Future of Forecasting: Collective and Artificial Intelligence

The original motivation for the Bayesian truth serum (BTS) may be dramatized by the ‘Cassandra predicament’ — a situation where a few experts hold unusual opinions. The truthfulness problem is to give such Cassandras a financial or reputational incentive to voice opinions that are certain to be greeted with disbelief. The truth problem is to confirm that Cassandras are genuine, that their judgments should override the opinions of the majority. BTS provides a solution to both problems under idealized conditions, assuming experts who are perfect Bayesian reasoners. I will present the logic behind BTS and review some recent results, touching also on the interplay with AI. .

To obtain accurate forecasts from collective judgments requires both effective elicitation and aggregation. I will primarily discuss an integrated mechanism for aggregating and incentivizing probabilistic beliefs, which relies on eliciting predictions about the average beliefs of others. Unlike most previous related work, our theoretical results on this mechanism hold for finite samples, e.g., expert panels. We exploit belief predictions to derive a common prior, which enables the computation of the exact full-information posterior over possible answers. On three datasets, our belief decomposition mechanism selects the correct crowd-wisdom answer with superior accuracy, penalizes hypothetical dishonest strategies with respect to how much they perturb beliefs, and tracks respondent expertise.

I have run commercial prediction markets and other crowd forecasting platforms continuously since 2000, and AI forecasting machines since 2024. This quarter century of natural experiments offers a view different from that of the laboratory. Four practical lessons stand out. (1) Accurate crowd forecasting does not require real-money wagering. In direct comparisons, play-money markets such as Hypermind match real-money counterparts such as Polymarket. Programs such as IARPA's Good Judgment Project and the Swedish Defense Research Agency's Glimt have also produced useful forecasts without wagers or significant rewards. (2) Brier scores are a poor way to communicate accuracy to decision makers. Measures that connect with how people actually think about predictions are needed. I will present some alternatives that have been well received by clients. (3) The Bayesian Truth Serum has made it possible to elicit long-term and ambiguous-resolution forecasts, which are in high commercial demand but unsuited to prediction markets and polls. Its value as an aggregator of probability forecasts, however, is unproven. (4) AI forecasting is about to transform the practice. It overcomes key commercial barriers of confidentiality, speed and cost, and it handles questions human crowds struggle with, such as rolling windows or very narrow topics. But this revolution will fail if one remaining pain point isn't solved: bridging the gap between how decision makers think about uncertainty and well-posed, resolvable MECE questions that human or artificial forecasters like to process. .

Many economically relevant decisions involve choosing among feature-rich options. Empirical data indicate that features that take extreme values loom large, exerting a disproportionate influence on choice, in a way that can lead to systematic biases. This pattern has been attributed to selective attention, for example, guided by the relative salience among features. We provide a normative foundation for this form of biased attention by showing that, under cognitive limitations, it is optimal—within a class of decision problems—to overweight extreme features. Specifically, we consider a model in which only a subset of features constitutes a deliberation sample used to guide choice. We characterize two objects jointly: a priority function—the probability with which a feature is included in the deliberation sample—and the decision function—the mapping from the deliberation sample to the chosen action—that together maximize the decision maker’s reward on average, over a prior distribution over possible decision environments. Under certain conditions, the priority function concentrates on extreme features and entirely disregards typical ones, while the decision function takes a particularly simple form—a tallying rule. While overweighting extreme events is optimal on average across environments, it can generate biased choices in specific settings. This comes as a possible explanation for biases such as the overvaluation of assets with positively skewed return distributions, the overpayment for insurance contracts, and the intransitivity of preference.

Perception and cognition can be understood as a process of building and using internal models of the world. These models are optimised for adaptive behaviour rather than truth, and can sometimes be misleading. As humans, we can draw on metacognitive capacities to recognise our misinterpretations, and to know when we need to learn more, seek advice, or reverse course. I will describe how this capacity to monitor and evaluate our own judgments can be measured with computational models that separate self-evaluation from performance. Convergent evidence obtained across rodents, non-human primates and humans indicates that metacognition depends on prefrontal circuits that track propositional confidence about possible decisions and actions. Applying the same psychophysical toolbox to AI systems reveals profiles that are recognisably metacognitive yet distinctly non-human. I will argue that these asymmetries matter for human–AI interaction and cognitive offloading, where fluent, confidently expressed outputs can result in negative metacognitive cycles of over-reliance.

Procrastination is widespread, costly, and persistent, yet it remains poorly understood. We develop a model of procrastination in stationary environments, with no deadlines, changing incentives, or new information. In this setting, chronic delay emerges as the only self-consistent policy: the agent completes the task with a fixed probability each period. This behavior resolves belief–action inconsistency but the agent gains nothing from delaying. We then extend the model to divisible tasks and show that allowing a voluntary break after completing a subtask never increases expected delay, and often reduces it. Enforcing such breaks can reduce delay even further, often by a substantial margin. .

Current efforts for online hate speech regulation mostly focus on explicit forms of hateful speech. This paper argues that harmful speech does not operate only at an explicit level, but that there are other forms of equally harmful language that are overlooked in the current regulatory paradigm: dogwhistles, figleaves and other forms of more implicit sexist, racist or homophobic speech fall outside the usual scope of what is deemed harmful, thus easily slipping under the radar. I argue that no single mechanism approach can adequately address all levels at which hateful speech operates. Since current content moderation practice alone is insufficient to detect the full range of harmful speech, the paper proposes three strategies to online speech governance: 1) swift platform-based responses to explicit harmful content, 2) counterspeech strategies based on perspective-taking, including a novel proposal of Kantian Maxim Mirroring, 3) preemptive institutional and group measures that can help mitigate the conditions under which hateful speech can arise.

Building upon the concept of hostile epistemology, I outline its definition and examine the inherent epistemic hostility of novel digital environments. Presuming that epistemically hostile environments exploit the inevitable vulnerability of our attempts to understand the world within the constraints of imperfect cognitive and contextual resources, I clarify how social media, nascent forms of native advertising, and overt disinformation manipulate our pursuit of clarity, understanding, and belonging. I locate a further vulnerability in the epistemological position that people are naturally inclined to believe what others tell them. After positing the incongruity of individuals prone to automatic trust and environments fabricated to disinform, persuade, or confuse them, I chart its implications for social polarization, democratic decision-making, and personal responsibility.

The premise that properly aggregated group judgments can outperform individuals underlies prediction markets, polls, and crowd forecasting. Yet in finite-state classification, error-free decision aggregation in the large-sample limit requires sufficiently informative response data. Prelec, Seung & McCoy (2017), building on insights behind the Bayesian Truth Serum, show that while votes alone do not suffice, votes paired with ‘peer predictions’ (respondents’ estimates of others’ voting behavior) are sufficiently informative for error-free two-state classification. Still open in the literature are: what other ‘side reports’ alongside votes are sufficiently informative for error-free classification; and how response data should optimally be aggregated. We answer these questions by developing an asymptotic theory of decision aggregation for general side information. Among other results, we derive: necessary and sufficient conditions on side reports for error-free classification; and asymptotically optimal aggregators under various objective functions, including general misidentification cost. Applied to peer predictions, these results yield aggregation methods with better asymptotic performance than those in Prelec, Seung & McCoy (2017). Of particular note, we show for the first time that peer predictions are sufficiently informative to guarantee almost-everywhere asymptotically error-free three-state classification, but insufficient for classification over four or more states. By contrast, ‘counterfactual outcomes’ are sufficiently informative to guarantee asymptotically error-free classification over any number of states. As crowds come to include AI agents drawing on shared data, votes grow correlated and less informative, making the design of cheap but sufficiently informative side reports increasingly central to reliable aggregation.

The ethics of artificial intelligence has developed rapidly, encompassing diverse theoretical approaches, practical guidelines, and regulatory proposals. However, unlike more established areas of applied ethics, such as medical or business ethics, it remains unclear whether AI ethics has a stable set of defining themes. This presentation examines whether a recognizable theoretical and thematic core can be identified in recent AI ethics literature and which issues remain peripheral, contested, or methodologically underdeveloped. The analysis draws on a comparative reading of recent philosophical and interdisciplinary work on AI ethics published over the past four to five years. Despite significant differences in methods, emphasis, and normative orientation, several themes appear central to current debates: responsibility for AI-driven decisions, transparency and explainability, privacy protection, the preservation of human autonomy, and algorithmic bias, fairness, and justice. Other topics, such as superintelligence and existential risk, the impact of AI on moral development, broader forms of technological governance, and social responsibility, appear prominent in some approaches but less central across the field as a whole. The presentation argues that AI ethics has not yet developed a fully established canon comparable to that of more mature fields of applied ethics. However, it has already begun to form a recognizable thematic core: a set of recurring problems that give the field its structure and coherence. Mapping this emerging core may contribute to a clearer methodological understanding of AI ethics and to its gradual differentiation from both technological hype and excessive alarmism.

Good decisions require good conversations but most groups have no way of knowing when their conversations are actually broken. One person dominates. Others go quiet. Ideas get lost. And nobody has the data to prove it. The Collective Intelligence Observer (CIO) is an AI-powered tool we built to change that. It analyzes meeting audio recordings and quantifies the structural conditions that enable or undermine collective intelligence. The pipeline moves from speaker diarization (pyannote.audio) through speech transcription (faster-whisper) to NLP-based idea extraction, computing three core metrics: an Interruption Score, a Participation Balance Score (via Gini coefficient), and an Idea Diversity Score combined into a weighted Collective Organisation Score visualized in an interactive dashboard. One of the hardest problems we faced was teaching the model to correctly identify and label individual speakers because without reliable diarization, every downstream metric collapses. That challenge revealed something important: measuring collective intelligence is not primarily a data problem. It is a signal problem. The conditions that make groups think well are subtle, overlapping, and easy to miss. This connects directly to the workshop's core concerns. If AI risks narrowing epistemic diversity in forecasting by pushing groups toward the same information, CIO offers a concrete diagnostic lens: what does homogenization actually look like in a conversation, and can it be detected before a decision is made? We propose that measuring process not just outcome is the missing layer in collective forecasting systems.

Prediction markets are increasingly treated as public probability engines: real-time summaries of dispersed beliefs about elections, macroeconomic releases, policy, litigation, technology, and geopolitics. As institutional capital and AI agents enter these markets, their role may shift from passive aggregation to active infrastructure for financial price discovery. Event-contract prices can improve equity, credit, rates, and crypto markets by translating uncertain future states into tradable probabilities; asset-market reactions feed back into prediction markets by revealing how informed capital prices those states’ consequences.The central question is whether this loop deepens collective intelligence or manufactures a new forecasting monoculture. Institutional participation can add liquidity, discipline, and arbitrage pressure, reducing convergence error and improving calibration. But if professional capital relies on overlapping models, data vendors, AI agents, retrieval systems, and risk constraints, prediction markets may grow faster without growing wiser—sharper prices in some domains, more fragile consensus in others. The monoculture risk here is not merely that models share biases, but that similar capital, positioned through similar automation, becomes correlated in a live, reflexive market that prices those beliefs back into the real economy, including industrial and maritime risk. Drawing on market microstructure, behavioral economics, and collective-intelligence theory, I propose a framework treating prediction markets as bidirectional information interfaces rather than isolated forecasting tools. Its claim: viewpoint diversity is an institutional and market-design problem, not only a modeling one—and preserving it determines whether automated, capital-backed forecasting sharpens collective intelligence or quietly narrows it.

This paper examines how AI-driven digital technologies affect democratic processes by influencing citizens’ epistemic and moral capacities, with particular attention to the challenges posed by misinformation, deepfakes, and increasing reliance on automated systems. Existing responses, such as education and regulatory interventions, appear insufficient to address the scale and adaptability of these problems. In response, the paper proposes a shift in perspective toward human–AI political symbiosis as a framework for mitigating epistemic degradation and improving the quality of democratic judgment and decision-making. Within this framework, human and AI systems are understood as cooperating agents that mutually enhance each other’s ability to operate within increasingly distorted informational environments. The paper distinguishes between two forms of symbiosis: a “weak” form, in which AI functions as an instrumental support for specific democratic tasks, and a “strong” form, in which human and AI capacities become interdependent for the stability and flourishing of democratic systems. The paper concludes by arguing that this symbiotic model offers a more adequate response to the evolving epistemic challenges of AI in democratic contexts, and that sustained human–AI cooperation may be necessary for preserving democratic resilience.

Recent advances in AI have raised the following controversy: to what extent can decision-making be outsourced without undermining autonomy? Our presentation addresses agentic AIs, systems capable of independently formulating goals, selecting means, and realising outcomes on behalf of human agents. While such systems promise significant efficiency gains and cognitive relief, they raise concerns that delegating decisions wholesale may undermine capacities associated with autonomy. We advance two claims. First, autonomous agency is compatible with substantial, though not unlimited outsourcing of decision-making. Drawing on accounts of autonomy by Christman (1988, 1991a, 1991b), Valdman (2010), and Double (1992), we argue that autonomy requires self-governance – the ability to shape one’s life in accordance with one’s values, commitments, and preferred “management style” – but not constant, hands-on deliberation. Autonomous agents may outsource many decisions, provided they retain evaluative control over the structures and processes that govern those decisions. Autonomy, however, is not fully outsourceable: wholesale delegation severs outcomes from an agent’s evaluative standpoint, thus undermining agency. Second, we assess whether agentic AI satisfies these conditions. While agentic AI could in principle support autonomy by conserving cognitive resources and handling low-stakes decisions, we argue that it faces structural risks not faced by another form of decision-making support: nudges. Nudges preserve agency more reliably because they involve at least minimal participation of agents in the pursuit of their values, and because they allow for resistance when assessments misfire. Agentic AI, by contrast, relies on outsourced value inference, but without offering opportunities for correction at the point of decision.

The phenomenon of online public shaming is pervasive, with profound consequences for individuals’ lives as well as broader effects on society. Attitudes toward this phenomenon are deeply divided. On the one hand, it is regarded as a dangerous instrument comparable to lynching. On the other hand, it is seen as a powerful tool for improving society and strengthening its moral norms. Perhaps this divergence in judgment reflects the diversity of cases. Paul Billingham and Tom Parr provide important criteria for systematizing our intuitions, assessing particular cases of public shaming, and subjecting the practice to normative constraints. They defend principles that restrict instances of public shaming by requiring, for example, proportionality, necessity, and respect for privacy and personal integrity. However, important doubts remain. I argue for a more restrictive view than the one proposed by Billingham and Parr. In my view, the enforcement of morality through sanctions (including those manifested in or resulting from public shaming) must be institutionalized and grounded in fair procedures. I then identify the conditions under which public shaming is justified, arguing that these conditions are analogous to those that justify civil disobedience.

The 'Surprisingly Popular' (SP) mechanism has proven effective at extracting hidden expertise and recovering ground truths by contrasting individuals' first-order judgments with their second-order predictions of the crowd distribution. In this paper, we explore whether the machinery of Bayesian collective intelligence can be extended beyond verifiable factual domains into normative and ethical decision-making. We introduce an experimental design spanning three epistemic environments: objective baseline forecasts, scenarios governed by classic cognitive biases (e.g., identifiable victim or omission effects), and non-verifiable moral trade-offs. Across these domains, we test whether meta-belief elicitation can reveal latent normative consensus or mitigate predictable behavioral distortions. Furthermore, we introduce randomized AI-generated arguments to examine how textual interventions shift private moral preferences relative to second-order expectations of peer persuasiveness.

In this paper, I argue that if we believe that capture by an invasive strategy is the central threat to an epistemically reliable system, then we must also believe that the selection of a random policy from some eligible set is epistemically more reliable than majoritarian decision-making. Institutional epistemologists working on the epistemically reliable design of democracy have rightly focused on capture by an invasive strategy as the central threat to epistemic powers of deliberation. However, this is insufficient to secure epistemic reliability of experimentalist democracy (which I believe best models the epistemic power of democracy). Instead, we must also protect decision-making from capture by an invasive strategy. This entails that we must forgo testing the most popular bet that has been chosen by a majority of deliberators, and instead, we must test a random one from a list of those we can all live with.

The starting dilemma of my discussion is whether radical belief polarisation (for instance, science denial, conspiracy theories) should be ascribed to individual cognitive failure, or to a structural feature of collective intelligence. While the 'epistemic deficit model' assumes that reducing ignorance and irrationality is enough to resolve polarisation, such interventions often fall short on their own, especially once beliefs become tied to identity. I argue that polarisation stems primarily from socially manufactured and digitally-mediated collective epistemic dynamics, not just from individual biases or epistemic vices. Digital and AI-driven communication amplify informational bubbles, echo chambers, and the process of de-factualisation. 'Bad beliefs' and belief polarisation, in other words, are not simply an emergent by-product of individual failings, but rather a democratic pathology. Such a diagnosis raises a further question: whether some forms of soft epistemic paternalism or protectionism might be justified as democratic remedies that prevent further erosion of epistemic infrastructure and democracy itself.

The growing use of artificial intelligence is transforming journalistic practices, newsroom operations, and audience perceptions of the media. Yet these developments are evolving at a much faster pace than legislative and policy frameworks can accommodate. This raises a broader dillema: how should regulation, public media policy, and professional practice interact in environment shaped by continuous technological change? These questions were partially explored through the Serbian project AI Media Hub (www.aimediahub.rs), established as an open platform for learning, research, and monitoring the impact of artificial intelligence on the media ecosystem. While the project’s initial focus has been on strengthening professional capacities and promoting the responsible use of AI in journalism, its findings have also highlighted the need to rethink the development of public media policies. Rather than asking only how strategic documents regulate emerging technologies, this presentation argues that equal attention should be devoted to how such policies are designed, negotiated, and continuously adapted through the participation of diverse stakeholders. The author explores whether AI literacy initiatives such as AI Media Hub can serve not only as educational interventions but also as experimental spaces for developing more participatory and adaptive models of public media policymaking. Focusing on Serbia and the Western Balkans, where institutional responses to AI remain fragmented, the presentation suggests that effective media governance requires not only updated regulatory instruments but also more adaptive, deliberative, and evidence-based policymaking processes capable of responding to the democratic challenges posed by artificial intelligence.

Can AI Increase Youth Political Efficacy? Framing a Research Agenda from Croatian Survey Data Drawing on the quantitative strand of the Political Competence of Youth at the Verge of Adulthood project (Institute for Social Research Zagreb, 2024-2027), this talk presents current findings on the political competence of Croatian upper-secondary students, focusing on political self-efficacy alongside political knowledge, information habits, and institutional trust. The dataset does not yet contain measures linking AI use to political efficacy, a gap shared across most existing youth political competence surveys. Rather than offer conclusive evidence, the talk uses the empirical findings as a starting point and proposes candidate approaches for capturing this link, including AI usage and trust items and perceived versus actual information diversity. The aim is to use workshop discussion to help crystallize a measurement strategy for AI's effect on political efficacy that could be tested in future survey waves.

Building on the concept of the hermeneutical turn in Technology Assessment as developed by Armin Grunwald and Bruno Gransche, this paper examines the limitations of scenario-based and prognostic approaches that conceive the future of technological development primarily as an object of prediction and extrapolation from existing trends. In contrast, the hermeneutic perspective emphasizes the openness of the future, the plurality of possible interpretations, and the reflective shaping of technological innovation through social processes. The second part of the presentation argues that artificial intelligence introduces a qualitatively new situation: technology is no longer merely the object of assessment and prediction but increasingly becomes the very medium and instrument through which predictions are generated. Predictive analytics, machine learning models, and generative AI actively constitute horizons of expectation and shape future decisions. This development transforms the epistemological foundations of Technology Assessment itself. Rather than assessing future technologies, attention shifts toward analyzing technologies that actively produce the future as the object of their own predictions. The paper explores the philosophical implications of this shift for our understanding of technology, anticipation, and responsibility in the age of artificial intelligence.