Delphi method

Some questions can’t simply be calculated. When will self-driving cars be ready for mass production? How many people will work from home in 2035? For such questions, one asks experts for their assessment. The Delphi method is an orderly way of doing this: a group of experts is questioned several times in succession, each answering individually and without giving their name. Between rounds, everyone learns what the group as a whole answered, and may change their own answer. The name goes back to the ancient Oracle of Delphi, which people in Greece consulted about the future.

Why you don’t just call a conference

In a normal discussion round, it’s often not the best argument that wins out. What decides is who speaks the loudest, who holds the highest title, or who speaks first. Anyone with a dissenting opinion tends to stay quiet. Experts call this group pressure. The Delphi method switches off this effect by ensuring no one knows who an answer came from.

At the same time, one doesn’t want to lose the exchange entirely. A single anonymous survey only delivers a snapshot of opinion, but no debate. The feedback between rounds is the core of the method. Anyone who gave an unusual figure must justify it, and everyone else reads that justification. This way the argument remains visible, but not the person.

An important distinction should be made here: the Delphi method is not opinion polling. One doesn’t survey a thousand random people, but rather twenty to fifty selected experts. The result is therefore not a statistic about the population, but rather the most thoroughly vetted expert judgment possible.

The process across several rounds

At the start there is a clearly formulated question, usually with a number or a year as the answer. A team selects the participants and sends out the questionnaire. Each person answers separately, often with a brief justification. The participants do not learn who else is taking part.

Then the team summarizes the answers. Typical is the mean or the median, i.e. the value exactly in the middle of all answers. Added to this are the most striking justifications from the outlying ranges. This package goes back to everyone, and the second round begins with the same question.

This is usually repeated two to four times. The answers often move closer together in the process, because arguments become convincing or misunderstandings disappear. If two camps remain, that too is a result: it shows that the expert community is divided. A common misconception is that the goal of the method is agreement at any cost. A forced compromise would be worthless.

Delphi in technology forecasts and AI debates

The method was developed in the 1950s at the US think tank RAND, originally for military forecasts. Today it is used by ministries, companies, and research institutes. When a government wants to know which technologies will become important in ten years, a Delphi study is often behind it.

In the AI industry, the principle is mainly encountered in surveys among researchers. Hundreds of scientists are regularly asked when machines will achieve certain human capabilities. Such figures then appear in news reports and investor presentations. Not every one of these surveys is a genuine Delphi study, since the feedback rounds are often missing.

For readers of business news, a closer look is therefore worthwhile. How many experts were surveyed, how were they selected, how far apart were their answers? A Delphi forecast is a structured judgment, not a measurement. It can be wrong, and historical examples show that experts have expected technical developments both far too early and far too late.