Inside Digital: What are the future implications of AI language models?
By Susanne Theimer, Group Head of Digital
Can Artificial Intelligence technologies like Chat GPT-3 help us write, and what is their full potential for the future of business?
I am an easy-going person who likes to look for simple as well as smart solutions. However, writing a blog post for our Thinking section has tipped me a little out of rhythm – because I don’t like writing. It means an immense effort for me to get this creative process going. Thoughts want to be sorted and put down on paper in a logical order. The blank page demotivates, the first sentence is the most difficult. Fortunately, I know that I am not alone. There are people – colleagues and friends – who report the same experiences. We are the ones who admire you linguistically gifted people for your talent.
To you polyglot, inspiring, creative copywriters and content creators who get straight to the point as if it were the easiest thing in the world: let me tell you that many of us face almost insurmountable challenges.
Artificial Intelligence: the future answer to the blank page?
Fortunately, thanks to artificial intelligence (AI) and new technologies, plenty of visions of the future are now coming into play that could finally provide a remedy. We are talking about AI-generated content that is indistinguishable from human-made content.
What’s the story behind this? The question is whether artificial intelligence will soon be able to produce perfect online texts, such as blog posts or social media posts, with the highest possible conversion probability, as if produced on an assembly line. And AI could thus replace creative work with numbers and calculations.
What is the disruptive potential of AI text generators and language models like Chat GPT-3?
GPT-3 is a language model developed by the American organisation OpenAI and is considered a breakthrough in this technology. The autoregressive language model GPT-3 stands for Generative Pre-trained Transformer 3. It is a software equipped with artificial intelligence or a kind of artificial neural network in which information is processed mathematically. Huge volumes of training data (and more every day) combine with machine learning to produce human-like texts.
How does this artificial intelligence model work?
The GPT-3 algorithm knows which words are written most often after a certain preceding word. So it puts the word next in the text that is at the top of the list of most frequently used words. This results in well-written text that takes into account the probability of certain word sequences. In practice, this means you only have to specify a topic and choose a few criteria about the output – e.g. text length or keywords – and the AI does the rest for you. So far, so good.
While the media (including NYT Magazine) are already reporting on the remarkable language abilities of these superhuman text generators, scientists are taking a more objective and detailed look at it. Everything revolves around the extent to which artificial intelligence is really capable of calculating creativity and generating texts at the push of a button.
How far is AI from being able to replicate human creativity?
Today, AI already reproduces an incredible amount of existing data in different ways. The written texts look impressively real. This has little to do with creativity and the ability to speak or the capabilities of the human intellect to create something completely new and unique. Humans are still superior to AIs in this respect.
So what qualities do AIs still lack in order to write convincing texts for users and search engines?
1. Spark emotions
An artificial intelligence can write paragraphs and short texts in which the sentences logically build on each other. However, it can’t yet experience or demonstrate emotions and expectations or be funny. This is because it lacks the stylistic and emotional tools that make texts worth reading and relevant for us.
2. Check facts
Language models like GPT-3 do not have a deeper understanding of subject matters and their contexts. The AI evaluates templates and summarises information anew. It cannot check whether the facts from the sources are correct at all. Even if the sources are trustworthy, one cannot blindly rely on AI tools for automated content creation because information can be presented in the wrong context or even described incorrectly by the AI.
3. Develop new ideas
An AI can rewrite information in other words. However, it does not do the brainstorming for us. Nor does it deliver usable results on topics for which there are no or few comparable texts.
4. Integrate AI-created text with other media formats
Good SEO texts are not the only important thing when it comes to website ranking. Search engines also evaluate other media formats such as images, videos, sounds and third-party content.
5. Perform live multi-language translation
Taking many different languages into account is still a major challenge. GPT-3 has so far been trained mainly on English. For languages spoken by fewer people – of which there are many in Europe – the application is only possible to a limited extent.
Other major challenges being faced by AI scientists
- How can language models be created that prevent the manifestation of unconscious and potentially offensive or bias thought patterns (i.e. racism, defamation, perpetuation of traditional mentalities and stereotyped roles, conscious or unconscious prejudices)? What does an ethical charter that establishes the values developers and deployers hold onto look like?
- How can an aspect of fairness be integrated into modelling?
- We are not even talking about a European standard for AI models that takes into account European values and laws, such as copyright and GDPR, at this point.
Open source and open science are the ultimate stages of technology that win over Big Tech and profit.
Despite the aforementioned limitations of AI, we can already make good use of it in text-based applications. For instance, we spend around 28% of our working time answering emails. That’s a quarter of our work – a relatively large amount. Much of this time could be automated with speech recognition and language models.
Let’s return to the beginning of this article: I spent a weekend writing this text. In doing so, the language models could have supported me in researching topics, as well as writing the copy. They could have helped me shape the topic, with alternative formulations and with better readability. For me, it would have been a dream come true.
Where do we go from here in the pursuit of human-like AI for writing?
We have already seen historical innovations being made with the advent of email, Internet 1.0 and social media. But we can be prepared for the next big communication challenge if we keep an eye on a few key areas:
- Continue to closely monitor market developments
- Acknowledge that all these AI models stem from the immense resources of private tech companies. Wealthy for-profit research labs exert absolute control over them. This must change
- Stay in touch with industry leaders
- Follow influencers such as:
- The European LEAM (Large European AI Models) initiative
- Gopher, Chinchilla, PaLM (other large language models)
- BLOOM (BigScience Language Open-science Open-access Multilingual)
- OpenAI and other tech corporates like Meta, Google
- Hugging Face, BigScience and other institutions
- Linguistics professor Emily M. Bender
- Other stakeholders and journalists
Things are happening in the field of AI language models – even if they don’t (yet) come close to real language intelligence. Still, we can be impressed by the progress…and I am finally overcoming my fear of the blank, white sheet, and writing texts that don’t have to go through too many edits.