures that already have their faces labeled to learn from.


Now let’s take a turn into patent law. You, brain-having human that you are, come up with an idea. Being the savvy and possibly ego-driven sort, you go to file a patent for your idea. Part of this process is that you sign a declaration, swearing that you are the inventor. In the United States, that word is defined to mean that you “invented or discovered the subject matter of the invention.”

To be an inventor, you must have contributed to the “conception of the invention.” The words “conception” and “inventor” end up holding an awful mess of assumptions about what it means to think and to be a person. For example, we acknowledge that a person could be involved in the inventive process, but not participate in the conception. Imagine a lowly laboratory technician mixing test tubes, without any idea of their purpose or what the results mean. That technician has not contributed to the conception of the invention and is not regarded as an inventor.

On the other hand, the threshold for conception is pretty low. If it was Ezekiel’s idea to add racing stripes, and that feature ends up in the claims of the patent application, then ol’ Zeke is going to be an inventor regardless of whether racing stripes are important to the overall invention.

Up to now, we have never given any thought to the tools used in the process of conception. The USPTO has no interest in the Muses. This is the federal government, after all, and inspiration has no currency. As a result, we have never stopped to ask whether a given invention was inspired by a computer, a poem, or a particularly gnarly looking cloud.

Behold, my muse. “Nekrogoblikon at Rock am Ring 2013” by Achim Raschka. CC BY 3.0


So you’ve got your machine learning, and you’ve got your conception of the invention. What if… what if we machine learned to concept invent? Enter the DABUS: Device for the Autonomous Bootstrapping of Unified Sentience. You can find the website for “The Artificial Inventor Project” here:

Their approach is to use a neural network to generate some useful output, and further to judge whether the output is useful. That second part is crucial, because it forms the basis for an argument that the role of the neural network was more than just that of a machine, that it had the capacity to understand what it created as an invention.

To help flesh this out, imagine a random number generator that is configured to spit out gene sequences. Inevitably, it will eventually spit out a gene sequence that is potentially useful. I would argue that the GENErator is not an inventor, because it has not conceived of this gene sequence as an invention. But what if a human is monitoring the output and sees that useful sequence and says, “Eureka!”? That human has recognized the utility of the gene sequence and has conceived of an invention.

If you can replace the human in that scenario with another machine, trained on a dataset of good inventions and garbage inventions, who is the inventor of the output when it identifies something as good? If I write that output up as a patent application, who signs the declaration?

The Artificial Inventor Project is testing this by filing patent applications that are ostensibly invented by DABUS, naming DABUS as the inventor. They have been rejected at the USPTO (you can find decisions on the subject here and here), with an appeal pending at the U.S. Court of Appeals for the Federal Circuit, but they have had success so far in South Africa and Australia.

In the U.S., our patent system is heavily focused on the individual inventor. So far, that has given the USPTO the ability to hold that the act of conception must be performed by a “natural person.” Other countries do not have that same individualistic emphasis. I will certainly be watching how the DABUS appeal in the U.S.


The question unanswered, and apparently unasked, is why would you even want to do this? The driving force appears to be a man named Stephen Thaler. Dr. Thaler is going through a lot of trouble to nail jello to the wall; whether the “inventor” is the machine or the owner of the machine, it won’t affect who owns the resulting patent, since the machine can’t own property. Their website’s copy suggests where the true motive lies:

Arguably, DABUS may be considered “sentient” in that any chain-based concept launches a series of memories (i.e., affect chains) that sometimes terminate in critical recollections, thereby launching a tide of artificial molecules. It is these associated memory sequences, and the accompanying simulated neurotransmitter rush, that are considered equivalent to subjective feelings in humans (i.e., sentience). In this way, DABUS has an emotional appreciation for what it conceives.

What we have here appears to be a True Believer, chasing the dream of artificial general intelligence, independent of any profit motive. After all, if Dr. Thaler just wanted to take DABUS’s inventions for himself, he could simply have named himself as the inventor and nobody would have questioned it. Instead, he is using this as a platform to force the question of machine personhood on an international scale.

So, even though I personally believe that Dr. Thaler’s crusade is a waste of time, money, and chutzpah that could be better spent elsewhere, I have to respect the effort. If nothing else, it’s fodder for smalltalk at all the patent lawyer cocktail parties I go to.

2 thoughts on “The Robot Inventors

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s