Rabbits and AI: When Innovation Multiplies Without Limits

Twenty-four rabbits

In 1859 twenty-four rabbits were released near Geelong in the Australian state of Victoria for hunting. They reached New South Wales by 1880, South Australia and Queensland in 1886 and eastern Western Australia around 1890, and by 1910 they occupied most of their present range. When the myxomatosis virus was released in 1950 it initially cut their numbers by about 95%, but their resistance to it then grew. The rabbits had found open country and plentiful food.

This invites a comparison with what we see in AI: applications born of ideas, and ideas breeding from applications. Is the comparison sound, and what does it hide?

What supplied the "food"?

The numbers support part of the story. Stanford HAI's 2025 AI Index reports that the inference cost of a GPT-3.5-level model "dropped over 280-fold between November 2022 and October 2024", that hardware costs fall by about 30% a year and its energy efficiency improves by about 40% a year. It also shows that 78% of organisations reported using AI in 2024, up from 55% a year earlier. On the training side, Epoch AI estimates that the training compute of frontier language models has doubled roughly every 5.2 months since 2020.

The 2026 edition adds that generative AI reached 53% of the population within three years, faster than the personal computer or the internet at the same stage. When costs fall and access widens, ideas multiply because trying them has become cheap.

Where the comparison fails

A rabbit neither decides nor answers for its actions; an application was built by people, who are accountable for it. Nor is every multiplication a flourishing: many applications do not mean many useful ones. The same report records a high of 362 documented AI incidents in 2025. It is a reminder that diffusion usually outruns the building of safeguards, whose foundations we set out in the article "Responsible Artificial Intelligence: Four Principles for Governance and Trust".

One part of the comparison does hold: ecosystems select. Crowded solutions mean that some fail or merge while others survive because they meet a real need. That selection improves if we take part in it with clear measures instead of leaving the noise to decide.

What to do

Before adopting any new application, an individual or an organisation can ask four simple questions:

1. What specific problem does it solve, and would a simpler tool solve it?
2. On what data was it tested, and what is its error rate in my case?
3. Who is responsible if it errs, and where does my data go? (See the article "Privacy in the Age of AI: How Your Secrets Are Inferred, and How to Protect Them".)
4. What is my plan if it stops working or its terms change?

Caution should not turn into paralysis, though. Lower costs are a real opportunity for small organisations and charities, provided they begin with a limited, measurable trial.

Caveats on the figures

These are figures about averages and trends; they do not tell us the quality of any single application. A fall in inference cost does not mean training has become cheaper: Epoch estimates that the cost of training frontier models rises about 3.5-fold a year. What tomorrow holds is in God's hands, and the honest course is to describe the present trend without stretching it in a straight line beyond what we know.

The rabbit story teaches that a creature finding open ground spreads faster than anyone expects, and that wisdom lies in asking beforehand what we want from this capability and what limits we set for it.