Stanford University’s Human-Centered Articial Intelligence has recently published its comprehensive annual review of global AI technology, investment and adoption trends.

What can AI do now?

AI continues to advance in leaps and bounds, and in many areas is substantially more skilled than human benchmarks. However, AI seems to be plateauing in more complex tasks – which the AI Index suggests means a breakthrough now depends on an as-yet undiscovered new AI technology.

The AI outperformers are:

  • Facial recognition: In 2017, some of the top-performing facial recognition algorithms had error rates of over 50%. But by 2021, the top-performing apps registered an error rate of just 0.1%, meaning that for every 1,000 faces, the model correctly identified 999. Even in a COVID world, the top apps only performed 5 to 16% worse on masked faces compared to unmasked ones.
  • Voice to text: In 2021, for every 100 words that top-performing transcription apps heard, they correctly transcribed 99.
  • Language translation: translation apps have now reached a workably proficient level. Since 2015, there has been a 23.7% improvement in English-French translation ability and a 68.1% improvement in English-German translation ability (after finally mastering Teutonic grammar).
  • Medical AI: Medical image segmentation refers to the ability of AI systems to segment objects of interest, such as organs, lesions, or tumours in medical images. AI systems are now capable of correctly segmenting colonoscopy polyps at a rate of 94.2%, representing an 11.9 percentage point improvement since 2015.
  • Reading comprehension: In 2021, the best AI scores 95.7 on reading comprehension, exceeding the human benchmark of 91.2.
  • Deepfakes: Many AI systems can now generate fake images that are indistinguishable from real ones. In 2012, the top-performing systems could correctly identify 69.9% of deepfakes but in 2021 that number increased to 97.7%.
  • Human pose recognition: This is the task of estimating different positions of human body joints (arms, head, torso, etc.) from a single image. Human pose estimation can be used to facilitate activity recognition for purposes such as sports analytics, crowd surveillance, CGI development, virtual environment design, and transportation: for example, an AI-controlled plane identifying the body language signs of an airport runway controller. In 3 dimensional pose recognition, back in 2015, the top-performing app was making an average per joint error of 16 centimetres, half the size of a standard school ruler, but in 2021 this fell to 1.9 centimetres, less than the size of an average paper clip.
  • Sentiment recognition: This involves AI recognising the sentiment of a piece of text (very negative, negative, neutral, positive, very positive). Sentiment analysis can be straightforward if sentences are worded clearly and unambiguously, such as “I dislike winter weather.” However, sentiment analysis can become more challenging when AI systems encounter sentences with flipped structures or negations, such as “to say that disliking winter weather is not really my thing is completely inaccurate.” By 2021, the top-performing apps estimate sentiment correctly 9 out of 10 times, compared to only 7 out of 10 times back in 2015.
  • Visual task recognition: This involves AI recognising activities like walking, waving, or more complex tasks like preparing a salad (which requires an AI system to recognize and chain together discrete actions like cutting tomatoes, washing the greens, applying dressing, etc). In 2021, AI correctly recognised 89.6% of simpler tasks but this fell to 82.2% accuracy for more complex tasks (pouring wine is apparently a complex task!). However, most strikingly, in the single year of 2021, the accuracy gap between AI recognising simple and complex tasks narrowed from 27 points to only 7.4 points, which means that improvements in AI performance on harder-to-recognise tasks is occurring more rapidly than performance on easier tasks.

However, on more complex tasks, AI still has a way to go. For example, Visual Common-sense Reasoning (VCR) involves AI systems answering challenging questions about scenarios presented from images, and also providing the reasoning behind their answers. In the photo below, the AI is asked why is one guest pointing his finger at another guest when the waiter turns up:

At the end of 2021, the best mark on VCR only stood at 72. While this is a 63.6% increase in performance since 2018, it is still a long way below human performance at 85%.

Investment and R&D trends

In 2021, the greatest AI investment globally came through private investment (totalling around US$93.5 billion), followed by mergers and acquisitions (around US$72 billion), public offerings (around US$9.5 billion), and minority stake (around US$1.3 billion).

Private investment more than doubled from 2020, which was the greatest year-over-year increase since 2014. Private investment in AI in Australia was US$1.25 billion.

The size of deals also increased markedly. In 2020, there were 4 funding rounds worth $500 million or more, but in 2021, there were 15. At the same time, the number of newly funded AI companies continued to fall year on year, from 1,051 companies in 2019 and 762 companies in 2020 to 746 companies in 2021.

Over the last 5 years, most private investment globally in AI has been poured into medical and healthcare AI ($28.9 billion); followed by data management, processing, and cloud ($26.9 billion); fintech ($24.9 billion); and retail ($21.95 billion). However, in 2021, the greatest private investment in AI shifted to data management, processing, and cloud (around $12.2 billion), 2.6 times the investment in 2020.

The number of AI-related patents filed in 2021 is more than 30 times higher than in 2015, a compound annual growth rate of 76.9%. After a late start, East Asia and Pacific (mainly China) accounted for 62.1% of all patent applications. The United States files patents at one-third the rate of China. Yet despite the geo-politics, the United States and China had the greatest number of cross-country collaborations in AI research in 2021, increasing five times since 2010.

AI skills

On building an AI-skilled workforce, Australia seems to lag behind comparable economies. In 2021, New Zealand had one of the world’s highest growths in AI hiring—2.42 times greater in 2021 compared with 2016. In 2021, Australia was only 1.14 times greater than its 2016 level.

The AI skill penetration rate in the workforce means the prevalence of AI skills across occupations (based on LinkedIn member's use AI skills in their job!). India led the world in the rate of AI skill penetration—3.09 times the global average from 2015 to 2021. But Australia was only at 0.77 of the global rate.

However, in some good news for Australia, among the 15 countries listed, the AI skill penetration rates of females are higher than those of males in India, Canada, South Korea, Australia, Finland, and Switzerland.

Slow AI adoption by business

While AI technology and investment races ahead, a survey of business leaders globally (by McKinsey) found that economy-wide adoption of AI is much slower. Globally, the adoption rate only increased by 6% in 2021. Developed Asia Pacific, which includes Australia, saw adoption increase by only 3%.

Globally, the greatest adoption was in product and/or service development for high tech/telecommunications (45%), followed by service operations for financial services (40%), service operations for high tech/telecommunications (34%), and risk function for financial services (32%).

The biggest barriers to business adoption in 2021 were cybersecurity (55% of respondents), followed by regulatory compliance (48%), explainability (41%), and personal/individual privacy (41%). But concerns over cybersecurity risks fell a little from the 2020 survey.


Read more: Artificial Intelligence Index Report 2022

Expertise Area