Paul Ford

Home » Blog » Quantum computing and AI

Quantum computing and AI

Last updated:

Could a quantum computer make artificial intelligence much more intelligent? It might one day make parts of AI training faster, but speed alone would not give an AI better judgement, better ideas or more reliable answers.

Today’s AI is already helping engineers control quantum computers and correct their errors. For now, AI may do more for quantum computing than quantum computing does for AI.

Quantum computers are specialist machines

An ordinary computer stores information in bits. Each bit is read as a zero or a one. A quantum computer uses qubits. A qubit is also a unit of information, but it obeys the rules of quantum physics.

Those rules let a quantum program handle some calculations in a different way. For the right problem, it can make some possible results more likely and others less likely before the answer is read. This can cut the amount of calculation needed.

That limit is crucial. A quantum computer does not make every calculation faster. It is being developed for a small group of tasks, such as modelling molecules, testing vast sets of possibilities and solving a few types of mathematical problem. Ordinary computers will still run email, websites, spreadsheets and the rest of our software.

The powerful quantum machines needed for some of the promised tasks do not yet exist. Qubits are extremely delicate. Small changes in heat, vibration or electrical conditions can spoil the calculation. This is what researchers mean by noise. It is not sound.

Engineers can correct some of these errors by using a group of physical qubits to make one more dependable unit, called a logical qubit. That is hard to do at scale. IBM says it aims to build a machine with 200 logical qubits by 2029. The figure describes a planned machine. It does not describe a general quantum computer available today, or prove that the machine will improve AI.

Faster calculation does not automatically produce better AI

Training an AI involves a great deal of arithmetic. Text, pictures and sound are turned into numbers. The computer makes the same kinds of calculation again and again while the model adjusts itself. Graphics processing units became important to AI because they are good at doing many of these small calculations at the same time.

Some proposed quantum methods could make certain parts of that arithmetic much faster. Three problems stand between that claim and a faster AI system.

The first is getting the information into the quantum computer. Text, pictures and sound are stored in the form used by ordinary computers. Before a quantum processor can use them, that information has to be prepared in a quantum form. If the preparation takes longer than the calculation saves, the whole process is no faster. A 2024 technical study begins with this problem and finds ways to reduce the cost for some well-ordered sets of data.

The second problem is training the quantum part of the system. The program has settings that must be adjusted until the answer improves. As the system grows, the signal saying “that change helped” can become so faint that the training process has no clear direction. Researchers call these flat regions barren plateaus. The small physical disturbances that cause errors in the hardware can make the problem worse.

The third problem is the pace of ordinary computing. A quantum method may look far ahead because it has been compared with a poor ordinary method. Then someone finds a better one. In 2018, Ewin Tang found a fast way for an ordinary computer to copy the main benefit of a proposed quantum recommendation system. A huge claimed lead for the quantum method disappeared.

Even if all three problems were solved, the result might only be cheaper or faster training. It might let a company build a larger model or try the same training process more times. The quality of the AI would still depend on how the model was built, what it was asked to learn, the information it was given and how its mistakes were corrected.

More computing power can improve an AI. It cannot decide what the AI should learn or what counts as a good answer.

Quantum computers may be better at studying quantum systems

The case is clearer when the subject being studied is itself quantum. A molecule, a new material or an experiment in quantum physics produces information that follows quantum rules. A quantum processor may be able to study that information before it is converted into the ordinary numbers stored by a normal computer.

In 2022, a team including Google Quantum AI researchers tested this idea on Google’s Sycamore processor. For the tasks in the paper, the quantum machine learned about another quantum system from far fewer experiments than the ordinary method used for comparison. It could do this because the information started in quantum form and was kept in that form while several copies were examined together.

This finding applies to learning about quantum systems. It does not show that a quantum processor can improve a language model trained on text, an image generator trained on pictures or a business system trained on sales records.

Quantum computers may still help chemistry and materials science. Some molecules and materials are very hard to model on an ordinary computer because they follow quantum rules. A future system might use a quantum processor to model the molecule, AI to choose the next test and an ordinary computer to control the process and store the results.

Researchers have also shown an advantage on other tightly defined learning tasks. These experiments prove that a quantum machine can beat an ordinary one in selected cases. They do not tell us whether the advantage will remain when the full process is run on dependable hardware at a practical scale.

AI is already helping quantum computers

A quantum computer needs constant control. The exact settings can change while it runs, and small physical disturbances create errors. An ordinary AI system can watch what is happening, spot changes and adjust the controls.

In 2026, Google Quantum AI used reinforcement learning to adjust the error-correction controls on its Willow processor. The AI ran on an ordinary computer connected to Willow. As conditions changed, it helped adjust the quantum machine and keep its errors under control.

A future computer may contain all three types of processor. Its main processor would run the software and control the system. Graphics processors would handle the calculations that suit AI. A quantum processor would be used only when a particular calculation suited it.

How to test a claim about quantum AI

When a company or research team announces a quantum AI result, four questions help establish what has been achieved:

  • What task did the quantum processor complete?
  • Did it beat the best ordinary method for the same task?
  • Does the comparison include the time and cost of preparing the data, correcting errors and running the test enough times to trust the answer?
  • What improved across the complete process: time, cost, accuracy or scientific knowledge?

If those questions are not answered, the claim may describe a laboratory result rather than a practical advance. For now, quantum hardware has beaten ordinary computers on selected research tasks, while ordinary AI is already being used to help keep quantum hardware under control.

Sources and further reading