When AI Predicts a Coup: The Ethics of Forecasting Instability

Takeaway	Detail
Self-fulfilling prophecy risk is the primary ethical hazard	A forecast of instability can trigger preemptive actions by governments or opposition groups, making the prediction come true regardless of underlying conditions.
Combine quantitative models with qualitative case comparisons	Pairing machine learning outputs with structural analogies from historical precedents improves forecast reliability and reduces blind spots from overfitting.
Assess model calibration by comparing predicted probabilities to observed event frequencies	A well-calibrated model will show that events predicted at 70% probability occur roughly 70% of the time over a validation period.
Use out-of-sample validation and cross-validation to prevent overfitting	These techniques ensure the model generalizes beyond the historical data it was trained on, a common failure mode in instability forecasting.
Precision and false positive rate are the key metrics to monitor	Precision measures how many predicted coups actually occur; false positive rate tracks how often the model cries wolf, which matters for trust and policy response.
UNESCO's ethical AI recommendations provide a governance framework	The global guidelines on responsible AI design apply directly to political instability forecasting systems, covering transparency, accountability, and human oversight.
AI increases the value of human judgment, not replaces it	Prediction alone is insufficient; human interpretation and decision-making remain essential for interpreting forecasts and choosing interventions.

When an AI model signals a high probability of a coup, the forecast itself becomes a political actor. The ethics of forecasting instability hinge on whether the prediction prevents the event or provokes it—a dilemma known as the self-fulfilling prophecy. This guide examines the core ethical considerations, technical limitations, and governance frameworks that practitioners must navigate when deploying instability forecasting models (IFMs) in high-stakes political environments.

Recent research from the Stimson Center identifies the short-term race for AI dominance as a top global risk for 2026, while a 2022 article in International Security argues that AI increases the value of judgment in international politics. These developments underscore that the challenge is not merely building accurate models, but deciding how to act on their outputs without destabilizing the very systems they aim to protect.

Assess model calibration by comparing predicted probabilities to observed event frequenciesA well-calibrated model will show that events predicted at 70% probability occur roughly 70% of the time over a validation period. Use out-of-sample validation and cross-validation to prevent overfittingThese techniques ensure the model generalizes beyond the historical data it was trained on, a common failure mode in instability forecasting. Precision and false positive rate are the key metrics to monitorPrecision measures how many predicted coups actually occur; false positive rate tracks how often the model cries wolf, which matters for trust and policy response. UNESCO’s ethical AI recommendations provide a governance frameworkThe global guidelines on responsible AI design apply directly to political instability forecasting systems, covering transparency, accountability, and human oversight. AI increases the value of human judgment, not replaces itPrediction alone is insufficient; human interpretation and decision-making remain essential for interpreting forecasts and choosing interventions.

What Ethical Frameworks Apply to AI Coup Predictions?

Three ethical frameworks dominate the discussion of AI coup predictions: consequentialist, deontological, and virtue-based approaches. Consequentialist frameworks evaluate forecasts by their outcomes, which directly raises the self-fulfilling prophecy problem. Deontological frameworks focus on the act of prediction itself, asking whether it is ever permissible to generate and disseminate a forecast that could destabilize a sovereign state. Virtue ethics examines the character of the institutions building and deploying these models, demanding transparency, humility, and accountability from the organizations involved.

The self-fulfilling prophecy mechanism works through several channels. A government that receives a high-risk alert may arrest opposition figures, suspend civil liberties, or deploy military forces to the capital. These actions can create the conditions for a coup attempt that would not have occurred otherwise. Conversely, a public forecast of instability can embolden plotters by signaling that the regime is weak or that international actors expect a transition. The model does not merely observe reality; it alters it. This feedback loop is the central ethical challenge that distinguishes political instability forecasting from weather prediction or disease outbreak modeling.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence, adopted by 193 member states in 2021, provides a procedural framework that applies directly to coup prediction systems. The recommendation calls for proportionality, transparency, and human oversight in all AI systems that affect human rights. For instability forecasting models, this means that outputs should never be treated as deterministic verdicts. The recommendation also requires that affected communities have access to meaningful explanation of how predictions are generated and the ability to contest them.

Precision and false positive rate are the two metrics that translate ethical concerns into technical requirements. Precision measures how many of the model’s positive predictions are correct. A model with low precision generates many false alarms, each of which carries real-world consequences: wasted resources, eroded trust, and potential overreaction by security forces. False positive rate measures how often the model incorrectly flags a stable period as high risk. In political instability forecasting, a false positive is not a harmless error. It can damage diplomatic relations, trigger unnecessary military deployments, or provide a pretext for political repression. The Stimson Center’s Top Ten Global Risks for 2026 report notes that the short-term race for AI dominance itself poses financial risks that may affect global stability, adding another layer of ethical complexity to the deployment of these systems.

Practitioners should adopt a tiered ethical review process before deploying any instability forecasting model. First, conduct a pre-deployment impact assessment that identifies all stakeholders who could be harmed by a false positive or a self-fulfilling prophecy. Second, establish a minimum precision threshold below which the model’s outputs are not acted upon. Third, implement a human-in-the-loop review for any prediction above that threshold, with explicit documentation of the decision to act or not act. Fourth, publish an annual audit of the model’s performance, including false positive rate and precision, disaggregated by region and time period. These steps do not eliminate the ethical risks, but they make them visible and accountable.

The concrete action you can take today is to review the UNESCO Recommendation on the Ethics of Artificial Intelligence and compare its principles against any instability forecasting system you encounter or build. Map each principle—proportionality, transparency, human oversight—to a specific design decision in the model pipeline. If a system cannot satisfy that mapping, it should not be deployed in operational settings.

How to Assess a Model’s Calibration and False Alarm Rate

To assess a model’s calibration and false alarm rate, you must compare its predicted probabilities against actual observed outcomes over a defined validation period. The false alarm rate, or false positive rate, measures how often the model incorrectly flags a stable period as high risk. You calculate it by dividing the number of false positives by the total number of actual negative cases in your test set.

The standard metrics for this evaluation are precision and the false positive rate. Precision, also called positive predictive value, tells you what fraction of the model’s positive predictions were correct. The false positive rate, by contrast, measures the proportion of all stable periods that the model mistakenly classified as high risk. Both metrics are essential because a model can appear accurate by rarely predicting any coups at all, yet still generate a high false alarm rate when it does issue a warning. You should compute both on a held-out test set that the model has never seen during training.

Calibration is typically assessed using a reliability diagram, which plots predicted probabilities on the x-axis against observed frequencies on the y-axis. A perfectly calibrated model produces points along the diagonal line. If the points fall below the diagonal, the model is overconfident — it predicts higher probabilities than what actually occurs. If they fall above, it is underconfident. For instability models, overconfidence is the more dangerous error because it can lead decision-makers to act on predictions that are less reliable than they appear. The Brier score, which measures the mean squared difference between predicted probabilities and actual outcomes, provides a single-number summary of calibration quality. Lower Brier scores indicate better calibration.

A common practitioner mistake is to rely solely on accuracy as a performance metric. You must instead focus on precision and false positive rate at the decision threshold you intend to use. The threshold itself is a policy choice, not a statistical one. A lower threshold catches more true events but generates more false alarms. A higher threshold reduces false alarms but may miss real threats. The acceptable tradeoff depends on the cost of acting on a false alarm versus the cost of missing a real coup.

One concrete method for evaluating calibration is to use a rolling window backtest. Take historical data from 2000 to 2010 to train the model, then test predictions for 2011 to 2015. Compare the predicted probabilities for each country-year in the test window against actual events. Repeat this process by moving the training window forward by one year at a time. This simulates how the model would have performed in real time and reveals whether calibration degrades as the political environment changes. The UNESCO Recommendation on the Ethics of Artificial Intelligence, adopted by 193 member states in 2021, provides a procedural framework that applies directly to this evaluation process. It calls for transparency in how predictions are generated and for human oversight of any system that affects human rights.

You can take a concrete action today: download a publicly available instability dataset, such as the Political Instability Task Force dataset or the Armed Conflict Location and Event Data Project (ACLED) data, and run a simple logistic regression model on a subset of features. Compute the predicted probabilities for a held-out test set, then plot a reliability diagram using any standard statistical software. Compare the observed event rate at each probability bin to the predicted rate. If the points deviate from the diagonal by more than 0.1 in any bin, the model requires recalibration before it can be used for operational decisions. This exercise takes about two hours and gives you direct experience with the calibration assessment process.

Which Data Sources and Features Drive Instability Forecasts?

Political instability forecasts, including coup predictions, rely on supervised learning models trained on historical examples of coups and non-coup periods to recognize associated patterns. The data sources that drive these models fall into three broad categories: economic indicators, political event data, and social signals. Economic indicators typically include GDP per capita, inflation rates, trade openness, and natural resource dependence, drawn from sources like the World Bank and IMF. Political event data comes from structured databases such as the Political Instability Task Force dataset, the Armed Conflict Location and Event Data Project (ACLED), and the Cline Center's Coup d'État Project, which code events like protests, government crises, and military movements from news archives. Social signals are increasingly drawn from social media feeds and news sentiment analysis, though these introduce significant noise and bias issues.

The feature engineering process typically transforms these raw sources into predictive variables. Common features include lagged economic growth rates, the number of coup attempts in neighboring countries in the prior year, the age of the current regime, and the presence of ethnic fractionalization. A known failure mode is overfitting to historical patterns, where a model learns spurious correlations from a specific decade or region. Mitigation requires out-of-sample validation and cross-validation techniques during development, as noted in the literature on instability forecasting models. The precise timing and nature of abrupt regime shifts may be beyond reliable AI prediction when the relevant data does not yet exist in the observable information set — a coup that emerges from a sudden, unanticipated military faction may have no prior signal in any open-source dataset.

Open-source data offers broad coverage but suffers from incompleteness, missing values, and reporting bias. Proprietary intelligence feeds, such as those from private risk consultancies or government agencies, may offer higher signal-to-noise ratios but introduce access restrictions and their own bias toward observable, quantifiable events. A documented approach to improve forecasting combines quantitative models with structural analogies — qualitative case comparisons that assess whether a country's current conditions resemble historical precursors to instability. The GSDRC report on this method advises caution on how to combine them, as qualitative judgments can introduce their own cognitive biases.

A key ethical concern specific to data sources is the risk of self-fulfilling prophecies. If a model flags a country as high risk and that prediction becomes public or influences diplomatic or economic actions, it may trigger capital flight, regime crackdowns, or preemptive military moves that increase the actual probability of a coup. This feedback loop means the choice of data sources and the transparency of their use are not merely technical decisions but ethical ones. You can take a concrete action today: open the ACLED data portal, filter for a region of interest, and examine the raw event types and their date stamps. Note how many events are coded from local news sources versus international wire services, and consider how that reporting bias might affect a model trained on that data. This exercise takes about thirty minutes and reveals the data quality constraints that every instability forecast must navigate.

How to Set Up a Human-in-the-Loop Review Process

A human-in-the-loop review process for AI-generated instability forecasts should be structured as a gated workflow where model outputs are flagged for mandatory review by domain experts before any action is taken. The core mechanism is a triage system that routes predictions into one of three tiers: low-confidence alerts that pass through without human review, medium-confidence alerts that require a single analyst check, and high-confidence alerts that demand a panel review. This prevents alert fatigue while ensuring the most consequential forecasts receive the scrutiny they deserve.

The workflow typically begins with the model outputting a probability score and a set of contributing features. An automated rule engine then applies thresholds to determine the review tier. For a coup prediction model, a score below 0.3 might pass through to a log file only, a score between 0.3 and 0.7 triggers a notification to a regional specialist who has 24 hours to confirm or override, and a score above 0.7 escalates to a three-person panel including a political scientist, an intelligence analyst, and a country expert. The panel must reach a consensus or document dissenting views before the forecast is considered actionable.

A documented failure mode in this process is the anchoring bias, where the human reviewer gives disproportionate weight to the model's initial probability. To mitigate this, the review interface should present the raw data features first, before showing the model's score. The reviewer should be required to write a brief contextual assessment that either supports or contradicts the model's logic. This forces the human to form an independent judgment rather than simply rubber-stamping the AI output. The Stimson Center’s Top Ten Global Risks for 2026 report notes that the short-term race for AI dominance itself poses financial risks that may affect global stability, adding another layer of complexity for reviewers who must consider whether the model's training data is already outdated by rapid geopolitical shifts.

An edge case that practitioners often overlook is the handling of contradictory signals from multiple models. If one model predicts a high probability of coup while another predicts low probability, the human-in-the-loop process should include a structured comparison step. The reviewer should examine which features are driving the divergence — for example, one model may weight military spending changes heavily while another weights social media sentiment. The panel should then decide which model's assumptions are more plausible given current on-the-ground reporting. This is not a vote but a reasoned judgment call, documented in a decision log that can be audited later.

A common mistake is treating the human review as a binary approve-or-reject gate. A more effective design includes a third option: "flag for monitoring." This allows the reviewer to acknowledge a credible but uncertain forecast without triggering a full response. The flagged alert then enters a watch list with a 30-day review cycle, during which the model's subsequent predictions for that country are compared against real-world events. This reduces the pressure on reviewers to make premature calls on ambiguous signals.

You can take a concrete action today: open the ACLED data portal, filter for a region of interest, and examine the raw event types and their date stamps. Note how many events are coded from local news sources versus international wire services, and consider how that reporting bias might affect a model trained on that data. This exercise takes about thirty minutes and reveals the data quality constraints that every instability forecast must navigate.

Statistical Models vs

Statistical models for instability forecasting differ from machine learning approaches primarily in their transparency and their handling of rare events. Traditional statistical models, such as logistic regression or survival analysis, specify an explicit mathematical relationship between input features and the probability of a coup. This means a reviewer can trace exactly how each variable — military spending changes, protest frequency, elite defections — contributes to the final probability. Machine learning models, particularly gradient-boosted trees or neural networks, often achieve higher predictive accuracy by capturing non-linear interactions that statistical models miss, but they do so at the cost of interpretability. A random forest model might correctly flag a high-risk scenario that a logistic regression overlooks, yet the reviewer cannot easily determine which combination of features drove the prediction.

The tradeoff between interpretability and accuracy becomes an ethical question when the forecast will inform real-world decisions. The reviewer can examine the historical data and note that this pattern held in only three of the past twelve coup cases, weakening the model's confidence. A black-box model offers no such recourse. The reviewer must either accept the output or reject it based on intuition, which introduces the anchoring bias described above. The UNESCO Recommendation on the Ethics of Artificial Intelligence, adopted by 193 member states, explicitly calls for transparency and human oversight in high-stakes AI systems, which includes instability forecasts that could trigger diplomatic or security responses.

Another structural difference lies in how each approach handles the extreme class imbalance inherent in coup prediction. Coups are rare events — fewer than 250 successful coups occurred globally between 1950 and 2020, against millions of country-month observations with no coup. Statistical models typically require explicit techniques to address this imbalance, such as weighting the minority class or using rare-events logistic regression. Machine learning models can sometimes learn from the imbalance directly through tree-based splitting, but they are also more prone to overfitting to the few positive examples in the training set. Precision measures the fraction of coup predictions that are correct, while false positive rate measures how often the model cries wolf. Both metrics are essential for evaluating whether a model is fit for operational use.

A documented failure mode specific to machine learning models is their sensitivity to distribution shift. A model trained on data from 1990 to 2015 may perform well on test data from that same period, but fail catastrophically when applied to 2026 conditions — for example, if social media dynamics or the role of foreign military aid have changed. Statistical models with fixed coefficients are also vulnerable to this, but their simpler structure makes the degradation easier to detect. A reviewer can compare the current distribution of key features against the training distribution and flag a warning if the model is operating outside its valid range. The Stimson Center’s Top Ten Global Risks for 2026 report notes that the short-term race for AI dominance itself poses financial risks that may affect global stability, a factor that no model trained before 2023 would have captured. This means any model used in 2026 must be retrained or recalibrated on recent data, and the choice between a statistical or machine learning approach should account for how frequently the organization can update the model.

You can take a concrete action today: open the Political Instability Task Force dataset, which is publicly available through Harvard's Dataverse, and compare the performance of a simple logistic regression against a random forest classifier on the same data. Use precision and false positive rate as your evaluation metrics, not accuracy. This exercise takes about two hours with standard Python libraries and will clarify which approach suits your specific use case. The key is to prioritize the metric that matters most for your decision context — if a false alarm triggers an unnecessary diplomatic intervention, you want a model with low false positive rate even if it misses some true events. If missing a real coup is the greater risk, you accept more false alarms and optimize for recall. No model eliminates this tradeoff; the ethical obligation is to make it explicit.

What Metrics Determine Real-World Utility of a Forecast?

The real-world utility of a political instability forecast is determined by its precision and false positive rate, not its accuracy. Precision measures the fraction of coup predictions that are correct, while false positive rate measures how often the model issues a false alarm. An organization deploying a forecast must decide which of these two metrics to prioritize, because they are in direct tension. A model tuned for high precision will issue fewer alerts, but may miss real events. A model tuned for low false positive rate will issue fewer false alarms, but may also miss real events if the threshold is set too conservatively.

The mechanism for this tradeoff is the decision threshold. A model outputs a probability score between 0 and 1 for each country-month observation. The operator chooses a threshold above which a prediction is considered a "coup alert." Lowering the threshold increases recall (more true events caught) but also increases the false positive rate. Raising the threshold does the opposite. The ethical obligation is to make this threshold explicit and to document the expected precision and false positive rate at that threshold before deployment. A 2022 article in International Security argues that AI increases the value of judgment in international politics, as prediction alone is insufficient without human interpretation and decision-making. The threshold choice is a judgment call that cannot be automated away.

Edge cases arise when the cost of a false alarm is asymmetric. If a false alarm triggers an unnecessary diplomatic intervention, sanctions, or military alert, the cost may be high in both financial and reputational terms. In that scenario, a model with a very low false positive rate is preferable, even if it misses some events. If missing a real coup could lead to a humanitarian crisis or regional war, the operator should accept a higher false positive rate and invest in a human review process to filter the alerts. The Stimson Center’s Top Ten Global Risks for 2026 report notes that the short-term race for AI dominance itself poses financial risks that may affect global stability, a factor that no model trained before 2023 would have captured. This distribution shift means that precision and false positive rate must be re-evaluated on recent data, not just on a historical test set.

A common practitioner mistake is to report only the area under the receiver operating characteristic curve (AUC-ROC) as a summary metric. AUC-ROC measures the model's ability to rank positive cases higher than negative cases, but it does not tell you the actual false positive rate at a given threshold. Two models with the same AUC-ROC can have very different precision at the same recall level. You should instead report a precision-recall curve or a table of precision and false positive rate at several candidate thresholds. UNESCO promotes ethical AI through global recommendations that guide responsible design, development, and use of artificial intelligence, which can apply to political instability forecasting systems. These recommendations emphasize transparency and accountability, which directly apply to the choice of evaluation metrics.

How to Implement a Tiered Alert System for Instability Warnings

A tiered alert system for instability warnings works by mapping a model's continuous probability output onto three discrete action levels: monitoring, warning, and action. Each tier triggers a distinct review protocol and response pathway, which prevents both overreaction to low-confidence signals and inaction on high-confidence ones. The system requires three explicit probability thresholds, each chosen based on the cost of false alarms versus missed events for that specific deployment context.

The monitoring tier typically covers probabilities below 0.3. At this level, no direct action is taken. The output is logged and reviewed periodically, often weekly, by a designated analyst. The warning tier, usually set between 0.3 and 0.6, triggers a structured human review. A small team examines the model's feature weights, checks for recent data shifts, and produces a qualitative assessment within 24 to 48 hours. The action tier, for probabilities above 0.6, escalates to a senior decision-maker or committee. This tier may activate pre-planned diplomatic, humanitarian, or security protocols, depending on the organization's mandate.

The thresholds themselves must be set using a precision-recall analysis on recent validation data, not on a historical test set from a different political era. As noted above, the distribution shift caused by new global dynamics means that a model's performance at a given threshold can degrade quickly. The ACM FAccT 2026 conference accepted papers include research on how middle managers in AI development experience ethical uncertainty, which is directly relevant to the governance of these forecasting tools. That work underscores that the people setting the thresholds and reviewing the alerts need clear ethical guidelines, not just statistical targets.

An edge case that practitioners often miss is the self-fulfilling prophecy risk. A public or semi-public alert at the action tier can itself become a signal that accelerates instability. If a government or media outlet learns that an AI model has flagged a high probability of a coup, that information may trigger capital flight, military movements, or preemptive arrests, all of which can increase the actual likelihood of a coup. The tiered system should therefore include a classification of alert sensitivity. Alerts at the action tier may need to be restricted to a small, trusted group and paired with a communication protocol that prevents leakage. The Stimson Center's Top Ten Global Risks for 2026 report notes that the short-term race for AI dominance itself poses financial risks that may affect global stability, a factor that no model trained before 2023 would have captured. This reinforces the need for a human review layer that can incorporate qualitative context the model cannot see.

A common mistake is to treat the three tiers as static. They should be recalibrated at least quarterly, or after any major political event in the target country. The precision and false positive rate at each threshold must be re-estimated on the most recent six months of data. If the false positive rate at the action tier rises above an acceptable level, the threshold should be raised, even if that means fewer alerts. Accountability interventions in states forecasted as high risk for instability are considered key to prevention, but societal systems often still fail due to both commission and omission, as noted in analysis from The Cipher Brief. A tiered system reduces the risk of omission by ensuring that no high-probability event is ignored, and reduces the risk of commission by forcing a review before any action is taken.

You can implement a basic version of this system today using any supervised learning model that outputs probabilities. Train a logistic regression or random forest classifier on the Political Instability Task Force dataset, available through Harvard's Dataverse. Generate a precision-recall curve and select three thresholds that correspond to a precision of at least 0.2 for monitoring, 0.4 for warning, and 0.6 for action. Document these thresholds and the expected false positive rate at each one. Then set a calendar reminder to re-evaluate the thresholds every three months. This workflow takes about two hours with standard Python libraries and gives you a working governance structure that can be adapted to any forecasting model.

What to do next

This guide has outlined the technical capabilities and ethical pitfalls of AI-driven instability forecasting. To engage with this topic critically, readers should move beyond the headline and examine the underlying models, their track records, and the governance frameworks that attempt to keep them accountable. The following steps offer concrete ways to verify claims, compare methodologies, and stay informed on emerging standards.

Step	Action	Why it matters
1	Review the precision and false positive rate of any published instability model you encounter. Check the model's documentation or the original research paper for these metrics.	High false alarm rates can erode trust and trigger unnecessary political or economic reactions. Precision tells you how often a "coup predicted" alert is actually correct.
2	Compare the methodology of two different forecasting models—for example, the Political Instability Task Force (PITF) model and a newer machine-learning approach like those from ACLED or the VIEWS project.	Different data sources, feature sets, and algorithms can produce wildly different probabilities for the same country. Understanding these differences reveals the limits of any single forecast.
3	Read the UNESCO Recommendation on the Ethics of Artificial Intelligence, available at unesco.org. Focus on the sections about transparency, accountability, and human oversight.	This is the most widely endorsed global framework for AI ethics. It provides a benchmark against which any specific forecasting system can be evaluated.
4	Set a calendar reminder to check the ACM FAccT conference proceedings (facctconference.org) each year for new papers on fairness, accountability, and transparency in predictive models.	FAccT is the leading venue for research on the societal impacts of AI, including work directly relevant to political instability forecasting and its governance.
5	Verify the source of any "AI predicts coup" news story by tracing it back to the original research organization (e.g., a university lab, a think tank like the Stimson Center, or a government agency).	Media coverage often exaggerates or oversimplifies model outputs. The original report will contain the actual confidence intervals, data limitations, and caveats that are stripped from headlines.
6	Read the 2022 International Security article "Prediction and Judgment: Why Artificial Intelligence Increases the Importance of Humans in War" by Horowitz and others (available via MIT Press Direct).	This paper makes the core argument that AI augments—but does not replace—human judgment in high-stakes political decisions, a thesis directly relevant to the ethics of coup forecasting.

Also worth reading: The Rise of Robot Ethics Education How Engineering Ethics Shapes Our Technological Future · 7 Ways AI Time Series Forecasting is Transforming Entrepreneurial Decision-Making in 2025 · The Bot and the Beautiful: How AI is Impacting Everything from Art to Ethics · The Ethics of Journalistic Neutrality Examining the Piers Morgan-Abby Martin Debate

Sources: unesco, plos, researchgate, gsdrc, mit