# Distributed forecasting platforms and the future of AI

Author: Fernando Nieto Lobato
Original publication: 2025-06-25
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-91-senales-debiles-decisiones
English URL: https://elcontemplador.github.io/estrategia-english/essays/091/
Status: Published translation

This is a translation of the original Spanish essay published on 25 June 2025. Its claims, examples and forecasts retain that historical context.

English publication: 2026-09-29

*Archive note: this essay describes forecasts and expectations in mid-2025. The figures and screenshots are historical snapshots, not current predictions. Where the prose and a screenshot give different estimates, both are preserved.*



In recent years, social forecasting platforms — websites where large numbers of people predict future events — have gained prominence as tools for anticipating trends in technology and politics. Leading examples include Metaculus, a collaborative forecasting community where no money changes hands, and Polymarket, a prediction market where users stake real money, usually cryptocurrency, on future outcomes. These platforms have achieved significant successes in the recent past. **Metaculus predicted with striking accuracy that there would be a Covid vaccine before the end of 2020**, when many experts still doubted that it could happen so quickly. **Polymarket correctly anticipated Joe Biden's victory**, adjusting its probability in real time as votes were counted state by state.



**These platforms draw on the “wisdom of crowds”, aggregating many individuals' estimates to produce forecasts that are generally better calibrated than isolated expert opinions or traditional polls.** Indeed, numerous studies show that prediction markets tend to outperform conventional methods in accuracy, because market prices quickly incorporate all the available information about an event.



In the case of **[Metaculus](https://www.metaculus.com/)**, founded in 2015, thousands of volunteers make probabilistic forecasts. Algorithms then weight these predictions, giving greater weight to people with better track records, to produce an aggregate consensus. Metaculus distils the wisdom of crowds through an open community that makes probabilistic predictions without money being involved, weighted by past accuracy and enriched by public debate. This gives it good calibration over short-term technological horizons.



By contrast, **[Polymarket](https://polymarket.com/)**, created in 2020, operates as a blockchain-based prediction market. Participants buy or sell contracts whose prices reflect each event's implied probability almost instantly. It provides fast and often accurate signals about well-defined events, but is susceptible to hype when information is scarce — as the delay to GPT-5, discussed below, illustrates. The two platforms therefore embody complementary approaches: deliberative aggregation based on reputation, and market signals driven by financial incentives. Other variants — from [Good Judgment Open](https://www.gjopen.com/), where expert “superforecasters” refine their judgements, to **[Manifold Markets](https://manifold.markets/home)**, with its virtual currency and high participation — broaden the range of methods and allow biases to be compared.



**Why are these tools gaining prominence in artificial intelligence? Above all, because AI development is currently both rapid and uncertain. In this situation, combining the forecasts of hundreds of informed people can provide an early signal of what to expect, and when. That is valuable when planning technological and political strategies.** Their predictions are also beginning to influence discussion beyond their own communities: reports by official institutions and specialist media increasingly cite these markets and forecasting forums as inputs to the debate about AI's future.



## Current predictions about AI development



What do crowds predict about the evolution of artificial intelligence? Collective forecasting platforms maintain active questions about its most consequential milestones: when we might achieve human-level or artificial general intelligence (AGI), social effects such as the automation of work, global AI regulation, and the pace of advances in new models. Below, we examine some of the most relevant and striking predictions from Metaculus, Polymarket and similar sites, reflecting expectations in mid-2025.



### The timeline towards artificial general intelligence



One central focus is estimating when artificial general intelligence might emerge, capable of performing any intellectual task at a human level. **These estimates have moved dramatically closer in recent years**, as the following chart shows. In 2020, for example, Metaculus users tended to think AGI was several decades away: their average forecast at the time placed it around 2053. Following recent advances, however, the community has shortened its timelines and now expects AGI around 2026–2035. In other words, the anticipated horizon for human-level AI has contracted from roughly 30 years to just a few, reflecting the increasingly widespread view that “AGI before 2030” is plausible.



[![Chart of falling forecast years until AGI, annotated with model launches.](https://elcontemplador.github.io/estrategia-english/assets/images/49fb1f46-2161-4aa0-9aff-d4f60587d50a_1600x1154.png)](https://substackcdn.com/image/fetch/$s_!5R6u!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fb1f46-2161-4aa0-9aff-d4f60587d50a_1600x1154.png)



*Original English chart by 80,000 Hours: “Years until first general AI system announced”. The horizontal axis is the date of the forecast, from June 2020 to beyond December 2024; the vertical axis is years until AGI, labelled as a logarithmic scale. Its annotations show a pre-GPT average of 80 years, roughly 50 years around the announcement of GPT-3 with its API in closed beta, roughly 18 years around ChatGPT's public launch, and five years at the latest data point. Other marked milestones are Google's announcement of LaMDA2 and the launch of GPT-4. The source is the Metaculus question, “When will the first general AI system be devised, tested, and publicly announced?” Its definition requires a two-hour adversarial Turing test, high accuracy on a question-and-answer expertise benchmark and the APPS coding benchmark, and general robotic capabilities sufficient to assemble a Ferrari from instructions.*



Not everyone agrees on an exact year. Views range from people who see AGI as possible within a couple of years to those who put it several decades away. But the collective median now points to the late 2020s or, at the very moment I am writing this newsletter, the early 2030s. The predicted date has slipped slightly in recent weeks, but, as the chart shows, a major new AI announcement tends to bring it forward. This trend towards shorter timelines followed milestones such as GPT-3 and GPT-4, which demonstrated unexpected capabilities and prompted a reassessment of how close we might be to truly general AI. We will have to see how the imminent arrival of GPT-5 shifts the forecast. It is worth stressing that even the forecasters acknowledge considerable uncertainty. None of these AGI projections is wholly reliable or definitive, given the unprecedented nature of the event. Even so, the fact that a significant share of the community considers AGI achievable this decade reveals the prevailing climate of expectations.



[![Two Metaculus predictions about human-machine parity and the first general AI.](https://elcontemplador.github.io/estrategia-english/assets/images/de7a6d7f-5119-43d5-ada1-a864ed47f0e3_1053x699.png)](https://www.metaculus.com/tournament/agi-horizons/?order_by=-vote_score)



*English reading of the two Metaculus cards: “Will there be intelligence parity between humans and machines before 2040?” — community prediction 95%, 2,541 forecasters, closing 1 January 2035. “When will the first general AI system be devised, tested, and publicly announced?” — August 2033, 1,692 forecasters, closing 24 December 2199. The second card shows a distribution with a long right-hand tail; these are the estimates displayed in this particular screenshot.*



### The evolution of AI models: the GPT-5 example



Over shorter horizons, predictions abound about the next generations of AI models. One specific case is OpenAI's GPT-5. After GPT-4 launched in 2023, enormous interest arose in when GPT-5 would arrive and what it could do. Forecasting platforms opened markets on the question: Polymarket, for instance, allowed people to bet on whether “GPT-5 would be released before the end of 2024”. During 2023, community sentiment was very optimistic — perhaps too optimistic. Forums as well as Metaculus and Manifold confidently forecast that OpenAI would announce GPT-5 towards the end of 2024. At one point, Polymarket prices implied a probability close to 90% that GPT-5 would be publicly available before 2025, effectively treating it as a near certainty. **But that prediction did not come true**: OpenAI had not released a “GPT-5” by the end of 2024, so those markets resolved to “No”. **The gap between the forecast and reality offered bettors a valuable lesson in humility, showing that even crowds can overestimate the speed of certain developments.**



Expectations have adjusted in light of the information now available. On Metaculus, for example, the community consensus puts the probability of OpenAI announcing GPT-5 by the end of August 2025 at around 50%. There are numerous rumours, and even statements from OpenAI, pointing to GPT-5 this summer. This case illustrates how forecasting platforms correct themselves dynamically as new signals emerge — OpenAI statements, leaks, intermediate developments such as GPT-4.5 — adjusting probabilities in real time. It also highlights a methodological issue: in the absence of official information, crowds tend to extrapolate recent trends, such as the rapid move from GPT-3 to GPT-4, and can become overconfident. That is an important factor when interpreting their predictions.



[![Historical Polymarket probabilities for GPT-5 release dates.](https://elcontemplador.github.io/estrategia-english/assets/images/f9a77090-1767-480c-8091-305885d8f7d0_1600x948.png)](https://substackcdn.com/image/fetch/$s_!K-z5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff9a77090-1767-480c-8091-305885d8f7d0_1600x948.png)



*Historical Polymarket screenshot, “GPT-5 released by...?”: 30 June, 2%; 31 July, 30%; 31 December, 94%. Markets for 31 March, 30 April and 31 May have all resolved to “No”. Total volume shown: $880,851. These displayed market probabilities are distinct from the Metaculus estimate discussed in the paragraph above.*



*Additional figures visible in this historical screenshot:*



| Deadline | Volume | Buy Yes | Buy No |

| --- | --- | --- | --- |

| 30 June | $311,428 | 2.6¢ | 97.9¢ |

| 31 July | $182,287 | 31¢ | 71¢ |

| 31 December | $82,305 | 94.0¢ | 6.6¢ |

| 31 March, resolved No | $87,684 | — | — |

| 30 April, resolved No | $189,205 | — | — |

| 31 May, resolved No | $27,942 | — | — |



*The displayed changes beside the open-market probabilities are down 0%, down 3% and down 0%, respectively. These are archived interface values, not live prices.*



### Automation and employment



Another critical area of forecasting is AI's effect on the labour market and the economy. Here, the questions concern how much human work intelligent machines will replace, and when. Crowd forecasting platforms have approached this with nuance, constructing quantitative scenarios such as: “What percentage of today's workers will be replaced by AI systems by 2030?” In a recent forecasting exercise involving experts in economics and AI policy, organised on Metaculus as part of the “Threshold 2030” tournament, the median prediction was that roughly 10% of current jobs could be completely replaced by AI in 2030. In other words, one in ten jobs is expected to be automated from end to end over the next half-decade.



The same forecasters stressed that this moderate percentage reflects several practical brakes on full automation: institutional inertia; the cost and time needed to implement new technologies; the availability of cheap labour in some sectors; and the fact that, in many roles, AI can increase productivity without replacing the human entirely. In short, current collective predictions paint a picture in which AI automation will be transformative but gradual over the next five to ten years. Beyond 2030, uncertainty grows and opinions diverge. For the near term, however, the forecast consensus as it stands is one of manageable disruption rather than immediate labour-market collapse.



[![Metaculus forecast of workers replaced by end-to-end AI systems in 2030.](https://elcontemplador.github.io/estrategia-english/assets/images/13b9e446-4ed8-4643-8f71-2f63aad96ac5_1158x728.png)](https://www.metaculus.com/c/threshold/35329/porcentaje-de-trabajos-actuales-realizados-por-ia/)



*English reading of the screenshot: “What percentage of current workers will be replaced by AI systems doing work from beginning to end in 2030?” Closing date: 31 December 2029. There are 22 forecasters. Community prediction: 14.1%, with a displayed interval of 6.23%–27.6%; the account's own prediction: 15.8%, with an interval of 12%–19.6%. The image's 14.1% and the prose's “roughly 10%” are both retained as they appear in the original.*



### AI regulation and electoral importance



Given the growing recognition of advanced AI's risks, another pressing question is whether a global governance framework or international agreements to control AI will emerge. Metaculus has hosted questions such as: “[Will an international AI regulatory body, analogous to the IAEA for nuclear energy, be created before 2030?](https://www.metaculus.com/questions/30249/international-tai-oversight-agency-before-2030/)” Strikingly, the current collective answer is highly ambivalent: around a 50% probability. The aggregate forecast is roughly a 50% chance that, by 2030, an operational international agency backed by at least 20 countries will exist to oversee transformative AI.



At the social level, public awareness of AI is also expected to grow: **on Manifold, the probability that AI will be among US voters' five most important issues in 2028 is estimated at about 65%**. There is even a similar probability of mass protests against AI, involving more than 100,000 people, before 2030.



[![Manifold forecast about AI among the five leading US electoral issues.](https://elcontemplador.github.io/estrategia-english/assets/images/a6a26a43-0513-4574-87e7-429d2ad202b6_1129x1225.png)](https://substackcdn.com/image/fetch/$s_!6LJv!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa6a26a43-0513-4574-87e7-429d2ad202b6_1129x1225.png)



*Original English Manifold question: “Will AI be among the top 5 most important issues for voters in the lead up to the 2028 election?” The screenshot shows a 65% chance. Its resolution criteria include AI, AI safety and externalities such as job losses or economic problems caused by autonomous systems, or autonomous weapons systems, as measured by Gallup polls. The issue must rank in the top five by the share of respondents rating it extremely or very important. The displayed example is “Top Issues for Voters 2022”: https://news.gallup.com/poll/244367/top-issues-voters-healthcare-economy-immigration.aspx.*



## A final reflection: their role in decisions as AI accelerates



**As artificial intelligence advances in leaps and bounds, sometimes unpredictably, social forecasting platforms such as Metaculus and Polymarket are emerging as valuable allies in both public and private decision-making.** They offer a window onto AI's plausible future, built from collective knowledge rather than an individual's crystal ball and, in Polymarket's case, from the financial incentives of many participants. **This collective intelligence can counterbalance hunches and institutional inertia: it forces assumptions to be made explicit, numerical probabilities to be assigned, and beliefs to be updated quickly as new evidence emerges.**



In terrain as slippery as anticipating AI's next decade, with its promises of prosperity and potential dangers, these aggregate signals can improve our preparedness. If experts and bettors converge on a substantial probability of AGI around 2030, for example, governments and businesses would have reasons to accelerate safety measures, regulation and adaptation now, avoiding being caught off guard. Equally, if collective forecasts suggest a technology is still far from maturity, that could prevent premature investment or unfounded panic.



In the highly uncertain circumstances we face, their practical value is considerable when they are used judiciously. Governments and businesses can treat them as **an early-warning system that quantifies the probability of critical scenarios and demands that plans be updated quickly**. However, **we need to watch for feedback effects — policies that act on a forecast and change it — biases within technology-enthusiast communities, and possible insider-information manoeuvres in betting markets**. Forecasting platforms do not replace expert analysis, but, used alongside it, they provide a dynamic, democratised measure for anticipating AI's future and making better-grounded strategic decisions.



**A critical, balanced approach is essential. Neither Metaculus nor Polymarket, nor any predictive tool, eliminates the uncertainty inherent in AI's future. Their forecasts should be interpreted cautiously: as indicators that can change, not a fixed destiny.** One danger would be policymakers seeing, for instance, “70% probability of mass unemployment in 2035” and reacting without understanding the assumptions behind that number.



**Nor should every decision be subordinated to a majority prediction. History offers examples of crowds getting things spectacularly wrong — the dot-com bubble, energy forecasts and so on — while, in disruptive technology, visionary experts in the minority sometimes prove right against prevailing opinion. Smart leaders and strategists will therefore use these platforms as one voice in the choir, not the entire score.** Ideally, they will combine information from prediction markets and forecasting forums with technical analysis, qualitative judgements from domain experts, and broader ethical and political considerations.



**For all that, these platforms' distinctive contribution is to democratise and quantify thinking about the future.** They allow not only senior executives and academics to express a view on what lies ahead, but any informed person with a sound argument, whose small contribution of probability adds to the common forecast. On a matter as consequential as the development of artificial intelligence, which will affect all humanity, there is value in open tools that make foresight a distributed activity rather than something confined to an ivory tower. This could produce better shared strategic intelligence: a start-up entrepreneur, a regulator in Brussels, a researcher in Tokyo and a curious citizen can all consult these collective predictions and draw insights for their own circumstances.



Ultimately, **social forecasting platforms today act as radar in the fog of AI's future: they cannot guarantee that we will avoid every obstacle, but they give us clues to where the icebergs may lie and which way the winds of change are blowing**. Used responsibly, they can guide public policy and corporate strategy alike, helping us navigate accelerating AI development with a little more clarity and foresight. In a world where AI is progressing rapidly, that collective capacity to look ahead, however imperfect, is preferable to moving blindly. Predictions from Metaculus, Polymarket and similar platforms should be seen as ongoing quantitative conversations about possible futures, not a destiny already written. Participating in them, or at least listening, can be an integral part of how societies and leaders prepare for what the AI revolution may bring in the coming years.



[Fernando Nieto Lobato](https://www.linkedin.com/in/fernandonietolobato/)



*Director of Digital Innovation at Institución Educativa ALEPH*