AI Flywheel
The AI Flywheel describes a cycle in which an AI product gets better through its users: more users generate more data, which can be used to improve the product, which in turn attracts more users. The term originates from business and explains why leads held by AI companies often keep growing larger and larger.
A flywheel is a heavy disc that is difficult to get moving. But once it’s spinning, it keeps turning with little effort and keeps getting faster. This is exactly the image behind the term AI Flywheel. It refers to a self-reinforcing cycle found in products based on artificial intelligence. Many people use an offering, and in doing so, records are generated of what worked and what didn’t. Using these records, the company improves its offering, it becomes more attractive, and even more people use it. The cycle begins again, only at a higher level.
Why the lead grows over time
In many industries, a lead is easy to close. A competitor builds the same machine, offers the same price, and catches up right away. With a functioning flywheel, this is harder. The lead then consists not only of technology but of the accumulated experience of users. A copycat cannot buy this experience. They would first have to win millions of users themselves in order to obtain it.
For investors and journalists, the flywheel is therefore a central argument when valuing AI companies. A company whose cycle is already running is considered hard to attack. This is referred to as a moat, meaning a lasting protection against competition. Part of the high stock market valuations in the AI sector rests on the assumption that exactly this effect will occur.
The beginning, however, is the hardest part. Without users there is no data, without data no good product, without a good product no users. This starting problem is called the cold start. That’s why young AI companies often give away their products for free for a long time. In doing so, they buy themselves the first turn of the flywheel.
The four stages of the cycle
The cycle usually has four stages. First, usage: people ask questions, click, correct, abandon. Second, collection: the company stores these signals. Third, improvement: new training data emerges from the signals, which is used to retrain or fine-tune the model. Fourth, delivery: the improved version goes out to all users.
Particularly valuable are signals that contain an evaluation. When a user gives a thumbs up or thumbs down to one of two answers, this is direct information about quality. Silent behavior counts too: someone who copies an answer was apparently satisfied. Someone who rephrases the question three times was not. Such feedback costs the company almost nothing and arises automatically.
An important distinction: a flywheel is not the same as a network effect. In a network effect, users benefit directly from one another, as with a messenger app in which all your friends are present. With a flywheel, the benefit runs through the product: your usage improves the model, and everyone else benefits from that. And a flywheel can also run in reverse. If quality declines, users leave, then data is missing, then quality declines further.
From Tesla to ChatGPT
A frequently cited example is driver assistance systems. Vehicles on the road report situations in which the driver had to intervene. These rare cases are particularly valuable for training because the system is still weak there. The more cars that are on the road, the more such cases accumulate. Tesla has been using this argument for years in its communication with investors.
Chatbots like ChatGPT are also described this way. Millions of conversations per day show which answers people find helpful. In quarterly reports and analyst commentary, you’ll usually encounter the term as a justification for why a market leader should be able to maintain its position.
Skepticism is nevertheless warranted. Not every amount of data is useful, and beyond a certain point additional examples bring hardly any further improvement. Furthermore, data protection rules set limits: conversations may not always simply be used for training. So when a company argues with its flywheel, it’s worth asking whether the cycle is measurably running or merely being claimed.