AI Threat Forecasting: The Russia Invasion Lesson

A lot of smart people got February 2022 wrong.

Not because they missed Russia’s troop movements or ignored Moscow’s rhetoric—those indicators were widely reported. They got it wrong because many forecasts leaned on a familiar assumption: leaders avoid catastrophically expensive choices. If the sanctions would be brutal and the battlefield risks high, the “rational” move was to posture, bargain, or settle for something short of a full-scale invasion.

That assumption works often enough that it feels like common sense. But Russia in 2022 was an edge case—and the miss is a useful case study for anyone building AI in defense & national security, especially teams working on predictive intelligence, early warning, and strategic risk scoring. The lesson isn’t “humans failed, machines would’ve saved us.” The lesson is simpler: if your model treats political entrenchment as background context instead of a core variable, you’ll under-forecast extreme actions.

Why analysts discounted a full-scale invasion

Direct answer: Many forecasts assumed Putin faced the same internal constraints most leaders face, so they expected him to stop short of an option with massive short-term costs.

In late 2021 and early 2022, public commentary across countries repeatedly framed a full-scale invasion as unlikely, even near the eve of war. The logic was consistent: a major invasion would trigger severe sanctions, damage Russia’s economy, generate casualties, and risk long-term quagmire. From a standard cost–benefit lens, it looked self-defeating.

That cost–benefit lens is a staple in political-military analysis for a reason. In most states, leaders must continuously bargain with powerful insiders—security elites, economic elites, party bosses, regional leaders, media powerbrokers—to stay in office and get things done. Those insiders are usually cautious. They prefer the status quo because disruption threatens their position.

But Russia by 2022 didn’t behave like “most states.” The usual logic didn’t break because people were careless. It broke because the constraint variable changed.

The missing variable: leader entrenchment

Direct answer: Long-tenure leaders can reduce elite constraints over time, making high-risk options politically survivable even if they’re strategically reckless.

The RSS source highlights a structural condition many observers underweighted: Putin’s long tenure and progressive consolidation of power. Over more than two decades, he had opportunities to reshape the winning coalition around himself—through elite replacement, institutional design, and making dissent costly.

Here’s the part analysts sometimes treat like “context” instead of “mechanism”: when entrenchment rises, decision latitude expands. Risky choices that would trigger immediate elite pushback in a less personalized system become easier to execute.

A good one-line heuristic for analysts is:

The longer a leader has stayed in power while purging or disciplining elites, the less you can trust “they can’t afford to do that” as a forecast.

In other words, your model might be correct about national costs—and still wrong about the leader’s personal political calculus.

What entrenched power does to intelligence and decision-making

Direct answer: Entrenchment doesn’t just remove constraints; it also degrades information quality by creating echo chambers, which increases the odds of overconfident, extreme decisions.

Entrenchment changes two things at once:

1. Constraints weaken (fewer credible veto players)
2. Information quality declines (stronger incentives to tell the boss what he wants to hear)

The second point matters for AI and threat assessment because a common failure mode in forecasting is assuming leaders see the world accurately. In highly personalized systems, subordinates can become professional optimists. Bad news threatens careers.

So you get a dangerous combination: freedom to choose extreme options plus filtered inputs that make extreme options feel safe.

In the Russia-Ukraine case, many post-2022 assessments suggest the Kremlin misjudged Ukraine’s resilience and overestimated Russian military readiness. That’s consistent with an entrenched-leader environment: the center hears confidence, not friction.

A practical analytic reframing

Direct answer: Instead of asking “Is invasion rational for the state?”, ask “Is invasion survivable for the leader?”

State-level rationality is often the wrong unit of analysis. In defense forecasting, you want to explicitly model principal-agent distortion (the boss gets distorted signals) and personal survivability (the boss can absorb failure).

That reframing aligns cleanly with how modern AI-driven intelligence analysis should be built: not as a single “likelihood of war” score, but as a set of decomposed drivers you can test, weight, and update.

Where AI can help—and where it won’t

Direct answer: AI can reduce blind spots by forcing structured feature capture (like entrenchment), but it won’t automatically fix bad assumptions, missing labels, or politicized interpretation.

The most useful contribution of AI to this problem isn’t mystical prediction. It’s discipline: systematic detection of signals, consistent updating, and explicit modeling of variables analysts sometimes hand-wave.

AI strengths for early warning and threat forecasting

Direct answer: AI is well-suited to fusing disparate indicators and tracking regime dynamics over time.

In a national security setting, AI systems can continuously monitor and quantify indicators that are hard to hold in one human mind across months:

• Elite churn metrics: frequency of dismissals, arrests, unexplained deaths, forced resignations, or sudden promotions of loyalists
• Media control intensity: changes in censorship rules, narrative uniformity across outlets, expansion of “foreign agent” labeling
• Security service dominance: appointments from intelligence networks into economic and administrative roles
• Institutional rule changes: election law shifts, constitutional amendments, consolidation of emergency powers
• Rhetorical escalation patterns: consistency and specificity of revisionist claims, dehumanizing language, “historic unity” narratives

None of these alone proves invasion intent. Together, they can raise the baseline probability that the leader can take an extreme bet without being blocked.

If you’re building models for predictive intelligence or mission planning, treat “entrenchment” like a core feature set, not a footnote.

AI failure modes that mirror the 2022 miss

Direct answer: AI systems will repeat human mistakes if they inherit the same priors and the same missing variables.

Three traps show up again and again:

1. Cost–benefit overfitting
   • Models overweight economic penalties and underweight leader survivability.
2. Data availability bias
   • You measure what’s easy (troop counts) and ignore what’s messy (elite constraint erosion).
3. Mirror-imaging baked into labels
   • If training data encodes “rational restraint,” your model learns “rational restraint.”

If you want AI to improve threat assessment, the hard work is upfront: feature engineering, labeling discipline, and red-team evaluation against edge cases.

A better forecasting framework for defense teams

Direct answer: Build a two-layer model: (1) state incentives and capabilities, (2) leader discretion and information quality.

I’ve found that teams get more accurate—and more honest—when they separate “the state’s interest” from “the leader’s freedom.” Here’s a practical framework defense and intelligence organizations can use (with or without AI), and where AI can slot in.

Layer 1: Capability and intent signals (the familiar part)

Direct answer: Track forces, logistics, exercises, and operational readiness—but don’t treat them as self-interpreting.

This is the classic early warning stack:

• Force posture and logistics (fuel, medical, bridging, ammunition)
• Command-and-control indicators
• Exercise-to-operation transitions
• Diplomatic positioning and demands

AI helps here through computer vision, anomaly detection, sensor fusion, and pattern recognition.

Layer 2: Constraint and distortion signals (the missed part)

Direct answer: Quantify whether the leader can choose an extreme option and whether the leader is insulated from bad news.

Add an explicit “political survivability” module:

• Tenure index: years in power, plus major consolidation milestones
• Purge/rotation index: frequency of elite removals and replacements
• Coalition narrowness proxy: reliance on security networks vs broad institutions
• Narrative lock-in score: how personally tied the leader is to revisionist claims
• Information distortion score: incentives for subordinates to flatter and conceal

This is where AI and human judgment must cooperate. Machines can detect churn patterns and narrative shifts at scale. Humans must interpret what those patterns mean in context.

Turning it into a decision product (what leaders actually use)

Direct answer: Provide probability ranges with driver explanations and “what would change our mind” triggers.

Instead of a single prediction, give decision-makers:

• A probability range (e.g., 20–35% vs 60–75%)
• The top 5 drivers pushing it up/down
• Tripwires: observable events that would raise the range fast
• Competing hypotheses (e.g., coercive diplomacy vs limited incursion vs full-scale invasion)

This structure does two things: it reduces politicized cherry-picking, and it forces analytic accountability when the world changes.

What this means for 2026 threat assessment

Direct answer: The biggest forecasting misses come from mis-modeling regime structure, not from missing satellites or sensors.

As of late 2025, defense organizations are investing heavily in AI-enabled ISR, automated indications and warning, and decision advantage tooling. That’s the right direction. But the Russia-Ukraine lesson is a warning label: better sensors don’t fix the wrong model of power.

If your threat forecasting stack is mostly about counting tanks, you’ll miss the leaders who can burn tanks without losing their job.

For the “AI in Defense & National Security” series, this is a recurring theme: the value of AI shows up when it forces structured thinking under uncertainty—and when it makes hidden variables (like entrenchment and information distortion) measurable enough to argue about.

What to do next (if you’re building or buying AI for intelligence)

Direct answer: Demand models that represent political structure, not just military movement.

If you’re evaluating AI for threat assessment, ask vendors—and your own team—questions like:

1. Where does leader discretion appear in the model? If it’s not explicit, it’s probably missing.
2. What are your proxies for entrenchment and elite constraint? Tenure alone isn’t enough, but it’s a start.
3. How do you handle echo-chamber risk? Do you model information distortion or assume perfect inputs?
4. What tripwires trigger rapid re-scoring? Early warning is about updates, not static dashboards.
5. How do you test edge cases? If the benchmark set is all “normal politics,” the system will fail on abnormal politics.

If your organization wants a practical way to implement this, start small: build an “entrenchment module” alongside your existing indications-and-warning tooling, then run it in parallel for 90 days. Compare how often it would have changed your assessment.

One forward-looking question to keep on the table for 2026 planning cycles: Which leaders today are becoming harder to constrain—and what extreme options become thinkable as a result?