Trust Score

A trust score is a number that expresses how trustworthy a computer system considers someone or something to be. What gets assessed can be people, companies, individual posts, or even devices. The scale typically ranges from 0 to 100 or from 0 to 1. A high value means: little speaks against you here. A low value means: there are irregularities here, so take a closer look. The number does not arise from a human judgment, but from a computational rule that combines many individual observations.

Why trust becomes a number

Large platforms cannot review every case individually. On a marketplace, millions of listings are created every day; on social networks, billions of posts. Humans could never review this volume. A number, on the other hand, can be compared instantly against a threshold. Everything below the threshold gets blocked or forwarded to a human, everything above it goes through without friction.

The second reason is money. Fraud costs payment providers and online retailers billions every year. Anyone who detects fraudulent orders early saves themselves chargebacks and legal disputes. At the same time, the check must not be too strict: every honest customer who is wrongly rejected represents lost revenue. The trust score is the dial between these two types of errors.

It’s important to distinguish this from a credit bureau score or a credit score. Those concern the question of whether someone can pay a bill. A trust score concerns the question of whether someone acts honestly. Both are point values, but they measure completely different things.

What the score is made up of

The basis consists of signals, meaning individual observable facts. For a seller account, these include the age of the account, the number of completed sales, complaint rates, and refunds. For a payment, technical signals are added: the device used, the approximate location, the time of day, the typing speed in the form. On its own, none of these signals says much. Only their combination produces a pattern.

In simple cases, the calculation is handled by a fixed formula with weights. More often today, a trained model does the work, meaning a program that has learned from many past cases. It is shown millions of completed transactions along with the known outcome, i.e., fraud or no fraud. The model itself searches for the features that typically occur together. It then fits new cases into this pattern and outputs a probability, which is displayed as the score.

A common misconception is that the score is an objective measurement. It is an estimate based on the past. If the training data was skewed, the model adopts that skew. If, for example, orders from certain regions were checked especially often in the past, more detected fraud cases will show up there. The model then considers the region suspicious, even though it was simply looked at more closely. Such feedback loops are the most difficult part of working on such systems.

Trust scores in everyday life and in the news

Trust scores are most visible where strangers do business with one another. Booking platforms, classified ad sites, and ride-hailing services display seller or driver ratings, behind which lies an internal value. They work invisibly in every online purchase: whether a card payment goes through immediately, requires an additional confirmation via app, or is rejected outright is often decided by such a value within milliseconds.

In media coverage, trust scores mainly come up in two contexts. First, in dealing with misinformation, where platforms rate sources and accounts in order to control reach. Second, in regulation: the European AI Act requires traceability and the possibility of human review for systems that decide on people’s access to services.

As a user, you rarely get to see your own value. It becomes noticeable indirectly, for instance when the option to buy on invoice suddenly disappears, or an account gets blocked after a move abroad. In many cases, there is a right to access the data being processed. However, providers are reluctant to disclose the exact computational rule, since doing so would make it possible to deliberately game the system.