Provided by Prof. Jon Whittle, Director CSIRO’s Data61
There are roughly two flavours of AI. Symbolic AI requires a human to write down a set of “rules” that the AI searches through to make a decision. Data-driven AI instead looks for patterns in large datasets – in essence, it figures out the rules for itself.
The vast majority of AI in industry today is data-driven AI. By definition, it needs plenty of quality data about your business to work.
Generative AI is a form of data-driven AI that creates content – text, images, video. Generative AI has become popular since the launch of ChatGPT in November 2022. ChatGPT is also a form of general-purpose AI as it can carry out a wide variety of tasks. In contrast, narrow AI is designed for a very specific task, and only works for that particular task.
AI has been used in industry for decades. Netflix has been using AI to suggest what you watch for over two decades. An early form of symbolic AI, Google Maps was invented (in Sydney!) in 2005. I was in the control room when the first AI software was used to command a spacecraft in 1998! In 2023, McKinsey said that 55% of organisations have implemented AI in at least one business unit or function [1].
Narrow AI is well embedded in industries such as automotive, manufacturing, and mining. Generative AI has impacted education, marketing and the creative industries. Some experts believe that most opportunities lie in healthcare (although health systems can be slow to adopt technology), the legal profession (one in two lawyers already use AI according to the AFR [2]), or banking (although there are issues around sensitive data).
Publicly-available generative AI tools like ChatGPT and Gemini can be used “out of the box” for a variety of business tasks, such as crafting emails, summarising documents or even writing code. More sophisticated tasks will require a bit more work. Some companies – such as KPMG and Bloomberg – have trained generative AI tools to work with their own internal documents.
Generative AI can be good at automating repetitive tasks, although it usually requires human oversight. More generally, if you have plenty of historical data about a particular business process or problem, you could potentially use narrow AI to solve a particular problem. But for any AI application, be aware that AI never produces right answers 100% of the time. AI is based on statistics so can get things wrong.
Maybe. It’s still early days in terms of scientific evidence on AI productivity [3]. A meta-review by Microsoft showed productivity increases of between 26% and 73% from its own Copilot tool. Harvard Business School showed a 12% increase in productivity for consultants using AI. And the National Bureau of Economic Research reported call centre agents could handle 14% more calls with AI.
However, these are isolated studies and it’s not clear if productivity increases translate at an enterprise level. It’s well known that productivity increases in one part of a business can lead to decreases elsewhere [4]. And Solow’s paradox [5] reminds us that, over the last few decades, the promises of digital transformation efforts often do not materialise in practice.
The most well-known are biases (where an AI is trained on limited demographic data and so discriminates against people outside that demographic) and hallucinations (where generative AI makes up facts – research indicates that ChatGPT fabricates unverifiable information in approximately 20% of its responses [3]). These can lead to significant reputational and financial risk for a company. For example, an Air Canada chatbot incorrectly offered a customer discount but was forced by a tribunal to honour it.
Good AI governance within a business is crucial to understanding and mitigating any potential risks. AI regulation is also coming – the EU has already passed an AI Act and regulation is actively being considered by the Australian government.
While many AI tools are free, the more advanced versions typically require a subscription model. These can add up to a significant investment across a large organisation. Given that the estimates for OpenAI to train its GPT-4 AI are $78M US [3], AI vendors will need to recoup costs somehow. For more specific AI applications in an enterprise, there will be additional costs: to collect and curate the data to train an AI, ongoing maintenance costs, consultancy and/or in-house AI expertise. These costs can be significant. On the other hand, McKinsey says that 42% of organisations report cost reductions overall [3].
The environmental cost of AI is somewhat under-reported. The datacentres that power AI require large amounts of electricity (not exclusively renewable) and water. Carbon emissions of generative AI systems come from training the AI model as well as use of the model. Meta’s Llama2 is estimated to have emitted 291 tonnes of carbon during training (compared to 1 tonne for a roundtrip flight from New York to San Francisco) [3]. By some estimates, a generative AI search requires ten times more electricity than a standard search. AI can also reduce environmental impact – AI is being used to reduce waste and to make cooling more efficient.
More generally, aspects of AI are starting to be considered in ESG frameworks. CSIRO and Alphinity Investment Management recently released an approach to evaluating ESG impacts of AI [7]. Early evidence suggests that those companies with a strong ESG track record are more likely to implement AI responsibly.
My advice to business looking to adopt AI is threefold: (i) Focus on the problem you are trying to solve rather than introducing AI for its own sake; (ii) AI is not appropriate for all problems – for example, AI will never achieve 100% accuracy because it is a statistical technique; (iii) Realise that AI takes effort to succeed – you need well-curated and sufficiently large datasets, AI technical expertise, and good AI governance practices.
For a general introduction to AI and its applications across a range of industries, check out my Everyday AI podcast, available on all streaming platforms. CSIRO’s National AI Centre provides resources about best-practice adoption of AI, safe and responsible AI practices, as well as practical support on how to get started.
[1] Chui, M., Yee, L., Hall, B., Singla, A. & Sukharevsky, A. (2023). The State of AI in 2023: Generative AI’s Breakout Year. McKinsey & Company. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023generative-ais-breakout-year#widespreadhttps://ceros.mckinsey.com/commentary-ai-2023-lareina-ye-desktop.
[2] ‘Get the job done’: One in two lawyers use AI, Euan Black, Australian Financial Review, Apr 16, 2024
[3] Nestor Maslej, Loredana Fattorini, Raymond Perrault, Vanessa Parli, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, and Jack Clark, “The AI Index 2024 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April 2024.
[4] Why CSCW applications fail: problems in the design and evaluation of organization of organizational interfaces. J. Grudin. CSCW '88: Proceedings of the 1988 ACM conference on Computer-supported cooperative work, page 85--93. New York, NY, USA, ACM, (1988)
[5] Bäck, Asta & Hajikhani, Arash & Jäger, Angela & Schubert, Torben & Suominen, Arho, 2022. "Return of the Solow-paradox in AI? AI-adoption and firm productivity," Papers in Innovation Studies 2022/1, Lund University, CIRCLE - Centre for Innovation Research
[6] Alex de Vries, The growing energy footprint of artificial intelligence, Joule, Volume 7, Issue 10, 2023, pages 2191-2194.
[7] Alphinity Investment Management & Commonwealth Scientific and Industrial Research Organisation (CSIRO), The intersection of Responsible AI and ESG: A Framework for Investors, CSIRO, 2024.
