Foreseeing the Future with ChatGPT?

In the late winter of 2022-2023, artificial intelligence (AI) made again the buzz. This time, it is not the amazing capacity an artificial intelligence agent, Alphago by DeepMind (now Alphabet/Google), demonstrated to win alone against a go Master, but the ability to discuss and do “many things as a human” displayed by the latest version of OpenAI (Microsoft) AI model, GPT-4.

Fear and fascination are once more unleashed.

As each time disruptive technology spreads, we may complain, struggle against it, or fear it. Yet, it is highly probable (over 80%, although not certain) that AI and more specifically the “GPT-type” of AI, will generate significant changes.*

The wisest behaviour when facing such wild changes is to find out how we can make sure the technology remains a tool to serve us rather than becoming slave to or victim of the technology.

This series of articles will thus explore concretely how AI and more specifically those AI based on GPT-like-models can help us anticipating the future, notably in the fields of geopolitics, national and international security.

The first article of the series first presents what are GPT models, ChatGPT and why they matter in terms of occupation and jobs. Then, we test ChatGPT on a specific question: “the future of the war in Ukraine over the next twelve months”.

After an initial disappointment we explain, we proceeded with the test to determine if ChatGPT can assist us in identifying variables and their causal relationships for strategic foresight and early warning analysis. We provide a transcript of our conversation with ChatGPT along with our comments, and conclude that there is sufficient potential to create an AI assistant for variables named Calvin (please visit the link).

ChatGPT, GPT, Generative AI… Let’s get things right first

Why does it matter?

On 5 and 7 April 2023 Goldman Sachs in its Insights and in the special tech edition of its Briefings warned about the “sweeping changes to the global economy” that “Generative Artificial Intelligence” (see below) would bring.

“They could drive a 7% (or almost $7 trillion) increase in global GDP and lift productivity growth by 1.5 percentage points over a 10-year period…”
However, “Shifts in workflows triggered by these advances could expose the equivalent of 300 million full-time jobs to automation…”
Generative AI could raise global GDP by 7%

OpenAI, for its part, also carried out research on the impact on the labor market of its AI models and more generally of Large Language Models (LLMs) (Tyna Eloundou, Sam Manning, Pamela Mishkin, Daniel Rock, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models“, ArXiv, 17 Mar 2023 (v1), last revised 23 Mar 2023 – v4). Note, however, that OpenAI also has a corporate and legal interest in promoting LLMs and more particularly its GPTs-models as “general purpose technology.” For OpenAI, to see its GPTs-models perceived as “general purpose technology” would, for example, imply a very strong advantage in the protection of its trademark and in preventing others to use the name GPT (e.g. Connie Loizos, “‘GPT’ may be trademarked soon if OpenAI has its way“, TechCrunch, 25 April 2023).

In general, OpenAI Research found that:

“… around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs, while approximately 19% of workers may see at least 50% of their tasks impacted…
Our analysis suggests that, with access to an LLM, about 15% of all worker tasks in the US could be completed significantly faster at the same level of quality. When incorporating software and tooling built on top of LLMs, this share increases to between 47 and 56% of all tasks.”
GPTs are GPTs

Generative AI is meant to particularly impact “software, healthcare and financial services industries”, and, more largely, the media and “entertainment, education, medicine and IT industry” sectors (Ibid., Goldman Sachs Insights, “Stability AI CEO says AI will prove more disruptive than the pandemic“, 31 March 2023).

For example, in May 2023, Hollywood writers went on strike to make sure their work and pay is protected from Generative AI (Dawn Chmielewski and Lisa Richwine, “‘Plagiarism machines’: Hollywood writers and studios battle over the future of AI”, Reuters, 3 May 2023). A Californian company, Chegg, selling specialised tutoring services is seeing its shares plummet because of ChatGPT’s use by its clients (Prarthana Prakash, “Chegg’s shares tumbled nearly 50% after the edtech company said its customers are using ChatGPT instead of paying for its study tools“, Fortune, 2 May 2023).

In general, it is estimated that 900 types of occupations will be impacted, and first among them knowledge workers, scientists, and software coders (Ibid.).

As human beings concerned with the future, or more specifically as scientists and practitioners of strategic foresight and early warning, we need thrice to consider these developments.

Portal to AI – Understanding AI and Foreseeing the Future AI-powered World

First, because, being concerned with the future, we must be able to input the development and use of AI into all our foresight and warning analysis. This is why, for example, we created our initial series of articles on AI.

Second, practitioners of strategic foresight, early warning, risk management etc. are “knowledge workers” and “scientists.” We are thus on the front line of those who will be hit by Generative AI, if Goldman Sachs is right. Hence it is better to make sure AI serves us rather than kill us.

Finally, and more idealistically, if Generative AI indeed helps easing the process of anticipation in many ways, then it should also help spreading widely the use of actionable foresight and early warning. As a result, obstacles to crisis prevention and threat mitigation will be potentially reduced. However, challenges related to the organisation of the anticipation process and the willingness to heed warnings may remain largely unaffected.

What are ChatGPT, GPT-based AI and Generative Artificial Intelligence

GPT-models

GPT is a language model (Natural Language developed by OpenAI (Microsoft) that uses deep learning to generate human-like responses to prompts. The “GPT” stands for “Generative Pre-trained Transformer,” which means that the AI has been trained on vast amounts of text data to understand the nuances not only of the English language but also of other languages. GPT-models are part of the Generative AI family.

The newest model still tested by OpenAI is gpt-4. The most recent model OpenAI commercialises for usage is gpt-3.5-turbo. Waiting for gpt-4 to be available, gpt-3.5-turbo is the model we use here at RTAS for our experiment with AI assistants for Strategic Foresight and Warning: Aria (always available throughout the website, bottom right-hand corner), Calvin, Kai, Regina and Pithia.

ChatGPT

ChatGPT is an application using a GPT model that can hold a conversation with a human and generate responses that sound like they were written by a person. This is what sets the application apart and also frightens people. In March-May 2023, the public version of ChatGPT uses the model gpt-3.5-turbo. The most advanced application uses the newest model gpt-4, but is only available to paying subscribers (via ChatGPT Plus). For external usage through API, developers and users must apply through a waiting list.

Generative AI

Generative AI belongs to the Machine Learning > Deep Learning > Unsupervised, Supervised and Reinforcement Learnings category of AI (for more details on the typologies of AI, see Hélène Lavoix, “When Artificial Intelligence will Power Geopolitics – Presenting AI“, “Artificial Intelligence and Deep Learning – The New AI-World in the Making“, “Inserting Artificial Intelligence in Reality“, The Red Team Analysis Society).

GPT latest models were trained through supervised and reinforcement learning with human feedback (see OpenAI, “Aligning language models to follow instructions“).

Generative AI is thus typically based on deep learning models, which are trained on large datasets of examples in order to learn patterns and generate new content. We have, for example, generative adversarial networks (GANs – Ian Goodfellow et al. “Generative Adversarial Networks“, 2014) and variational autoencoders (VAEs – defined in 2013 by Kingma et al. and Rezende et al. – for an explanation J. Altosaar, “Tutorial – What is a variational autoencoder?“). For instance, a GAN might be trained on a dataset of images, and then used to generate new images that are similar in style and content to the original dataset, but are not exact copies.

Generative AI is thus a type of AI that is designed to create or generate new and original content, such as images, videos, music, or text. You can see below examples of images generated with OpenAI Dall-E model for this article, on three topics: deep sea security, space security and steampunk architecture.

Usage, applications and impacts

Generative AI has a wide range of potential applications, from creating realistic 3D models to generating personalised content for users. However, it also raises ethical and security concerns around the potential misuse of generated content, such as deepfakes or fake news.

The world into which an application such as ChatGPT actually evolves, as well as most GPT-based applications is digital. In other words, most GPT-based apps have no capability to act directly concretely in reality… except if human beings become their actuators (see Helene Lavoix, “Actuator for AI (1): Inserting Artificial Intelligence in Reality“, The Red Team Analysis Society, 14 January 2019).

Thus, those activities that will be most impacted by GPT-based applications and their likes are those located within the digital world. Actually, we may rephrase this as: the more insubstantial – in the meaning of not having physical existence – our activity, the more likely it will be impacted by Generative AI applications such as GPT-based apps.

For example, if you are a plumber or a mason, it is highly unlikely ChatGPT or similar GPT-based application will make a serious impact on your activity. It may help people understanding drains or masonry, and give them advice and steps to follow to do something, but the real activity of repairing a basin or removing a part of a wall to make a cupboard will still be made by a real human being and likely by a specialist.

On the contrary if your work is directly related to and carried out in the world-wide-web from web-marketing to coder and developer, your work is very likely to be heavily impacted by ChatGPT or its likes.

In the world of activities related to ideas, including knowledge, similarly, the impact of Generative AI will highly likely be important.

The logic is similar to what we identified and explained previously regarding the importance of sensors and actuators for the development of AI (see Lavoix, “Actuator for AI (1): Inserting Artificial Intelligence in Reality“, Ibid, 2019). This corresponds to what Eloundou, et al. (“GPTs are GPTs, 2023) found:

…information processing industries (4-digit NAICS) exhibit high exposure, while manufacturing, agriculture, and mining demonstrate lower exposure [to LLMs and GPTs].
GPTs are GPTs

“Occupations with the highest exposure to GPTs or GPT-powered software”, Eloundou, et al. (“GPTs are GPTs, 2023), Table 4, p.16

Indeed, OpenAI’s research paper presents a table of impacted occupations, as reproduced here.

Logically, it appears that there are also variations in the degree of exposure of “ideational” activities according to the types of skills most necessary for an activity. Science and critical thinking are the least exposed and impacted by LLMs, whilst programming and writing are part of the most exposed activities:

“Our findings indicate that the importance of science and critical thinking skills are strongly negatively associated with exposure, suggesting that occupations requiring these skills are less likely to be impacted by current LLMs. Conversely, programming and writing skills show a strong positive association with exposure, implying that occupations involving these skills are more susceptible to being influenced by LLMs.”
GPTs are GPTs

“Skills exposure to GPTs or GPT-powered software”, Eloundou, et al. (“GPTs are GPTs, 2023), Table 5, p.17

Eloundou, et al. (“GPTs are GPTs, 2023) provide a table, reproduced here, estimating the exposure of various types of mainly cognitive skills.

This will allow each of us to appraise which of our tasks can be eased by GPT-like models, or threatened by them, according to the way AI is perceived, integrated within society and used.

Our concern with the use of GPT-based applications for strategic foresight and warning is thus a case of a larger issue regarding the way Generative AI will impact knowledge and ideas-based activities.

Now, concretely, the type and extent of impact of Generative AI models and their applications will depend upon the specifics of each model and its applications in regard to specific activities. Let us thus turn practically to such a case, the use of ChatGPT for strategic foresight and warning, and the first test we did, focused on the future of the war in Ukraine.

Testing ChatGPT on the future of the war in Ukraine

Not (yet) a tool to monitor current events

The first challenge we faced was that gpt-3.5-turbo, thus the model used by ChatGPT, ended its training in September 2021. The second was that ChatGPT is a closed system, unable to access the internet or any source of information after the end of its training.

In the words of ChatGPT, asked if it could give a political assessment on current events:

I can provide political assessments on current events to the best of my abilities based on the information available up until my last training cutoff in September 2021. However, please keep in mind that my responses might not reflect the most up-to-date developments or the current state of affairs, as the world is constantly evolving.

Hence, ChatGPT and its likes cannot be used, for now, for all the steps of the strategic foresight and early warning process or risk management system that need up to date information. For example, it cannot be used for monitoring and surveilling early warning issues, nor to estimate the likelihood of scenarios or use the latter to steer policy. In terms of horizon scanning, it would need to be completed by other tools.

For those who would be disappointed, please note that OpenAI has opened its models to developers working on plugins for ChatGPT that would, for example, allow to “retrieve real-time information”. It is thus a question of time before GPT-based models and applications can be up-to-date. Indeed, the “New Bing“, Microsoft new search engine, uses GPT-4 and, de facto, accesses the world-wide-web.

Even though developments are very fast in the field, for now, our test will be done with what is available: a Generative AI which “knowledge” was cut off in September 2021.

If ChatGPT, as it is currently (Spring 2023) is not a tool we can use to monitor current evolutions, does that also mean it is useless for our purpose? Alternatively, even cut off from most recent events and knowledge, can it nonetheless be useful?

First milestones for a tool contributing to build a model for strategic foresight and early warning

Considering the methodology of strategic foresight and early warning, we shall test in which way ChatGPT and GPT-models can become tools for different aspects and parts of the analytical process.

Knowing that most serious methodologies of strategic foresight and warning, indeed any analysis, are grounded in explicit modeling, we shall examine first the capability of ChatGPT to help us with the creation and development of our model (see Hélène Lavoix, “Modeling for Dynamic Risks and Uncertainties (1) : Mapping Risk and Uncertainty“, The Red Team Analysis Society, 2018 [2011]; Joshua Epstein, Joshua M., “Why Model?“, Journal of Artificial Societies and Social Simulation 11(4)12, 2008; The Millennium Project: Futures Research Methodology,Version 3.0, 2009).

Can ChatGPT assist in identifying the crucial factors and actors relevant to our question? More precisely, can it aid in discovering the variables and causal relationships that drive the dynamics of our strategic foresight and warning issue?

We cut and pasted the conversation with ChatGPT below, and added comments in the column next to the answers.

Conclusion of the test

As it is, ChatGPT and gpt-3.5-turbo model cannot directly output a good enough model for a strategic foresight and early warning’s question on an issue pertaining to national and international security.

For this purpose, there is however a great potential in GPT and its apps, and thus probably in Generative AI.

As it is, we estimate that ChatGPT can be nonetheless very helpful to beginners and non-specialists. Iterations and asking increasingly more precise questions appear as the best systematic way to use ChatGPT for variables and causal relationships. We may also imagine the AI can help specialists if the questions asked are precisely formulated – and thus well understood by ChatGPT.

This also highlights the key importance of prompts (the questions asked to a Generative AI) when interacting with such AIs, and suggests thus a new type of activity that may emerge alongside traditional ones, for strategic foresight and warning in particular and more generally for good usage of generative AI.

We also identified other challenges impacting the identification of variables and causal relationships, such as the problem of bibliographic references, which are lacking here in ChatGPT answers. Indeed, variables and causal relationships should be grounded in facts, logic, as well as scientific research. Thus, ideally, we would need a reference for each variable and each relationship. We shall deal with this key aspect in another article. Indeed, if erroneous or plainly fake variables and links were included in models, or if variables were omitted either willingly or because o biases, then the model would become useless, or biased, with severe consequences in terms of prevention.

As a whole, we estimate that there is enough potential here to create a corresponding AI Assistant, Calvin, that our members and readers will be able to use. We intend to continue improving this AI assistant. Notably, in the future, further training of the AI for our specific purpose – called fine-tuning – could yield very interesting results. However, it seems best to wait until the full release of GPT-4 to start such a process, after carrying out a full test again with GPT-4.

Finally, assuming a Generative AI could indeed create the conceptual models we need, then one potential danger would be that practitioners would stop making the effort to think (see also ** below). We may wonder if this would lead to a decrease in cognitive capacities. In that case, it would be important to make sure that thinking continues being exerted, and that the capacities of Generative AI are properly embedded into a process that does not harm human beings.

On the positive side, the use of Generative AI could so much increase the speed of creation of a proper model that it would ease and spread the use of proper strategic foresight and warning and as a result enhance human capacities to prevent threats.

Notes

… unless other even more upsetting events take place. The specifics of the changes that will emerge from the use of AI in general, ChatGPT-like AI in particular, may also differ from our current expectations. These two points will be addressed in future articles. Here we concentrate on the highly probable and the evolution generated by ChatGPT-like AI.
*… this task is also used in workshops to open-up the minds of participants, to generate new ideas. For analysts, the very fact to make the cognitive effort to find out variables and their causal relationships enhance understanding and comprehension of an issue. Thus, being able to carry out this task alone or in a group has great value in itself. To see it being completely done by an AI-agent would is thus likely to have serious negative consequences. It will thus be key to creatively integrate the AI tool, once operational, in a way that minimise drawbacks.