What machine learning can say about protection by copyright or design patent: Example of typeface designs (fonts)

Behind fantasies and myths, artificial intelligence (AI) is a technical reality that is progressively gaining ground in just about every industry sector, this including intellectual property (IP). Carefully designed algorithms and clever analytics allow useful measurements to be performed. In particular, it is possible to measure distances between works, between IP rights (patents, trademarks), or between a given IP right and an alleged infringement.

Here is an example of how machine learning techniques can be used to compare works such as typeface designs (i.e., fonts). In addition to computerized learning, simple statistics allow the originality of a particular work to be assessed, which is helpful in determining whether the work can enjoy protection by copyright or a design patent.

Some comments first:

• The terms “font” and “typeface” are often used interchangeably. However, a font more specifically refers to a particular size, weight, and style of a typeface. In modern usage, a font typically means a digital file used by a computer system to display corresponding glyphs. For example, the Arial typeface comprises many styles, including Regular, Italic, and Bold. See, e.g., this Wikipedia article.
• Not all countries allow the artistic design of a typeface to be protected by copyright law. E.g., the law in Germany and U.K. protects typeface designs, whereas copyright in the U.S. only protects the corresponding digital files. Besides copyright, protection for typeface designs may be obtained via design patents. Trademarks only protect the name of a typeface. All this is explained here in detail.
• Originality is a prerequisite for protection, whether by a design patent or by copyright. A minimal degree of creativity is needed in both cases, though the threshold of originality required depends on jurisdictions, topics, and the actual type of IP protection sought (copyright or design patent).

In fact, this work began a few months ago, when a former colleague asked me whether machine learning (ML) could be used to assess copyright. He had actually created his own typeface and wanted to know to what extent his creation could be protected by copyright laws.

The following describes a simple pipeline addressing this question. First, a set of images are generated, to be able to compare the typefaces. Second, features of the generated images are extracted. Third, a distance matrix is computed according to the extracted features. Finally, simple statistics are gathered in view of assessing the originality of the new font.

So, we first need a comparison basis. We may for example use a pangram— a sentence that involves all letters of the English alphabet, to which the 10 digits are added, as routinely done to compare fonts. Instances of the resulting text are displayed below, using a few popular fonts.

The list goes on. In total, 101 popular fonts were selected. Dingbat fonts were discarded, as well as redundant fonts (i.e., resulting in same glyphs for the English alphabet). Only regular font styles were included, for the sake of consistency in the comparisons with the new font below. As I cannot divulge the font designed by my former colleague, I will use a new font designed on purpose:

This typeface was drawn after Arial as a bitmap. Elementary image processing was then used to dwindle ends of the glyphs, with very minimal creative inputs, on purpose.

We now want to assess the originality of this font, in order to conclude as to the possible extent of protection, be it in terms of copyright or design patents. To that aim, we first automatically extract features of all pangrams, before computing the distance matrix according to the extracted features. An excerpt of the resulting matrix is shown below.

The darker the closer. Zero distances on the diagonal mean that any font is identical to itself. All numerical values were renormalized (so that the value 100 corresponds to the average distance) and then rounded.

Analytics show that the distances between the prior fonts are quite narrowly distributed about their average value, i.e., 100. Another important indicator is the minimal distance between any pair of prior fonts, which, on average, is 72. For example, the average distance between Arial and other prior fonts is 99, but its closest font (besides the new font) is at a distance of 41. In comparison, the distance between Arial and Arial Narrow (a stylistic variant) is 52, and the maximal distance (207) is found between Lucida Console and Impact.

In contrast, the distance between the new font and Arial is only 29, i.e., smaller than any of the distance values evoked above in respect of the prior fonts. In fact, only one pair of prior fonts was found at a lower distance (27).

As expected, the values obtained indicate that the creative effort to create the new font is significantly less than the usual efforts in designing new fonts.

So, the adopted pipeline makes it possible to quantify the originality, a thing that can be used in an assessment made by an IP attorney, rather than as a substitute for her/his judgment. That is, the IP attorney is now given additional facts s/he may want to consider, in order to assess the extent of IP protection or, conversely, the risk of infringement.

Can this new font enjoy protection by copyright or design patent?

On the one hand, 29 is strictly larger than zero, i.e., the new font slightly differs from Arial. But, on the other hand, even stylistic variants within a same font family can be further apart. In fact, all prior font pairs are, subject to one exception, the minimal distance between prior fonts being 72 on average.

So, maybe a talented lawyer could manage to defend the originality of this new font. However, considering all the numbers above, one should perhaps not expect too much in terms of IP protection in that case.

To conclude, this short study illustrates how machine learning and analytics can be exploited to assess distances between works and use the latter to assess the extent of IP protection or, similarly, the risk of infringement. Interestingly, the same approach can be used to assess distances between any type of datasets (e.g., 3D images, 3D printer files) or, in fact, any kind of works (audio, images, e.g., paintings, text).

Note, this study adds to a previous work directed to patent monitoring, available here. In that case, a pipeline was devised to rank utility patents by order of relevance, based on ratings from a user. An advantage of this pipeline is that it does not require any description of the invention. On the contrary, it can capture multiple activities of a company. Moreover, similar techniques can be exploited to compare trademarks, as to be reported soon.

Thus, machine learning and related techniques may potentially benefit to various sectors of IP law. Such techniques are appealing as they can learn from and then quickly process vast amounts of data. However, while the outcomes are often remarkable, the underlying pipelines can hardly capture all subtleties of legal analyses. So, we can perhaps trust, but we must also verify.

Sébastien Ragot

#machine learning #AI #Intellectual property #copyright #design patent #typeface #font