Will ChatGPT / AI replace experts?

My interest in this question derives from many years working with large organisations internationally in both the public and private sector, who strive to improve their curation and uptake of institutional knowledge by the experts within the organisation. So, I am not thinking in terms of the general use of tools such as ChatGPT as simply a better Google, for use on the internet.

AI and Jobs

Will Artificial Intelligence (AI) tools make people redundant? 

Some people, undoubtedly, particularly if they are doing mundane desk-based jobs collating information or giving advice that is based on a body of accepted information. In some cases, the jobs may not be regarded as mundane, but may be be subject to procedures that are relatively easy to emulate.

But will ChatGPT or subsequent AI tools replace experts? It depends a little on what we mean by an expert.

Imagine a community centre where experts are paid to offer advice to people on their entitlement to benefits. One can well imagine an AI tool very quickly replacing the basic role of synthesising the information needed to help someone know their rights.

Is this enough? I don’t think so. Context is crucial.

The person’s situation may be complicated by multiple factors including complex domestic relationships, mental health problems, access to transport, and many others. The adviser will need to display emotional intelligence not only knowledge of the current state of legislation. It may be that in being able to get support in checking knowledge of the latest legislation on benefits, the adviser can spend more quality time with someone who is in desperate need of support.

In this example AI would not be replacing a human, but would be helping redefine the role of the human, to increase their true value to society.

AI as an on-demand wikipedia without editors

In the context of knowledge in general, the AI models that are emerging today, are being seen as ‘oracles’, able to dispense accurate and well sourced summary answers to complex questions.

However, it is worth noting that AI models such as ChatGPT have to feed off source material.  Who chooses the scope of this source material? If the material is of varying quality and provenance, what rules are embedded in the AI to help it decide on the level of trust it gives to different sources?

When such tools try to synthesise knowledge from public resources on the internet they confront the very real issue that much of what is written on a subject is copied; and often without being checked first. Even trusted actors can fall foul of using poorly researched information and unwittingly give it credibility by propagating it.

If you ask Google ‘when did the Titanic sink?’ you’ll get the basic answer pretty quick. An AI tool does not do much more. They will refer to the collision with an iceberg and broken compartments taking on water.

If I ask the question ‘why did the Titanic sink?’, things get more interesting, and the AI tool can do better than Google. However, it may not anticipate the 5 whys that inevitably result. Why was it going too fast, and so on.

I had to experiment with one AI tool to get the right question in order to get to the primary cause I was interested in:

What were the company pressures that influenced the captain of the Titanic to maintain high speed despite the risk of iceberg collisions?

The answer I received was:

There were allegations that Captain Edward J Smith was under pressure to maintain high speed despite the risk of iceberg collisions. It was claimed that he ordered to ‘make a record crossing’ thus indirectly causing the collision with the iceberg. However, it is also suggested that the Titanic hit an iceberg in 1912 because of a basic steering error, and only sank as fast as it did because an official persuaded the captain to continue sailing.

I still feel I would need a historian with expertise in this area to really delve into the history. Having a very smart on-demand dictionary is great, but is Wikipedia-without-editors (as some may see it) really enough? What about the intentions that lie behind a question or the specific context that it is being raised in? 

A historian might conceivably point me to diaries written by the captain long before the ship sailed, and only recently discovered – and are not fully digitised – that reveal the pressures he felt under to go as fast as possible.

AI and complex questions of science and history

Let’s try a much more difficult question. 

When was the science settled that the world is warming as a result of man-made emissions of greenhouse gases?

I got the following answer from an AI tool (a ChatGPT derivative):

The science of climate change has been studied for over a century. In the late 19th century, scientists first argued that human emissions of greenhouse gases could change Earth’s energy balance and climate1. The United Nations Environment Programme (UNEP) was formed in 1972 as a result of growing concern about the impact of human activities on the environment. The Intergovernmental Panel on Climate Change (IPCC) was established in 1988 by the World Meteorological Organization and the United Nations Environment Programme to provide scientific information on climate change. The IPCC Third Assessment Report in 2001 found “new and stronger evidence” that humanity’s emissions of greenhouse gases are the main cause of the warming seen in the second half of the 20th Century.

Not a bad answer, but not one I feel would satisfy an expert climate scientist. 

This is what Spencer Weart writes in The Discovery of Global Warming, referring to the seminal 1967 paper by Manabe and Wetherald:

This was the first time a greenhouse warming computation included enough of the essential factors, in particular the effects of water vapour, to seem plausible to experts. Wallace Broecker, who would later play a major role in climate change studies, recalled that it was the 1967 paper “that convinced me that this was a thing to worry about.” Another scientist called it “arguably the greatest climate-science paper of all time,” for it “essentially settled the debate on whether carbon dioxide causes global warming.” Experts in a 2015 poll agreed, naming it as the “most influential” of all climate change papers.

So I went back to the AI tool, to ask the question more precisely, and anticipating something closer to what I expected:

When were climate scientists finally convinced of the greenhouse gas theory for global warming on Earth.

I got the following answer:

The greenhouse gas theory for global warming on Earth was first proposed by Swedish scientist Svante Arrhenius in 1896. In his 1938 study, he showed that humans had already emitted enough CO2 into the atmosphere to increase the average temperature on Earth and that the Earth was responding: Over the previous 50 years, the average temperature had indeed increased as per Arrhenius’s proposed greenhouse effect. Many other theories of climate change were advanced, involving forces from volcanism to solar variation. In the 1960s, the evidence for the warming effect of carbon dioxide gas became increasingly convincing.

By around 2000, many climate scientists were convinced that the rise in temperature was likely to continue as greenhouse gases accumulated. Some predicted that an unprecedented global warming would become apparent.

Let’s forgive the fact that the tool confused Callendar with Arrhenius when  referring to the seminal 1938 paper (it was Callendar who first showed empirical evidence of man-made emissions causing global warming), or the suggestion that Arrhenius was the first to understand the carbon dioxide driven greenhouse effect (Tyndall certainly did, much earlier, but Arrhenius was the first to calculate its magnitude). 

The bigger surprise was that the AI tool did not cite or allude to the 1967 Manabe and Wetherald paper.

I am sure that new versions of ChatGPT and other tools will fix such problems when answering complex and nuanced questions.

AI in an organisational knowledge context

There is however a more fundamental issue with the use of such tools within organisations – whether public or private – that are aiming to become learning organisations and develop  a body of trusted knowledge.

Chinese walls, confidentiality and other factors mean we cannot simply let rip with open access to the information held in organisations.  And while quantity is useful to AI it is much  less important to an expert than the quality of the information and insight being parsed.

Let’s consider a scenario.

A multi-national consulting engineering company has done thousands of projects around the world. It partners with diverse international and local companies – experts in specific disciplines such as new materials, acoustics, carbon accounting, and much more – in design, project management and construction 

On the one hand, the consulting company wants its intellectual property respected, and in many cases, kept confidential. Clients and partners want the same for their contributions to projects. A complex Venn diagram emerges of private and shared information, and the insights (knowledge) that emerges from these experiences. Document management systems are used to apply both open access but also need-to-know policies, and often at quite a granular level.

Documents never get printed and left on trains because people who by virtue of their role need access to certain collections of information, get it – by design. Documents that are needed to be retained and never unintentionally lost, never are – by design. This is just basic content management good practice – notwithstanding the inability of Governments and many companies to apply these 20th Century capabilities effectively.

The issue for AI is that it would need to be able to navigate these complex access rights when providing answers to questions. The same question would have to give different answers to different people; even within the same organisation if chinese walls are not to be breached. This is the Achille’s heal of AI if it is to be commercialised in an institutional setting.

I am grateful to a relative (Jon Hayter) for making the following observation:

 Isaac Asimov clearly gave some serious thought to this when he wrote “I,Robot” 

At one point when the hero is speaking to a holographic version of his deceased mentor the programme gives him information but can only answer specific questions. At one point when he has made a statement based on his own thought processing the hologram says “That, is the right question”

On the other hand, the consulting organisation also wants to parade their experience and say that it uses its unique collective know-how on past projects in the conduct of new ones. This is in part through the tacit knowledge of their expert employees, as well as the codified experience within the organisation (guidelines, technique papers, anonymised project summaries, etc.) embodied in the lingua franca of knowledge: documents.

Resolving this tension between confidentiality and reuse is part of the art of working in complex organisations, and especially in consulting.

It begs a question as to the source set of information that an AI tool can or should use to answer a queries like:

We’ve been asked to design and manage the construction of a new theatre in north east China that will be a  showcase for regional Chinese culture, and an exemplar of sustainable construction and operation. What projects should we learn from and who would we ideally partner with?

Financial constraints, unrealistic expectations, political interference and resulting scope creep will be at least as important as innovative design and engineering, and all have to be factored into the answer.

Much of what is most useful as source material will be the tacit knowledge that is often not written down, and by definition, unparseable.  This is gold dust.

To counter the ‘not written down’ issue, some organisations conduct informal review interviews and workshops at the end of each project to tease out insights. For those enlightened consultancies that actually make time to do this, these reviews would aim to provide not only an overview of what was done (the what), but also why it was done that way.

Those candid reflections; those serendipitous encounters; those lightbulb moments – none of which appeared in the project file – might be scribbled in notebooks or might surface in those informal post-project reviews. Sometimes it has to wait till the exit interview or even retirement (to save the blushes)!

As things stand, only true experts can navigate the intersection between technical know-how, personal testimony, historical and current context, emotional factors, politics, deep insights, and much more, that explain the why’s and wherefore’s of key decisions on complex endeavours. 

Would ChatGPT conjure up a Sidney Opera House design out of the blue if nothing remotely similar existed beforehand?

You know the answer.

That does not mean that the AI of the future cannot play a role as an assistant in these endeavours – taking on some of the mundane tasks that exist in the curation of information and even knowledge.

For example, in the business of applying subject-specific metadata based on controlled vocabularies, AI could certainly prove a powerful assistant by making time-consuming tasks more efficient, if not quite a complete replacement for the expert knowledge curator.

However, I am confident that for the foreseeable future, it will not replace the true expert within an organisation.

Your job is safe.

(c) Richard W. Erskine, May 2023