AGI in 2027: what would the odds really mean?

Imagine a forecasting club assigning a 25% chance to AGI arriving by the end of 2027. That number is invented for this thought experiment, not a live market price, an expert consensus, or a measured forecast.

The club's members agree on the number for about thirty seconds. Then someone asks what arriving means. A company announcement? A demonstration? A system that can take on unfamiliar work and finish it reliably? Before they can discuss the odds, they have to decide what would make the prediction come true.

Define the finish line first

For their fictional exercise, the members choose a practical target: by December 31, 2027, an independently evaluated system must learn and complete a broad range of unfamiliar intellectual tasks at a competent human level, with limited supervision. They want evidence of adaptation across tasks, rather than a collection of carefully selected successes.

Even that definition leaves questions. How broad is broad? Which humans provide the comparison? How much help counts as limited supervision? The group writes those questions down instead of hiding them inside a confident percentage. Changing the test halfway through would change the prediction itself.

What one chance in four says

In the club's hypothetical forecast, 25% means the event is possible but not their most likely outcome. It does not mean that a system is a quarter of the way to general intelligence. It also does not identify the year in which success becomes certain. The number expresses uncertainty about one defined event before one deadline.

A single outcome cannot establish whether that probability was well calibrated. An event assigned a low probability can happen without making the original forecast absurd. The club would need to examine a larger record of comparable predictions to learn whether its confidence tends to match reality.

What would move the odds up?

The optimists propose evidence they would find persuasive: a system taking on genuinely unfamiliar work, carrying a plan through several stages, identifying its own mistakes, and recovering without a person supplying every correction. They would want independent evaluators to reproduce those results under conditions the system's creators did not select.

They also care about how much assistance is hidden around the model. A polished result could depend on extensive human preparation, many discarded attempts, or narrowly designed tools. Those contributions may still produce something useful, but the club wants to distinguish the overall system's success from the particular ability its forecast is meant to measure.

What would move the odds down?

The skeptics look for recurring failures under small changes in context. Suppose a system succeeds when a task is neatly specified but loses track when requirements conflict. Suppose it can generate a plan but cannot notice that an early assumption has become false. In their exercise, those weaknesses would weigh against a claim of dependable general capability.

Time also matters. An impressive prototype near the deadline might leave little opportunity for independent testing. The club separates the possibility that a capability exists privately from the possibility that its chosen public evidence standard is met. These are related events, but they are not interchangeable.

A useful forecast can change

The members agree to record their assumptions and revisit them when new evidence arrives. They will explain which observation changed their view and why. Revisions should follow evidence, rather than the excitement of a product name or the repetition of a persuasive headline.

Our invented 25% is only a starting point for that conversation. The more useful questions are what counts as AGI, what would demonstrate it, and which observations would make us reconsider. A forecast becomes easier to evaluate when its reasoning is as visible as its number.