OSince last year, governments, academia and industry have invested considerable resources into investigating the harms of advanced AI. But one important factor seems to be continually ignored: right now, primary AI tests and models are confined to English.
Advanced AI can be used in many languages to do harm, but focusing primarily on English may leave us with only part of the answer. It also ignores those most vulnerable to its damage.
Following the launch of ChatGPT in November 2022, AI developers expressed surprise with a capacity displayed by the model: it could “speak” at least 80 languages, not just English. Over the past year, commentators have highlighted that GPT-4 Outperforms Google Translate in dozens of languages. But this focus on English for testing leaves open the possibility that assessments may be overlooking the capabilities of AI models that become more relevant for other languages.
As half the world goes to the polls this year, experts are voicing concerns about the ability of AI systems to not only be “superspreaders of disinformation”, but also its ability to to threaten the integrity of elections. The threats here go by “deepfakes and voice cloning” to “identity manipulation and AI-produced fake news”. The recent launch of “multi-models” – AI systems that can also talk, see and hear everything you do – like GPT-4o and Gemini Live by tech giants OpenAI and Google, look set to make this threat even worse. And yet virtually every policy discussion, including the one in May Historic AI Safety Summit in Seoul and the launch of the long-awaited AI Roadmap in the US Senateneglect languages other than English.
It’s not just about leaving out some languages at the expense of others. In the USA, research consistently demonstrated that English as a Second Language (ESL) communities, in this predominantly Spanish-speaking context, are more vulnerable to misinformation than English as a Primary Language (EPL) communities. Such results were replicated for cases involving migrants in general, both in the United States and in Europe where refugees have been effective targets – and subjects – of these campaigns. To make matters worse, content moderation barriers on social networking sites – a likely forum where such AI-generated falsehoods would proliferate – are heavily biased towards English. Although 90% of Facebook users are outside the US and Canada, the company’s content moderators I just spent 13% of their work hours focusing on misinformation outside the U.S. The failure of social media platforms to moderate hate speech in Myanmar, Ethiopia and other countries involved in conflict and instability further reveals the linguistic gap in these efforts.
Even as policymakers, business executives and AI experts prepare to combat AI-generated disinformation, their efforts cast a shadow over those most likely to be targeted and vulnerable to such false campaigns, including the immigrants and those living in the Global South.
See more information: OpenAI used Kenyan workers earning less than $2 an hour to make ChatGPT less toxic
This discrepancy is even more concerning when it comes to the potential of AI systems to cause mass human casualties, for example, by being used to develop and launch a biological weapon. In 2023, experts expressed fear that large language models (LLMs) could be used to synthesize and deploy pathogens with potential pandemic potential. Since then, a multitude of research articles investigating this issue have been published within and outside the industry. A common conclusion from these reports is that the current generation of AI systems are as good as and no better than search engines like Google at providing dangerous information to malevolent actors that could be used to build biological weapons. Research from leading AI company OpenAI yielded this discovery in January 2024, followed by a RAND Corporation report that showed a similar result.
What is surprising about these studies is the almost complete absence of testing in languages other than English. This is especially disconcerting given that most Western efforts to combat non-state actors are concentrated in regions of the world where English is rarely spoken as a first language. The claim here is not that Pashto, Arabic, Russian or other languages can produce more dangerous results than English. The claim, instead, is simply that the use of these languages is a capability leap for non-state actors who are better versed in languages other than English.
See more information: How the global dominance of English fails us
LLMs are often better translators than traditional services. It is much easier for a terrorist to simply enter their query into an LLM in a language of their choice and directly receive a response in that language. The counterfactual point here, however, is to rely on clunky search engines in your own language, use Google for your in-language queries (which often only produces results published on the Internet in your own language), or go through an arduous translation and re-elaboration. translation to obtain information in English with the possibility of loss of meaning. Consequently, AI systems are making non-state actors as good as if they spoke fluent English. How much better this makes them is something we will find out in the coming months.
This notion – that advanced AI systems can provide results in any language that are as good as if requested in English – has a wide range of applications. Perhaps the most intuitive example here is “spearphishing,” targeting specific individuals using manipulation techniques to obtain information or money from them. Since the popularization of “Nigerian Prince“fraud, experts postulate a basic rule to protect yourself: if the message appears to be written in broken English, with inappropriate grammatical errors, it is a scam. Now these messages can be worked by those who have no experience with English, simply type the prompt in their native language and receive a fluent response in English. For starters, this says nothing about how much AI systems can drive scams where the same non-English language is used at input and output.
It is clear that the “linguistic issue” in AI is of paramount importance and there is a lot that can be done. This includes new guidelines and requirements for testing AI models from government and academic institutions, and encouraging companies to develop new benchmarks for testing that may be less operable in languages other than English. Most importantly, it is vital that immigrants and those from the Global South are better integrated into these efforts. Coalitions working to keep the world safe from AI should start looking more like this.
This story originally appeared on Time.com read the full story
