Don’t Get Lost in Translation: How Google Translate and Other AI Tools are Transforming Trademark Law

In-house lawyers who are responsible for their company’s trademarks need to understand the latest developments with regard to translation tools. No longer cumbersome programs that generate gibberish, these tools are changing the way trademark applications are processed and approved. By looking at one such service, Google Translate, it is clear that in-house counsel should become familiar with its capabilities before the trademark examiner uses it against you.

In its article entitled “The Great A.I. Awakening,” the New York Times chronicled Google Translate’s history and explosive growth. Originally launched in 2006, Google Translate has grown to serve more than 500 million monthly users. For the first 10 years of its existence, it was considered good but not great. It wasn’t until Google announced last year that it had upgraded Google Translate using deep neural machine translation technology, a form of artificial intelligence, that it became surprisingly good.

In fact, Google claimed its new Neural Machine Translation system was so good that it reduced translation errors by 60 percent for several popular language pairs. Rather than relying on translation tables, Google Translate uses machine learning to make probabilistic translations based on patterns located in large amounts of training data — a large portion of which is largely taken from translated UN documents. The accuracy of the translation, or the output, is therefore only as accurate as its training data, or input.

Unlike a human translator, Google Translate does not “understand” language in the sense that it interprets the meaning of a word or phrase. Instead, it relies on deep neural learning to make predictions based on a “pattern of patterns” that it finds in its training data and user experience. Before adopting artificial intelligence, Google used an inferior phrase-based technology system that translated phrases by piecing together each individual word, which often resulted in incomprehensible and sometimes absurd results.

Thanks to significant improvements in both accuracy and speed, use and acceptance of Google Translate and other online translators has skyrocketed — including use by trademark offices worldwide. When trademarks containing foreign language words or phrases are filed for protection, the judicially created Doctrine of Foreign Equivalents requires the mark’s foreign translation to be taken into consideration during the prosecution process if an “appreciable number of US consumers” would stop and translate a foreign-language trademark into English. Because the Trademark Office considers many American consumers to be multilingual, it uses the Doctrine of Foreign Equivalents to protect these consumers from confusion resulting from similar trademarks in different languages.

Most often, the Doctrine of Foreign Equivalents is applied when a trademark office or a court assesses the scope of a trademark. This scope is determined by whether a trademark is found to be “confusingly similar” to another trademark in the minds of consumers — and is therefore an infringement. Some of the factors taken into account when evaluating similarity (known as the DuPont factors) include:

• Overall appearances of the trademarks (referred to as “sight, sound, and meaning”);
• Goods or services offered in connection with the trademarks;
• Overlapping channels of trade; and,
• Classes of consumers.

The Doctrine of Foreign Equivalents plays a role in the first DuPont factor: whether two trademarks exhibit similar overall appearances. Since a trademark’s meaning is considered in the context of its overall appearance, the Doctrine of Foreign Equivalents requires that trademarks in different languages be compared if consumers are likely to translate them in their heads, even if the trademarks do not share visual or phonetic similarities. On this basis, for example, the trademark MARCHE NOIR was found to infringe the trademark BLACK MARKET MINERALS even though the trademarks had different overall visual appearances. Because the trademark office considered French a common language, it believed some consumers would believe both trademarks originated from the same source because they both shared the phrase “black market.”

Before the existence of Google Translate and other online dictionaries, trademark examiners needed to guess whether a mark might have a foreign language translation before they consulted dictionaries to verify their suspicion. But online dictionaries now make it easier for examiners to quickly type the trademark into Google Translate and determine whether any foreign meaning exists. This has resulted in some far-fetched translation requests that may not have been made in the pre-Google Translate age.

Issues with the Doctrine of Foreign Equivalents are usually raised early in the application process. A foreign translation of the trademark is requested on the application form itself. If none is specified, and the examiner suspects the trademark to include foreign matter, an Office Action letter requesting further clarification is sent to the applicant. Yet many applicants, especially foreign trademark owners filing through the Madrid Protocol, fail to appreciate the consequences of the request and instead accede all too easily to an examiner’s suggested translation.

Once a foreign translation is submitted to the trademark office, the examiner will use it for two purposes. First, they will use it to determine whether the applied-for trademark is confusingly similar to a prior-filed trademark. If a confusing similarity is found, the applied-for trademark is prohibited from registration. In the past, examiners have found a confusing similarity between an English mark and a foreign mark, two foreign marks in the same language, and also two foreign marks in two different languages.

For example, the Trademark Trial and Appeal Board found the Italian language trademark DUE TORRI (meaning “two towers”) for wines confusingly similar with the Spanish trademarks TORRES and TRES TORRES (meaning “towers” and “three towers”) for wines and brandy. Because the trademarks’ English translations were considered similar and they were being used for identical goods, the Trademark Office believed that consumers who understood both English and Spanish might inadvertently confuse them, believing both products to be produced by the same company.

One way in-house counsel can defend against such a finding involves demonstrating that the trademark in question does not possess only one literal, direct translation. Even though, for example, the Trademark Office considered the trademark PALOMA, primarily translating from Italian to “pigeon, dove” in English, to be confusingly similar to the trademark DOVE, the applicant was able to successfully overcome the refusal by demonstrating that the word “paloma” had English connotations outside of its translation for “dove.” In-house counsel would be wise to consult native speakers, in addition to Google Translate, to assess all possible English connotations associated with the trademark.

Agreeing to an inaccurate or unwarranted translation can jeopardize registration by expanding the pool of prior trademarks that the applied-for trademark is compared against for confusing similarity. Returning to the example of DUE TORRI: The trademark would not only be compared to other trademarks with similar overall commercial impressions, like TORRES, but also to trademarks with similar English meanings but different commercial impressions, like the trademark TOWER. The potential obstacles to registration effectively double.

Second, the foreign translation will also be used by the examining attorney to determine a trademark’s distinctiveness. This is evaluated using the Abercrombie spectrum of distinctiveness, which identifies four categories of possible trademarks: generic, descriptive, distinctive, and arbitrary. Suggestive trademarks, like IPHONE for a smartphone, and arbitrary trademarks, like APPLE for computers, are capable of trademark protection because they are unique for the goods and services. Generic and descriptive trademarks cannot be registered because they involve words and phrases that are needed by others to describe their own products or services.

Because a trademark’s meaning is critical for evaluating distinctiveness, a slight variation in translation can swing a trademark into descriptive, and unregistrable, territory that otherwise might have been found suggestive and registrable.

Take, for example, the Spanish-language trademark BAJATUSEGURO.COM — an online insurance brokerage and underwriting service based in Mexico City. The examiner initially refused the application because he believed US consumers would translate the trademark from Spanish into English to mean “lower your insurance.com,” which he found descriptive of insurance brokerage services and therefore unregistrable. In support of his position, the examiner submitted screenshots from several online dictionaries such as freetranslation.com, spanishcentral.com, and bing.com. He also used the online dictionaries to reverse translate the terms, thereby confirming the translation “loweryourinsurance.com.” Lastly, he submitted a translation made by a human translator.

However, the applicant argued that when the word BAJA is combined with “.com,” it means “download” rather than “lower.” He submitted a declaration made by a native Spanish speaker, a partner at the Mexican law firm that represented him, to support this translation. He also attached a few exhibits to the declaration. One included a screenshot from Google Translate that translated the word “bajar” to mean “get off,” “lower,” “bring down,” and “drop” but not “download.” Another exhibit contained another screenshot from Google Translate — this time translating the phrase “bajar musica” to mean “download music.”

In the end, the Trademark Trial and Appeal Board acknowledged that both translations could be correct as they were supported by human translation. Turning to the online translations, the board warned that “online or automatic translations generated by software may not be totally accurate, or may combine the translations of multiple individual words.” It further explained that one of the online dictionaries, WordLingo.com, even contained an explicit warning message on its homepage to this effect.

As noted by the board, many online translators, especially those who have not yet incorporated advanced neural machine learning technology, translate phrases word by word. This has led to odd situations where examiners have been known to translate a two-word trademark into separate languages. For example, the trademark NOVO NORDISK was translated by the office to mean “new nordic” because online dictionaries translated the first word, NOVO, to mean “new” in Portuguese while translating the second word, NORDISK, to mean “nordic” in Norwegian. The Appeals Board has yet to address such a situation where a trademark contains content from multiple languages. If, according to 2013 US Census data, only 39,430 US residents speak Norwegian, how many of these Norwegian speakers also speak Portuguese? The chances seem slim that the number of Portuguese and Norwegian speakers could end up representing an “appreciable” number of consumers.

Thanks to online translators like Google Translate, and their increasing acceptance by trademark offices worldwide, practitioners are more likely than ever to encounter Doctrine of Foreign Equivalents issues while prosecuting trademarks. Here are some tips in-house counsel can employ to proactively anticipate translation issues for organizational trademarks:

Think twice about accepting a foreign translation. Remember that accepting an inappropriate or inaccurate translation can jeopardize future use and/or the registration of your trademark. Foreign translations are considered when comparing an applied-for trademark with prior marks for confusing similarity and for distinctiveness purposes.

Proactively translate your trademark. Translate potential trademarks using online translators like Google Translate before performing legal clearance, no matter how fanciful the trademark appears. Make sure to translate the trademark as a whole, the constituent words separately, and how the words pair with other words for context. In the BAJATUSEGURO.COM case, the word “baja” was translated by Google Translate differently when translated on its own and when combined with the word “musica.” This may have to do with Google Translate’s algorithms for assessing verbal connotations. While performing legal clearance, clear both the trademark and its foreign translation.

Engage a human translator. As demonstrated by the BAJATUSEGURO.COM decision, the Trademark Office considers human translations to be more persuasive than online translations. This can take the form of a basic declaration made by a foreign speaker who is also fluent in English and should be substantiated with supporting evidence.

Pay special attention to translations in uncommon languages. Online translators are only as “good” as their training data — meaning that the scarcity of reliably translated data for uncommon languages negatively affects the accuracy of those translations. Besides, Google’s new Neural Machine Learning technology has only been implemented for eight language pairs so far: English, Spanish, French, Portuguese, German, Korean, Japanese, Chinese, and Turkish. Translations involving other language pairs are therefore less likely to be accurate — particularly language pairs not involving any of these languages, like Finnish to Afrikaans, where Google Translate needs to translate from Finnish to English and then English to Afrikaans.

Avoid arguing that the language is obscure. While the Doctrine of Foreign Equivalents does not require foreign translation of dead or obscure languages, it is difficult in reality to establish a language as “obscure.” Other than Latin, which is considered a “dead” language, the doctrine is used for most other languages including Spanish, French, Italian, German, Japanese, Chinese, Russian, Polish, and Hungarian. This also includes more uncommon languages like Serbian and Yiddish.

Examiners often point to US census data when establishing whether the Doctrine of Foreign Equivalents applies to a given language. What may seem like a slim percentage of speakers is often sufficient for the Trademark Office. The doctrine applies to Russian, for example, even though only 706,000 people understand Russian in the United States. Because the trademark office considers the doctrine to apply to most languages, arguing that a language is obscure is unlikely to be a winning argument.

While some online translators like Google Translate have experienced significant advancements in recent years, they have yet to achieve human accuracy. In fact, when members of Google’s AI department, known as Google Brain, published a journal paper debuting its Neural Machine Learning software for Google Translate, they acknowledged these limitations in the paper’s title: “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation.”

As with any emerging technology, Google Translate presents both risks and opportunities for in-house counsel tasked with trademark protection. Because trademark examiners in the United States, and increasingly abroad, rely on Google Translate to detect possible foreign translations of trademarks, in-house counsel should be prepared to confront Doctrine of Foreign Equivalents issues during the prosecution process, and should also be aware of the potentially damaging consequences associated with accepting an inaccurate foreign translation or a foreign translation that fails to capture the full spectrum of English connotations. At the very least, in-house counsel should always consult Google Translate early in the branding process and arm themselves with this information when performing pre-filing legal clearance and communicating with the Trademark Office during prosecution to remain one step ahead.