Development in AI systems can be generally grouped by approach:
- Systems that think exactly like humans do (‘strong AI’).
- Systems that work and achieve the goal, without figuring out how human reasoning works (‘weak AI’).
- Systems that use human reasoning as a model, but not necessarily the end goal (machine learning).
What is AI? – a flowchart (2018)
- As AI and machine learning technology evolves, more administrative and process-driven tasks will be able to be automated or partially automated, creating efficiency or freeing up people to focus on more complex and meaningful work.
- AI and machine learning driven-analysis can help organisations understand and anticipate user needs.
- AI and machine learning can enable many other revolutionary technologies.
- Automated processes can potentially be designed to be as fair or more fair than human-made decisions. For example, automated processes can be designed to only consider relevant information, to be transparent and accountable, and to deliver a decision together with the detailed reasons for the decision and the evidence taken into account.
- AI architecture is open-ended – there is no upper limit to how smart AI can become.
The majority of AI processes need to be ‘trained’, using large datasets. There is a risk that AI processes may be trained on data that is incomplete or biased. This can result in an AI process that reinforces or automate biases including sexism, racism and other discriminatory practices.
- Self-driving cars more likely to hit people with dark skin. Study (2019).
- Amazon scrapped an AI recruitment tool that it had developed to shortlist suitable candidates, as it was severely biased against women. The system had been trained on a data of CVs provided mostly by men (2018).
- Image recognition trained on biased dataset can reinforce sexist views of women (2017).
- Google photos image recognition algorithm routinely identified images of black people as ‘gorillas’ (2015).
Revolutionary solutions will cause industry disruption. Increasing automation may lead to significant job losses.
- Roundup of recent studies on job loss predictions (2018)
- Adzuna: Up to 30% of jobs in Australia at risk of automation (2018)
AI-derived solutions may lead to problematic methods for achieving the defined goal.
- A list of machine-learning solutions providing what was literally asked for, rather than what was intended.
AI could be used to create a system of interconnected and pervasive surveillance using, for example, facial recognition and video analysis, voice recognition, and monitoring of online and systems activity.
- Deloitte Insights: Pervasive intelligence (2018)
- Wired: China’s social credit system explained (2019).
The ‘creepy’ factor - users may find it intrusive if information about them is used in ways they do not expect or want.
Future of Life Institute: Benefits and Risks of AI
The NSW Government is currently exploring opportunities presented by AI, automated decision making and machine learning with a key focus on achieving ethical and equitable outcomes for the people of NSW.
Hyperanna - "Artificial intelligence powered data analyst”
Transport for NSW have been looking at delivering meaningful data analysis with the help of Hyper Anna, an AI-powered data analyst that streamlines and accelerates the journey from raw data to actional insights (Feb 2019).
Optimising NSW trains
Downer, which is responsible for maintaining NSW’s Waratah train fleet, has partnered with Microsoft to use sensors to gather data, and to use AI and machine learning to analyse that data to assess the risk of a failure, anticipate required maintenance and proactively schedule downtime required for that maintenance, increasing safety and reliability.
Bushfire prediction with CSIRO Data 61 “Spark”
Bushfires are complex processes, making it difficult to accurately predict their progress across the landscape. Spark is a software platform that uses simulation science to predict the future location of bushfires and the spread of those already burning. Spark can be used for real-time modelling of fire spread by emergency management decision makers for predicting risk, deploying firefighting resources or planning evacuation routes.
Teaching machines to assess cattle condition
Researchers at the University of Technology Sydney (UTS) have developed technology using off-the-shelf cameras to analyse cattle as they move through a crush to determine a condition score for each animal to determine if individual animals meet market specifications (2018).
Water pipe failure prediction
CSIRO Data 61 are working with more than 30 utilities from around the world to develop data-driven predictive analytics technology that accurately predicts pipe failure. Their technology equips utilities with the ability to better target repair and renewal programs, reduce operational costs of unexpected failure, and minimise the disruption to water supplies and the community.
An AI system can diagnose childhood diseases better than some doctors
In February 2019, a deep-learning system beat less experienced paediatricians at detecting illnesses including meningitis and flu. The system was trained on medical records from 1.4 million visits by 567,498 patients under 18 to a medical centre in Guangzhou, China. A team distilled this information into keywords linked to different diagnoses, and then fed these into the system to help it detect one of 55 diseases. The system managed to diagnose conditions ranging from common ailments like influenza and hand-foot-mouth disease to life-threatening conditions like meningitis with 90% to 97% accuracy. Its accuracy was compared with that of 20 paediatricians. It managed to outperform the junior ones, but more senior doctors had a higher success rate. The findings are described in a paper in Nature Medicine (February 2019).
Supergovernance: A Candian Algorithmic Assessment Tool (2018)
Supergovernance: The Government of Canada’s Algorithmic Impact Assessment: Take Two (2018)
- NSW Department of Education: Occasional Paper: Toby Walsh -The AI Revolution (2018)
- Australian Government Rural Industries Research and Development Corporation: Fact sheet - Artificial intelligence and agrifurtures (2016)
- The Partnership on AI to Benefit People and Society was established to study and formulate best practices on AI technologies, to advance the public’s understanding of AI, and to serve as an open platform for discussion and engagement about AI and its influences on people and society.
- The Gradient Institute is an Australian non-profit established by Data61 to progress the research, design, development and adoption of ethical AI systems.