You might ask what Star Trek, Norway and LLM’s have in common. Rob Price weaves a path through his week of conversations and travel to explore the question of bias in AI technology, a key part of the broader conversation around #DigitalResponsibility.
It was a pleasure to travel to Oslo, in Norway, this week at the invite of the GoForIT (linked to IKT-Norge, and including organisations such as Bouvet, Microsoft and Sopra Steria) team, who have been looking at CDR as a relevant topic area to explore for the Norwegian IT industry. With the help of Leif, Pia and Ida, they have translated the International CDR Manifesto into Digitalt Samfunnsansvar, in essence Digital Social Responsibility. This was felt to be more aligned to the collaborative group of public and private sector, plus academia organisations involved in the association as the “Corporate” tag could be seen to be more relevant to business, rather than – as intended – all organisations. I have some sympathies with this perspective as I have also seen Public Sector organisations question if CDR is right for them due to the “Corporate” label.
It was fantastic to hear about the challenges and opportunities for the tech industry in Norway. Some of the trends were indeed common (to the UK or Europe) such as ageing populations and consequential strains placed on Health and Social Care. Some were specific, especially in this regard the access to volume of affordable talent. With a 5.2 million population, perhaps too few universities/colleges delivering technology/digital/data courses and a greater challenge to source further afield in near shore/off shore models given cultural and language challenges, then there is a strong need for more support in delivering transformational change in those industries that will be under increasing pressure to evolve – for example, Oil and Gas and the shift to renewables or Fishing, and the shift to the Circular Economy and sustainable practices.
You can find the GoForIT translation here and I look forward to seeing how it develops and resonates with private, public and academic organisations in Norway over the coming months.
Having returned to the UK, I then had a dinner with AutogenAI, a UK based business focusing on the application of Large Language Models (LLM) on Enterprise Sales & Bidding. I had already seen a demo, which looked fascinating. In essence, AutoGenAI build on the core content with your own Enterprise content (previous bids, reports, methods, collateral, marketing) and provide a structured process to pull together an answer to the question you prompt. Over the years, I have yet to see an effective knowledge management system/process in an enterprise organisation so have always favoured an effective search approach. In principle I like it, and the way in which the user is guided through options means that this is firmly in the intelligent accelerator to help the author do a great job in a shorter amount of time, rather than simply replacing the role of a bid author. It seems – and indeed, we will continue to check – a good and effective application of the technology.
Of course there are questions. Some are obvious and much written about. For example, questions in regard to the environmental impact of the increased energy use, or the debates around interpretation of what is ‘right’. In that context we talked about hallucinating, where the LLM in essence makes things up, seemingly presented as fact – which I’ve already personally experienced with ChatGPT – and debated the ease of recognising and indeed writing sarcasm.
However, I was quickly drawn back to my conversations in Norway when we talked about the content with which the LLM was fed. We discussed the type of content (internet, books, Wikipedia, academic papers etc.) and we also discussed the language of that content (lots of English) that had been ‘learned’. It made me think what would be the consequential impact of highly effective and mass adopted LLMs that had been fed the global content of knowledge. I don’t have access to the split of written knowledge by language – I would love to know that if someone does have it – but even if we just look at the top spoken languages (first or second language) of the world, we can see that English (16.5%), Mandarin (14.6%), Hindi (8.3%) and Spanish (7.0%) dominate. So if we are feeding LLMs with lots of content that is in those languages, how do we ensure that there is not massive bias against those languages with fewer speakers? Norwegian for example, or as we also discussed on the night, Aboriginal languages. On one hand, you would hope it was a way to preserve those languages rapidly being lost, according to https://languageconservancy.org, currently 9 languages per year cease to be spoken, expected to rise to 16 by 2080. On the other, you can foresee that the more English and Chinese are using AI (LLM or otherwise), the more they can effectively compete and dominate versus those that have less content to learn and inform the LLM (noting that language is also representing cultures and behaviours as well as know how and beliefs – so simply translating English content is not maintaining identity).
As is the way in 2023, I decided to ask ChatGPT for guidance.
As large language models (LLMs) are primarily based on English content, niche languages such as Norwegian may face several challenges. The ultimate impact may be a reduction in the representation of Norwegian language, culture, and expressions, leading to a biased and limited understanding of the language.
Over time, to ensure the adequate representation of languages such as Norwegian, a few suggestions can be made:
1. Creating more data in Norwegian: It is essential to create more Norwegian content to train LLMs. This can be achieved by encouraging content creators, language enthusiasts, and technology companies to create more Norwegian text, speech, and other digital data.
2. Collaborating with diverse communities: LLMs can be biased when trained on a limited dataset. Therefore, it is essential to collaborate with diverse communities, including different age groups, genders, and cultures, to create more comprehensive and inclusive Norwegian data.
3. Promoting multilingualism: Instead of solely relying on English, promoting multilingualism can help improve the representation of niche languages. This includes training LLMs in multiple languages, promoting the use of multiple languages in digital media, and encouraging individuals to learn and use different languages.
4. Developing specific Norwegian LLMs: Instead of relying on general-purpose LLMs, developing Norwegian-specific LLMs can help improve the language’s representation. This can be achieved by training LLMs on Norwegian data, making it more appropriate for Norwegian speakers and their expressions.
5. Regularly auditing LLMs: Regularly auditing LLMs can help identify biases and improve the representation of languages such as Norwegian. This includes auditing the data used to train LLMs, the algorithms used to develop them, and the outputs generated by them.
In summary, ensuring adequate representation of niche languages such as Norwegian requires creating more data, collaborating with diverse communities, promoting multilingualism, developing Norwegian-specific LLMs, and regularly auditing LLMs. These steps can help reduce language bias and ensure the preservation and growth of Norwegian language and culture.
OK, so first point to note. It seems ChatGPT agrees with me, which is reassuring – yet clearly not reassuring. Secondly, its primary suggestion is to create more Norwegian content. Excellent, I will ensure that the entirety of Norway is immediately set to work creating a volume of work to parallel that (in volume) of English and Chinese. In parallel, I will also get those other nations to do the same. No doubt in 2000 years, assuming no further works are produced in English and Chinese, we will have caught up!!
You might detect my sceptism. And sarcasm. But what is the answer? Are we on an inevitable track towards homogenisation and assimilation?
Are we in fact, ultimately, Borg?
Don’t get me wrong, this is not intended as an anti-AI or LLM rant. It isn’t. I am genuinely interesting in AutogenAI and their solution. It is undoubtedly an accelerator for a specific and clear use case in business – and I have seen first hand the long days and nights and weekends of stressful bidding cycles. This would help – not replace – those to deliver an effective response more quickly, and likely pulling on their organisational know how in totality rather than what they happen to find.
Indeed that is a specific example of the Intelligent Automation that I have spoken about and written about previously, helping employees to do their role more effectively in the available time and resource constraints. In a time where we cannot easily find additional people needed in many businesses – public and private – it will inevitably be a hugely valuable tool that helps accelerate and inform positively.
As ever, I just remain interested in the governance and controls that we might need to put in place, combined with an awareness of the unintended consequences that could arise – because thy are undoubtedly there too. This is #DigitalResponsibility. It is still critical (more than ever) to innovate using Digital, but to do so in an informed, governed and conscious (and yes, effectively regulated) way – trying to help people, communities and societies and to minimise the harms or negative effects.
But let us try to do so whilst maintaining the diversity and culture that humanity has evolved. And in doing so, avoid traversing the Galaxy in a Borg Cube.
