Teaching the machine
In order to teach an AI machine to create an artistic work, it's necessary to provide it with data. If the input data consists of works of art where the copyright has expired (because it's old), any copying during the process will not infringe. The Next Rembrandt project fed data from numerous authentic Rembrandt paintings into a computer1. Given that the paintings were over 350 years old, there was no danger that this process would infringe copyright in the original paintings.
If the input is not capable of being a copyright work (or other protectable subject matter), then copying this would not infringe either. Data taken directly from the weather, oceans, plants, animals, the planet Mars, etc. would not amount to copyright works (assuming that the data is collected at source2). Other types of data, such as sports and financial data, or copying data which has previously been recorded by a third party who may own rights in it, may or may not be protected. For example, some sports data may be protectable by database right.
On the other hand, if the input data comprises copyright works, does the person teaching or operating the AI infringe without a licence? If the process of teaching (whether during the preparation phase or during the actual process of machine learning) requires that all or a substantial part of an earlier copyright work be reproduced (whether directly or indirectly), then this would likely infringe copyright unless the use was licenced or a defence applied3.
However, if teaching of the machine does not actually involve any copying of the earlier work (say only certain statistics are recorded such as the distance between the eyes and nose of a figure in a painting), then it is possible that there is no infringement. If none of the expression of the author's own intellectual creation (EU test) or any of the skill, labour and judgment (traditional UK test) has been taken from the earlier work, then there may be no infringement. This is a question of fact.
The machine itself
It seems unlikely that the AI machine itself, once taught, would infringe copyright because the machine may only comprise an algorithm and a network of weighted connections. If the machine contains a digital record of earlier copyright protected art, then this would probably require a licence. In other words, if the machine has been trained on numerous artistic works, this does not entail that the resulting machine itself contains any such works (whether in digital form or otherwise). The taught machine essentially contains rules (in the form of a neutral network) and not artworks, even though it developed the rules from analysing many artworks.
The machine's output
Will the machine's output infringe an earlier copyright work? If the machine produces a work of art which is a direct or indirect copy of an earlier copyright work or a substantial part one and there is no defence, this will be an infringement. However, it is important to note that if the work was independently created by the AI (ie if it was not a direct or indirect copy of a copyright work), however similar it is to a pre-existing work of art, then there will be no infringement. So, if the operator, the machine and the programmers had not seen or used the claimant's work, then there should be no infringement however similar the output is to the claimant's work. If AI machines produce billions of works of art, some of them are likely to resemble earlier works, even if there had been no copying. This is more likely to occur than the scenario in the infinite monkey theorem4.
Even if the AI generated art resembles the claimant's work and causation is proved (eg because the claimant's work was part of the input data), infringement does not follow automatically. The claimant will need to prove subsistence and title and sue the correct human and/or corporate defendant. Furthermore, even if all these things are proved, the question of substantiality can be difficult to determine with certainty, as the red bus case5 below perhaps illustrates (see the images below, with the claimant's work on the left).
Conclusions on AI generated art
In this 6-part series, we have looked at the crucial tests of originality and considered whether a human author is necessary for copyright to subsist. This may depend on whether the test is the EU or traditional UK test. We have also looked at authorship, term and infringement issues. At present, this topic feels relatively new and untested. However, as AI becomes increasingly common in the creative industries, the answers to the questions discussed in this series will become well-established.
1 See https://www.nextrembrandt.com/. The resulting work is a new 3D-printed work of art which looks like it could have been painted by Rembrandt.
2 If the data is copied from a third party recording of it, then it is possible that rights may apply to that. However, that is beyond the scope of this article.
3 Whether the temporary copies exception applies is a question of fact. If any permanent copies of artistic works are made during the process, then the exception would not apply to such copies for example.
4 "The infinite monkey theorem states that a monkey hitting keys at random on a typewriter keyboard for an infinite amount of time will almost surely type a given text, such as the complete works of William Shakespeare": wikipedia.org.
5 Temple Island Collections v New English Teas [2012]. The court held that the defendant's work had infringed the copyright in the claimant's work.
