Children can still beat AI at four of the five senses

Simon Rasalingham

3 days ago

8 min read

Author: Simon Rasalingham, 2026

The most frightening audience I ever faced was a bunch of kids aged between 5 and 11. My first two children were in the audience.

A week before, we had won Best Digital Health Solution at the Prix Galien UK awards at the Natural History Museum. My children's school was just up the road. Mrs Melrose, the formidable headteacher, asked me to come in and talk to the children about artificial intelligence.

I told her I was terrified. She gave me the most sage advice I have ever received on how to explain technology. "To explain something difficult to children, you have to strip it back to its basics, declutter the ideas and get it into something a child would understand."

So I started with a question. What are the five senses?

Hands shot up. See. Hear. Smell. Touch. Taste.

Then I told them I had built a computer program that was very good at one of them. It could see. It looked at X-rays and picked up things doctors sometimes miss, so they could save more patients.

I did not imagine then that what I had built would one day pick up my kids mum's lung cancer.

That talk gave me the frame I still use for every conversation about artificial intelligence and the hype around superintelligence.

AI has mastered one sense and is learning a second

Human intelligence is built on five senses. Every concept a child learns arrives through at least one of them.

AI is now very good at seeing. Vision models read X-rays, CT scans and photographs. It is getting better at hearing, through speech recognition and voice.

It is nowhere near the other three.

Can AI smell?

Not in the way a dog or a person does. The best published work is a 2023 paper in Science from Osmo and the Monell Chemical Senses Center. The model learned from 5,000 known compounds and was tested on 400 new ones. For 53% of the new molecules it predicted the panel's description better than any single one of 15 trained human panellists.

The model reads the molecule's chemical structure, as a chemist would read a diagram. It never smells anything.

Predicting a smell from a physical sensor in the room is, by Osmo's own account, still an open problem.

The UK has now put money behind that gap. In August 2026 ARIA, the Advanced Research and Invention Agency, opened Hypersensory Intelligence: Olfactory Perception.

It is a £65 million programme over four years, run under ARIA's Nature-Inspired Computation opportunity space. Its starting point is the one I gave those children: AI can describe what is in a x-ray but it still cannot smell.

ARIA's focus is around the data. Smell datasets today are small, proprietary or incompatible with each other. So up to £25 million goes on an Open Olfactory Resource, which ARIA describes as an ImageNet for smell.

The programme ends with a challenge on three tasks the teams have not seen. The bar it names is a trained dog. I very much look forward to seeing the results.

Can AI touch or taste?

Tactile sensors and electronic tongues exist in laboratories and in some robots. As far as I can find, none of them is part of the general models people mean when they say "AI". A model can describe the touch of skin. It has never felt any.

ARIA is working on touch too. Its Robot Dexterity programme, led by Jenny Read under the Adaptive Machines opportunity space, launched in 2024 and now has £88.5 million.

It starts from the point that robot bodies have not kept pace with computation. Robots still cannot match a human hand for flexibility, speed and precision. The teams it funds include electronic skins for touch. Its current call makes standards for tactile data a priority, which is the same data problem ARIA found with smell.

Taste has no ARIA programme at all.

The best definition of AGI measures sight and sound, then stops

In late 2025 a group of 33 researchers, including Yoshua Bengio, Gary Marcus and Eric Schmidt, published A Definition of AGI. They define AGI as an AI that matches the cognitive versatility and proficiency of a well-educated adult.

Their method borrows from how psychologists test people. The Cattell-Horn-Carroll theory splits human intelligence into broad abilities, and almost every major clinical IQ test is built on it. The authors pick ten of those abilities and weight each at 10%. A system that passes every test scores 100%.

By that measure GPT-4 scored 27% and GPT-5 scored 57%. The profile is jagged. GPT-5 is near perfect at reading, writing and maths. It scores nothing on storing new memories.

Redrawn from Hendrycks et al., A Definition of AGI (arXiv:2510.18212), Figure 1 and Table 1. GPT-5 tested in Auto mode.

Look at the senses in that picture. Visual processing and auditory processing are there. Smell, touch and taste are not. The authors say so openly: they leave out motor skills and tactile sensing on purpose, because they want to measure the mind rather than the quality of its sensors.

That is my point about all the hype around superintelligence from the AI labs. There is a huge blindspot and quite a lack of taste.

The hardest test of general intelligence still only uses the eyes

The best-known attempt to measure general intelligence in machines is the ARC-AGI benchmark. François Chollet introduced it in 2019 to measure how quickly a system learns a new skill, not how much it has memorised.

What does ARC-AGI-3 measure?

The third version launched in March 2026. There are no instructions. The player has to explore, work out what the goal is, and then reach it in as few moves as possible. Humans solved all 135 environments on their first go.

Think of a video game you have never seen, with no rules sheet. A screen of coloured blocks appears. You have a handful of controls. Nothing tells you what winning looks like.

A person starts pressing buttons. They notice what moves, guess at the goal and test the guess. Within a few minutes the rules click. ARC-AGI-3 asks an AI to do the same thing, game after game.

It is scored on efficiency, not only on winning. Each game counts how many moves the AI needs compared with people. An AI that stumbles on a win after thousands of random moves scores poorly. One that works the rules out as fast as a person scores 100%.

At launch, every frontier model scored below 1%. Gemini 3.1 Pro Preview managed 0.37%. GPT-5.4 scored 0.26%. Claude Opus 4.6 scored 0.25%.

Did GPT-6 Astra just pass it?

Six months on, the picture is very different. On 3 September 2026 ARC Prize reported that OpenAI's GPT-6 Astra scored 99.9% using a harness OpenAI built for its own model.

On ARC Prize's standard harness, the same model scored 62.7%.

The gap between those two numbers is the harness. A harness is the software wrapped round a model: what it remembers between moves and how it keeps track of what it has already tried, a fancy word for tasked memory. OpenAI's own harness was built for its own model. ARC Prize's standard harness is the same for everyone, which is why ARC Prize treats it as the fairer test.

ARC Prize itself says that saturating the benchmark would not prove AGI. Its games have fixed, closed rules. The real world does not.

The hardest test we have for general intelligence uses just one sense.

Medicine shows why the missing senses matter

Doctors diagnose with all five. I recently spoke with an American GI cancer surgeon about the use of robotics in his field. He said that they prefer to use the human hand, particularly when identifying smaller lesions and masses, because they can feel and touch them, which they can't do with a robot.

Radiology is the one specialty that runs almost entirely on sight. That is why AI arrived there first. It is not a coincidence.

The missing senses cannot be scraped from the internet, that data is like drawings in a cave

The internet is like drawings in caves. We've been drawing in caves for thousands of years, some people do it on walls now for a living. We have recorded events usually 'wins' rather than 'losses'.

That my worry about superintelligence it doesn't score how we learn from failure. When we see a baby trying to walk for the first time or the number of bruises from failing of your bike. The recorded memory is only the first steps or the first wobbly bike ride. It is not the a series of failures that provided you with intelligence to walk or ride and how you learnt from them.

You cannot download the feel of a swollen lymph node or the smell of an infected wound, you remember the time you missed the cancerous node or the wound that become infected. That data has to be collected by sensors in the physical world, one patient and one object at a time and all your mistakes.

In my view that is the real bottleneck to superintelligence. It is not the size of the model.

I have built AI that sees, trained on things that doctors miss as well as what they got right. I have not built one that smells, and I do not know how hard that will be. I also personally think that reinforcement learning approaches to get to superintelligence are NOT right methodology, either. It is a lot more complex than that.

Recording all five senses is a data and feedback problem too big to put a number on

A machine that could see, hear, touch, taste and smell would need a record of all five. I cannot put a number on the size of that dataset. Five senses, recorded together across every object and every patient, with a failure feedback loop for each instance, is a several different orders of magnitude problem.

That is why I have such an interest in quantum computing and what it could do for artificial intelligence. I see it as one possible route to superintelligence.

It is a not here now. I have written about it in why your 1990s desktop still beats a quantum computer. It is an exciting journey all the same.

For now I am exploring something nearer: how quantum physics can be brought into diagnostic medicine. More on that in posts to come.

Three takeaways

AI can see and is learning to hear. It cannot yet smell, touch or taste. The leading definition of AGI does not even test for them.
ARC-AGI-3 went from under 1% to 99.9% in six months on a vendor's own harness. It is still a test of seeing using a harness as a notebook.
The missing senses are a data problem. ARIA is putting £65 million into smell and £88.5 million into touch to start building that data. Recording all five may need a new kind of computing.

Which is why I am watching quantum, and why I am trying to figure out how I'm going to explain quantum entanglement to a bunch of kids too.