By David Stephen
An approach to AI safety could be a derivative of language translation, where access to the original content is accessible to the receiver. In a lot of use cases for language translation, an individual would have the original text translated, then send, but the receiver only gets the translation, which conveys the message – but has no access to the original.
Machine Translation
Often, translating a message from a language to another and translating back shows some differences from the original, and could even continue to change in pieces, over several iterations, depending on the language. While language translation is competent enough to provide the communication, it could be viable, for AI safety, to have translations come with an ID, so that the original message is accessible or retrievable from the platform, within a timeframe – by the receiver.
Could language translation model some AI Alignment?
The necessity for this, as a language translation option, may be small percentage, especially if the receiver wanted extra clarity or needed to check the emphasis in some paragraphs, or even knows the original language too, but the importance could be a channel towards AI safety.
One of the questions in AI safety is where do deepfakes come from? There are often videos with political or cultural implications, or some AI audio for deception, or some malware, some fake images, or texts. There are several AI tools, just like translation platforms, that indicate that they do not store data, or the data is removed after some time. This appears appropriate, ideally, for privacy, storage, as well as for several no-harm cases. But it has also made misuses easier and several – with consequences.
For prompts, IDs, selectively, may provide token architecture for misuses in ways to shape how AI models categorize outputs, then possible alerts, delivery-expectation or even red-teaming against those.
Also, several contemporary use cases can assist AI models become more outputs-aware, not just output-resulting. This means the possibility to prospect the likely motive or destination of the output, given the contents [by reading parallels of token architecture, conceptually].
AI Alignment?
How can AI be aligned to human values in ways that it knows what it might be used for? One angle to defining human values is what is accepted in public, or in certain spheres of the public, or at certain times. This means that AI may also be exploring the reach or extents of it outputs – given the quality, timing, destination and possible consequences.
Outputs could be an amplified focus of AI safety, using ID-keeps-and-reversal, advancing from some input-dominated red-teaming. Language translation with access to the original could become a potent tracker for what else could be ahead, for safety towards AGI.
Language is a prominent function of human memory and intentionality. Language is a core of cooperation. Language for AI is already an open risk, for unwanted possibilities with AI connivance, aside from predictions of AGI. Deepening into the language processing could have potential for AI alignment.
There is a recent analysis in The Conversation, To understand the future of AI, take a look at the failings of Google Translate,, stating that, “Machine translation (MT) has improved relentlessly in the past two decades, driven not only by tech advances but also the size and diversity of training data sets. Whereas Google Translate started by offering translations between just three languages in 2006 – English, Chinese and Arabic – today it supports 249. Yet while this may sound impressive, it’s still actually less than 4% of the world’s estimated 7,000 languages. Between a handful of those languages, like English and Spanish, translations are often flawless.
Yet even in these languages, the translator sometimes fails on idioms, place names, legal and technical terms, and various other nuances. Between many other languages, the service can help you to get the gist of a text, but often contains serious errors. The largest annual evaluation of machine translation systems – which now includes translations done by LLMs that rival those of purpose-built translation systems – bluntly concluded in 2024 that “MT is not solved yet”. Machine translation is widely used in spite of these shortcomings: as far back as 2021, the Google Translate app reached 1 billion installs.
Yet users still appear to understand that they should use such services cautiously: a 2022 survey of 1,200 people found that they mostly used machine translation in low-stakes settings, like understanding online content outside of work or study. Only about 2% of respondents’ translations involved higher stakes settings, including interacting with healthcare workers or police.”
David Stephen currently does research in conceptual brain science with focus on the electrical and chemical signals for how they mechanize the human mind with implications for mental health, disorders, neurotechnology, consciousness, learning, artificial intelligence and nurture. He was a visiting scholar in medical entomology at the University of Illinois at Urbana Champaign, IL. He did computer vision research at Rovira i Virgili University, Tarragona.
See more breaking stories here.
