Internationalization and official translations of international documents related to ethics and ethics of AI/S

[fititnt]

There are some important documents to help implementers of artificial intelligence and autonomous systems and for countries that do not speak English natively this could be an additional challenge. And even for countries that are not among the 6 official languages of the UN, it seems that the challenge may be slightly higher, since some human rights documents may not be simple to find in other languages.

In this topic I will copy and paste some texts that I spent working writing about this subject while interacting with other groups. I will not copy answers that are not mine.

It is worth remembering that my opinion can evolve over the years.

[fititnt]

On Apr 30, 2018

---

This is proposal for a free discussion about Internationalization and official translations of international documents related to ethics and ethics of AI/S. Internationalization/Localization are always challenging in any organization who works with normative documents, so is reasonable assume it would be at IEEE global initiative and any other working groups and organizations who promote documents related to ethics of AI/S.

 

This discussion is not exclusive about a possible translation of final version of Ethically Aligned Design (EAD) document. Let's think ahead: what documents (that allows free translation/sharing) could be very useful to those who implement laws, teach about the subject or would influence the public opinion about issues related to the EAD document? For documents that do not have translations: is possible to build a sorted list that could be used as reference to other people prioritize translations? Are there specific issues that do not have permissive license documents that could be suggested to third parties to create?

 

 

By free discussion it explicit means

 

1. Each person express their own personal opinion that do not necessary represent the organization or group you make part of
2. Suggesting does not imply being the person who will implement or lead; in special those who have internationalization experience do not need feel forced to take on work.
3. It is ok to disagree with the way or even the need of translations in addition to English; (some examples: to discuss the need for minimum standards to consider official translation; argue that certain documents do not need a very strict translation by the nature of the document itself, etc.)
4. Be creative; it's ok to use some inductive reasoning; also not all documents need to be translated using the same method

 

Some external links that could be useful to read

• Internationalization and localization, Wikipedia
• List of languages by number of native speakers, Wikipedia
• List of languages by total number of speakers, Wikipedia
• Official languages of the United Nations, Wikipedia

 

---

 

I will start the discussion with some personal comments about the impact that translations could do in non-natively speaking countries of English.

 

• Recommendations or normative documents on ethics can make sense to be accessible also to humans who do not work in the area
  • Humans who work with implementation, even if well well-intentioned
    • could fear retaliation, but general most likely to guarantee implementation
    • could fear misinterpretation, as saying one very specific part that is not good, could make general public force block good cases
  • Hypothetical example: a country who breaks laws that one international document strongly recommends could, at best, do not cooperate with translations, and at worst, change it's definitions
• English is the most common language for international documents, but English is not the most native spoken language
  • and, when, as second language, can have several problems on interpretation
• Humans with very proficient translations skills can lack of understand technical terms; 
  • ...and the humans who understand those technical terms can be less skilled at translation.
  • if a full document is not possible to translate, what additional documents could be useful? 
    • Maybe glossary?
• Could the lack of translations validated by humans make documents on ethics and ethics of AI/S be prone to adversarial AI?
  • The way some translations engine work as today can make it a vector it prone to very extreme errors: a bad actor on future could make more minimal changes on some translations in several translations engines
    • Open question 1: could validated translation be used as stronger parameter for automatic translations (same language)?
    • Open question 2: could other validated translations be used as additional parameters for translations engines be more resistent to adversarial attacks?
      • Maliciously trying to fool an ethical statement in Portuguese and the engine using some anchorage in official translation in Spanish and English.
  • Machine translation soon will be better than average non-native English speaker, if is not already at 2018.
    • And as better it gets, more expectation humans with low skills will have about it.
    • And some machines (like a big company who make products on several country) rely on automatic translations
• Reference documents can be heavily used to make local laws, so the language is one additional factor on influence to avoid delay
  • Countries with more English speakers and/or are on some area of influence (think European Union) possibly could be more heavily influenced by international recommendations written only in English.
    • It means some countries are more affected by missing translation. No Portuguese translation affects more Brazil than Portugal; No Spanish translation could affect more Bolivia than Spain; No French translation could affect more some african countries than France
  • No translation could hypothetically make a subject like artificial intelligence even more elitist (promote "the bad side" of Technocracy / Meritocracy) and after some time eventually force the bad side of populism.
    • Note: I'm not a social scientist, but I know that when the economy of a country is affected, the knowledge of the population on a subject and the need of something to blame can force decision makers take less optimal decisions
    • On a extreme scenario, the lack of translation of a international document (and only the existence of short or even optionated versions by local organizations who could do "something unpopular") could also lead to aversion on the subject as the locals could be induced to reprove anything promoted by some other local organizations
      • As reference, in 2018 human rights defenders in cities like Rio de Janeiro in Brazil can be perceived as unpopular fueled by false dichotomies (i.e. by protecting the rights of a prisoner in jail to be seen as disrupting police service)
      • If there is no international ban (or national, in the case of Brazil) it is plausible that in the future the population of cities such as Rio de Janeiro will be in favor of autonomous weapons for use within the city
      • In short, is not just about ethics of AI/S, but ethics in some places of the world.

 

---

PS.: The author of this discussion lives in a country that natively speaks the seventh most spoken language in the world, Portuguese. One of the organizations that manages human rights documents, the UN, has 6 official languages, and Portuguese is not one of them.

[fititnt]

May 2, 2018.

---

I will use as example one impact of web accessibility caused by missing / wrong translations and use inductive reasoning on the impact of missing / wrong translations of international documents on ethics and ethics of AI/S.

 

Years ago, the web was transitioning from HTML 4.01 to "HTML5" (HTML 5.0). When I was on the road to becoming a web accessibility expert one day I was shocked to see that one of the sites I had done had some specific accessibility errors as per the HTML 5 documentation in English. I went to investigate and realized that most Brazilian sites had the same errors, except those written in the previous version, in HTML 4. I spoke with other Brazilians, and even those who were also speakers at technical events were shocked, because some specific mistakes we were even teaching wrong.

 

Websites using HTML 5 were less accessible than those written with HTML in older versions. That is, there was a regression in implementation of a new technology (in HTML semantics), which made it less user-friendly to screen readers.

 

The probable main reason for this was the following: there was an error of translation in technical documents of people who where experts and no one was aware of this impact, even they. On the following years, many examples of implementation also induced people to write wrong code, less semantic HTML 5 than would be with the old version, even in articles claiming to explain semantic HTML. In the countries of Portuguese speakers (or at least in Brazil), this became an echo chamber.

 

Not surprisingly, websites that did not have these error patterns were made with code outside of Brazil: a wordpress website with free international template without local customization could easily be more accessible by default.

 

Even after this specific type of issue become discussed, some people updated their websites, but the ones who where more interested in changes where more the new developers. But even they where afraid of make some errors, as they also had one additional mental load: some key terms where in English and most material in Portuguese lack of deep and focused more on other features of the newer version.

 

If at start some translations errors or misinterpretations lead to some summarized or new created documents that amplified more misinformation years later, then the lack of translation of original documents also lead to new people have a hard time when these errors where perceived. On this specific example, the more senior developers preferred to stay silent. And at least one person who sent material to explain how to solve such problems at a national technical conference was not accepted. This is at least an uncomfortable situation, as the root problems where not complex.

 

> Inductive reasoning with possible lack of translations on ethics and ethics of AI/S

 

Some similarities:

 

• Web accessibility is generally accepted as something that every web project should have, even if the implementer is not a specialist ; those specialists also agree that web development courses should explain the basics of accessibility when do not have some specific non-compulsory course
  • We expect AI/S implementors to also have baseline knowledge of AI/S ethics
• Web accessibility specialists are very scarce, perhaps because market demand is low and high time to learn; such experts may even be perceived as possibly causing product delivery delays when called in the last step
  • For now I will assume that the people who knows both ethics and AI/S are scarce too
• Web accessibility specialists can make errors, but web developers could amplify by not questioning
  • Any human can make errors

 

So, my inductive reasoning here is that without some strategies that could mitigate some human errors, the area of ethics of AI/S could have some smaller misinterpretations reinforced with echo chambers. 

 

Most of the ideas of documents such as Ethically Aligned Design, and other human rights documents are very reasonable, people tend to agree. It is possible that people in the early years in their countries of origin discuss topics that draw more attention, but the problem is precisely in the details where blind spots may occur. The example I gave about web accessibility in a country was a blind spot. Probably when it started people were busy with other immediate issues, maybe similar on AI/S ethics about no autonomous weapons & basic human rights is today.

 

 

> The ethics documents not specific to AI/S

One of the main reasons of my post here and my effort to make clear for everyone that this beyond AI/S ethics and that by listing reference documents it can be used later for people on their countries.

 

I was looking for translations to Portuguese (the 7th most spoken language in the world, and the UN officially use 6 languages) at https://github.com/EticaAI/docs.etica.ai/issues I got scared when I found that some documents could have multiple translations (like one from Organization of American States have differences from the document at Brazilian Chamber of Deputies) and that is even possible that some laws around the world for those countries outside the 6 UN languages may have been influenced by opinions based on inaccurate translations.

 

Be aware of the big picture is think that even the UN will probably need help. They are discussion a lot of AI impact but exist documents 50 years old that are important for the local implementation and that needs some working on it, it is not just the Universal Declaration of Human Rights.