Seldon's conditions

The idea

Isaac Asimov’s Foundation was published as a book by Gnome Press, New York, in 1951. It collects stories that first ran in Astounding between 1942 and 1944, and two further volumes followed in 1952 and 1953 (SFE: Asimov). At the 1966 Hugo Awards the series won Best All-Time Series, in a field that included The Lord of the Rings (The Hugo Awards).

The setting is a Galactic Empire in its last centuries. Hari Seldon, a mathematician, has built a science he calls psychohistory. It assumes that the behaviour of humans in the mass, and with it the course of history, can be predicted by purely statistical means (SFE: Psychohistory). Seldon’s numbers say the Empire will fall and 30,000 years of disorder will follow. His plan aims to cut that to 1,000 (SFE: Asimov).

The science has a second condition, which the encyclopaedia quotes from the series: “the human conglomerate be itself unaware of psychohistoric analysis in order that its reactions be truly random.” The plot of the trilogy depends on it. The same entry notes Asimov’s own “strong scepticism about the workability of psychohistory” (SFE: Psychohistory).

My reading: the novels rest on two bets. Crowds are predictable even when people are not. And a forecast stays true only while the people it describes do not act on it.

The bridge

Replace Seldon’s equations with a language model and psychohistory turns into three research questions, each with data behind it.

The first is whether a machine can put useful probabilities on real future events. Forecasting systems built on language models are now scored against human forecasters on questions that had not resolved when the forecast was made.

The second is whether the mass is easier to predict than the person. One line of work uses models to simulate the answers of whole populations. Another measures how well any method can predict a single life.

The third is whether a forecast changes its own target. Machine learning has a name for Seldon’s second condition, performative prediction, and a body of theory around it.

The analogy holds where the claims are statistical: forecasts given as probabilities, scored after the fact, over many questions. It does not require the model to understand history, only to be right more often than the alternatives.

What research says now

Machines that forecast

In 2024, Halawi and colleagues built a system that retrieves news, has a language model write forecasts and aggregates them. They scored it on questions from forecasting platforms that opened after the models’ training cutoff. Its Brier score was 0.179 against 0.149 for the human crowd; lower is better. In every setting they report, the average of system and crowd beat both (Halawi et al., 2024).

ForecastBench, run by the Forecasting Research Institute, made the test continuous. It asks only about events with no known answer at the time of submission, so no answer can sit in the training data. On a 200-question sample in the original paper, expert forecasters beat the best model, with p under 0.001 (Karger et al., 2025). The benchmark now opens a round of 500 questions every two weeks. Half come from prediction markets, half are generated from time series with horizons from 7 days to a year (ForecastBench).

The gap has closed quickly. In January 2026 superforecasters still ranked first, 0.017 Brier points ahead of the best models. The team put that at about a year of progress and projected parity for November 2026, with a 95% interval from January 2026 to November 2027 (FRI, January 2026). By July they reported several systems statistically indistinguishable from superforecasters. One system, from Cassi AI, ranked above the superforecaster median on market questions, a first. The caveats are theirs. The superforecaster forecasts date from 2024 and are extrapolated. The confidence intervals overlap so much that the results fit parity better than outperformance (FRI, July 2026). By late October, six entries ranked above the superforecaster median on the live leaderboard, all within its confidence interval (ForecastBench, accessed 27 October 2026).

Not every test agrees. Forecast-Dojo, posted in September and revised on 26 October, replays 1,568 resolved Polymarket events alongside 18.8 million dated news articles. Research tools improved all 12 models tested. Every model still trailed the historical market forecasts (Ye et al., 2026).

Crowds of models

Seldon worked with aggregates, and so do most of the systems that score well. Schoenegger and colleagues pooled 12 language models on 31 questions from a three-month Metaculus tournament in which 925 people forecast. The model crowd was statistically indistinguishable from the human crowd. It also showed a human bias: the models leaned towards “yes”, although fewer than half the questions resolved that way. Showing two models the human median improved their accuracy by 17 to 28%, yet a plain average of human and machine forecasts did better still (Schoenegger et al., 2024).

The leading ForecastBench entries aggregate too. xAI’s entry in the January standings averaged eight forecasts per question. Cassi’s pipeline combines several models and, on market questions, lets a model revisit any forecast that differs sharply from the market price. That step alone improved its score by nearly 0.01 Brier points (FRI, January 2026).

My read: across these results, the machine does best when it reads the crowd, not when it replaces it.

Simulating the population

Forecasting predicts events. Psychohistory predicts people in aggregate, which is closer to another line of work: models as stand-ins for survey populations. Argyle and colleagues conditioned GPT-3 on thousands of demographic backstories from real survey respondents in the United States. It reproduced the response distributions of many subgroups, a property they called “algorithmic fidelity” (Argyle et al., 2023).

The fidelity is partial. Santurkar and colleagues compared model opinions with those of 60 US demographic groups. They found misalignment on the scale of the Democrat-Republican divide on climate change, and it persisted when the model was steered towards a group (Santurkar et al., 2023). SimBench, presented at ICLR 2026, unifies 20 datasets of human behaviour. The best model scored 40.80 out of 100. Instruction tuning improved results on questions where people agree and worsened them where people disagree. Models struggled most with specific demographic groups (Hu et al., 2026).

Two results go further. Park and colleagues built agents for 1,052 Americans from two-hour interviews, surveys, or both. On held-out survey items the agents reached 82% to 86% of the consistency people show with their own answers two weeks later. Agents given only demographics reached 74% (Park et al., 2026). Centaur, published in Nature in 2025, is Llama 3.1 70B fine-tuned on Psych-101: 160 experiments, more than 60,000 participants and more than 10 million choices. It predicted held-out participants better than existing cognitive models in almost every experiment (Binz et al., 2025).

A reply in National Science Open tested Centaur on four tasks after removing the instructions or replacing them with a misleading one. Centaur still beat the cognitive models in most conditions. The authors read this as reliance on statistical cues in the data, not understanding of the task (Liu and Ding, 2025).

Predicting one life

Seldon’s science leaves the individual out; in the trilogy, a single person is the wildcard that disrupts the plan (SFE: Psychohistory). One large study suggests he had reason to be careful. In the Fragile Families Challenge, 160 teams tried to predict six outcomes, among them a child’s grade point average and a household’s eviction. They had data on 4,242 families with 12,942 variables each. The best predictions reached an R² of about 0.2 for two outcomes and about 0.05 for the other four. That was only slightly better than a regression on four variables chosen by a domain expert, and that simple model beat many of the submissions. Errors depended strongly on which family was predicted and weakly on the method (Salganik et al., 2020).

A study with far more data reports a different result. Savcisens and colleagues trained a transformer on Danish registry data, treating each life as a sequence of events in health, education, work, income and address. It predicted outcomes from early mortality to personality nuances and beat established models by what the authors call “a wide margin” (Savcisens et al., 2024).

The forecast that changes its target

Seldon’s second condition has a formal counterpart. Perdomo and colleagues put it in one sentence: “When predictions support decisions they may influence the outcome they aim to predict.” They call such predictions performative. Their equilibrium, performative stability, describes predictions calibrated against the outcomes that follow from acting on them, not against past outcomes (Perdomo et al., 2020). A later review by Hardt and Mendler-Dünner separates learning about the world from steering it, and asks how much a platform can steer people through its predictions (Hardt and Mendler-Dünner, 2023).

Forecasting with language models has not caught up with this. An August 2026 survey of forecasting agents lists “methods for handling feedback between deployed forecasts and the outcomes being forecast” as future work (Xu et al., 2026).

Keeping the forecaster honest

Evaluation faces Seldon’s condition in reverse: here the forecaster must not know the outcome. Paleka and colleagues catalogue how it leaks. Take a 2021 question on whether Queen Elizabeth would live to 100. Used as a resolved test question in 2025, it can only have resolved “no”, because a “yes” could not be known before 2026. Search engines restricted to old dates also return documents shaped by later events (Paleka et al., 2026). Telling a model to ignore what it learned after a date does not fix this. Across nine models, instructed models kept inflated accuracy and confidence on questions they had been told to forget (Li et al., 2026).

Where the analogy breaks

Time scale. Seldon forecasts across centuries. The longest ForecastBench horizon is a year, and a live benchmark has to wait for its questions to resolve. My read: the evidence covers weeks and months. Nothing yet tests forecasts across decades, and nothing can test that quickly.

Mechanism. Psychohistory is a theory of how societies move. The systems above search, read and average, and some consult the market price. My read: they resemble a diligent analyst with a good news feed more than a set of social equations.

The mass. Psychohistory treats the mass as the easy part. For models, the simulation results point the other way. SimBench found that instruction tuning favours consensus and loses the spread of opinion, and specific groups are simulated worst. A population with its disagreements smoothed out is a different population.

Secrecy. Seldon kept his analysis hidden. Machine forecasts are published, scored against markets and, in some pipelines, adjusted towards market prices. Speculation: if model forecasts become cheap and widely used, the markets they are scored against will start to contain those forecasts. Benchmarks that compare models with markets would then partly measure models against themselves.

The individual. My read: here the book and the research mostly agree. The Fragile Families result is what Seldon would expect. The registry result shows that “unpredictable” can also mean “not yet measured”, and that turns a forecasting question into a privacy question the novels never had to face.

Follow the thread

• ForecastBench: the leaderboard updates nightly (about) and the forecast data can be downloaded. FRI announced a fresh superforecaster round for October and calls it a rematch, and ForecastBench lists the tournament on 25 October 2026 (FRI on X, 22 September 2026; ForecastBench). Results will take weeks, because models only join the leaderboard 50 days after their forecasts (FRI, July 2026). Watch whether parity holds against the new human forecasts.
• Pitfalls in Evaluating Language Model Forecasters (Paleka et al., 2026): read it before trusting any claim that a model beats human forecasters.
• Simulated Ignorance Fails (Li et al., 2026): why backtests on events before a model’s cutoff give inflated results, and a recommendation against them.
• Forecast-Dojo (Ye et al., 2026): resolved market events replayed with dated news, built for longitudinal study and training. Check it against the two leakage papers above.
• SimBench data: test any model’s population simulation, and compare consensus questions with contested ones.
• Psych-101 and the Centaur critique code: the dataset is public under Apache 2.0, and Liu and Ding published their analysis scripts. Both the claim and the critique can be rerun.
• The Fragile Families Challenge (Salganik et al., 2020): a model for honest benchmarks, with a simple expert baseline that many complex submissions failed to beat.
• Performative Prediction: Past and Future (Hardt and Mendler-Dünner): the theory in review form. The open question for forecasting agents: how should a public forecaster account for its own effect on what it predicts?