Artificial intelligence has changed form in recent years.
What started in the public eye as a burgeoning field with promising (yet largely benign) applications, has snowballed into a more than US $100 billion industry where the heavy hitters – Microsoft, Google, and OpenAI, to name a few – seem intent on out-competing one another.
The result has been increasingly sophisticated large language models, often released in haste and without adequate testing and oversight.
There’s no doubt AI systems appear to be “intelligent” to some extent. But could they ever be as intelligent as humans?
There’s a term for this: artificial general intelligence (AGI). Although it’s a broad concept, for simplicity you can think of AGI as the point at which AI acquires human-like generalized cognitive capabilities. In other words, it’s the point where AI can tackle any intellectual task a human can.
AGI isn’t here yet; current AI models are held back by a lack of certain human traits such as true creativity and emotional awareness.
We asked five experts if they think AI will ever reach AGI, and five out of five said yes.
But there are subtle differences in how they approach the question. From their responses, more questions emerge. When might we achieve AGI? Will it go on to surpass humans? And what constitutes “intelligence”, anyway?
Here are their detailed responses:
AI and Philosophy of Technology
AI has already achieved and surpassed human intelligence in many tasks. It can beat us at strategy games such as Go, chess, StarCraft and Diplomacy, outperform us on many language performance benchmarks, and write passable undergraduate university essays.
Of course, it can also make things up, or “hallucinate”, and get things wrong – but so can humans (although not in the same ways).
Given a long enough timescale, it seems likely AI will achieve AGI, or “human-level intelligence”. That is, it will have achieved proficiency across enough of the interconnected domains of intelligence humans possess. Still, some may worry that – despite AI achievements so far – AI will not really be “intelligent” because it doesn’t (or can’t) understand what it’s doing, since it isn’t conscious.
However, the rise of AI suggests we can have intelligence without consciousness, because intelligence can be understood in functional terms. An intelligent entity can do intelligent things such as learn, reason, write essays, or use tools.
The AIs we create may never have consciousness, but they are increasingly able to do intelligent things. In some cases, they already do them at a level beyond us, which is a trend that will likely continue.
Computational Neuroscience and Biomedical Engineering
AI will achieve human-level intelligence, but perhaps not anytime soon. Human-level intelligence allows us to reason, solve problems and make decisions. It requires many cognitive abilities including adaptability, social intelligence and learning from experience.
AI already ticks many of these boxes. What’s left is for AI models to learn inherent human traits such as critical reasoning, and understanding what emotion is and which events might prompt it.
As humans, we learn and experience these traits from the moment we’re born. Our first experience of “happiness” is too early for us to even remember. We also learn critical reasoning and emotional regulation throughout childhood, and develop a sense of our “emotions” as we interact with and experience the world around us. Importantly, it can take many years for the human brain to develop such intelligence.
AI hasn’t acquired these capabilities yet. But if humans can learn these traits, AI probably can too – and maybe at an even faster rate. We are still discovering how AI models should be built, trained, and interacted with in order to develop such traits in them. Really, the big question is not if AI will achieve human-level intelligence, but when – and how.
AI and Swarm Intelligence
I believe AI will surpass human intelligence. Why? The past offers insights we can’t ignore. A lot of people believed tasks such as playing computer games, image recognition and content creation (among others) could only be done by humans – but technological advancement proved otherwise.
Today the rapid advancement and adoption of AI algorithms, in conjunction with an abundance of data and computational resources, has led to a level of intelligence and automation previously unimaginable. If we follow the same trajectory, having more generalised AI is no longer a possibility, but a certainty of the future.
It is just a matter of time. AI has advanced significantly, but not yet in tasks requiring intuition, empathy and creativity, for example. But breakthroughs in algorithms will allow this.
Moreover, once AI systems achieve such human-like cognitive abilities, there will be a snowball effect and AI systems will be able to improve themselves with minimal to no human involvement. This kind of “automation of intelligence” will profoundly change the world.
Artificial general intelligence remains a significant challenge, and there are ethical and societal implications that must be addressed very carefully as we continue to advance towards it.
AI and Data Science
Yes, AI is going to get as smart as humans in many ways – but exactly how smart it gets will be decided largely by advancements in quantum computing.
Human intelligence isn’t as simple as knowing facts. It has several aspects such as creativity, emotional intelligence and intuition, which current AI models can mimic, but can’t match. That said, AI has advanced massively and this trend will continue.
Current models are limited by relatively small and biased training datasets, as well as limited computational power. The emergence of quantum computing will transform AI’s capabilities. With quantum-enhanced AI, we’ll be able to feed AI models multiple massive datasets that are comparable to humans’ natural multi-modal data collection achieved through interacting with the world. These models will be able to maintain fast and accurate analyses.
Having an advanced version of continual learning should lead to the development of highly sophisticated AI systems which, after a certain point, will be able to improve themselves without human input.
As such, AI algorithms running on stable quantum computers have a high chance of reaching something similar to generalised human intelligence – even if they don’t necessarily match every aspect of human intelligence as we know it.
Machine Learning and AI Alignment
I think it’s likely AGI will one day become a reality, although the timeline remains highly uncertain. If AGI is developed, then surpassing human-level intelligence seems inevitable.
Humans themselves are proof that highly flexible and adaptable intelligence is allowed by the laws of physics. There’s no fundamental reason we should believe that machines are, in principle, incapable of performing the computations necessary to achieve human-like problem solving abilities.
Furthermore, AI has distinct advantages over humans, such as better speed and memory capacity, fewer physical constraints, and the potential for more rationality and recursive self-improvement. As computational power grows, AI systems will eventually surpass the human brain’s computational capacity.
Our primary challenge then is to gain a better understanding of intelligence itself, and knowledge on how to build AGI. Present-day AI systems have many limitations and are nowhere near being able to master the different domains that would characterise AGI. The path to AGI will likely require unpredictable breakthroughs and innovations.
The median predicted date for AGI on Metaculus, a well-regarded forecasting platform, is 2032. To me, this seems too optimistic. A 2022 expert survey estimated a 50 percent chance of us achieving human-level AI by 2059. I find this plausible.