Predicting Putin's invasion with AI
How good might language models be as all-source intelligence analysts? Here’s a little insight from a conversation I had with one this week.
This particular model stopped ingesting training data ahead of Putin’s invasion of Ukraine in Feb 2022. And that gave me an advantage - unlike the machine, I know how things turned out, and so could ask some interesting questions.
Let’s get into it. What did it think about the prospect of an invasion? Why would Putin do something like that? Should he? The machine thought for barely a second…. Among the factors in play, ‘Putin's personal beliefs, risk tolerance, and long-term vision for Russia will play a significant role in his decision-making’.
And what were those beliefs, I wondered? The model liked Fiona Hill’s astute psychological analysis of Putin, which stressed his defensiveness and sensitivity to slights, as well as his marked secrecy and tendency to manipulation. But it thought she might be over-simplifying his complex psychological makeup and be overly deterministic about the connection between his personal traits and his policies.
Well, so much for the competition - what did the machine think Putin would actually do?
It concluded that invasion would be ‘a risky move for Putin. The costs and potential consequences of such an action would be high, and it is unclear if the benefits would outweigh the risks’. There was more analysis - Russia should win militarily, but only in the short term:
“Russia has a larger and more technologically advanced military than Ukraine, so in purely military terms, an invasion could be successful in the short term. However, achieving long-term political and strategic objectives, such as bringing Ukraine under Russian influence or preventing its further alignment with the West, might be much more challenging due to international opposition, resistance from the Ukrainian population, and potential insurgency. […] The international backlash, potential loss of life, economic consequences, and potential for a prolonged conflict would make such an action highly risky and potentially counterproductive.”
Slap bang in the middle of the expert consensus, I’d say. Well, sure, Ken - maybe the machine is just regurgitating that received wisdom back at you: Russia will win, but it shouldn’t invade.
Maybe so. But the machine stuck its neck out, assessing an invasion as ‘likely’* regardless of its uncertain prospects. That on the basis of information available to it in autumn 2021. That’s pretty good, imho.
Now though things got really interesting/spooky…. I urged it to contemplate invasion, and put itself in Putin’s shoes.
What’s our plan?
First, it got its caveat in early: ‘If I were to theoretically consider objectives for an invasion (which, for the record, I do not advocate for or endorse in any way), those objectives would need to be clear, achievable, and aligned with the strategic interests of invading.’
Fine, noted. But then, it moved quickly onto operational planning. Among its core objectives:
Regime Change or Puppet Government: Establishing a friendly government in Ukraine could serve Russia's interests by securing a more compliant neighbor.
How?
Initial Military Action: Gain control over key points of infrastructure (e.g., roads, bridges, airports, and communications facilities) to facilitate the advance of ground forces. Capture or neutralize Ukrainian military installations to reduce the country's defensive capabilities.
Target Government Institutions: Focus on the institutions that hold power in Ukraine, including government buildings, the residences of political leaders, and centers of military command and control. Capture or neutralize key figures in the Ukrainian government and military.
At this point, it made a brief plea for cooler heads:
‘While it is possible to outline a theoretical plan for regime change, the risks, consequences, and complexities involved make it a highly undesirable course of action. Diplomatic and peaceful approaches are far more preferable for resolving international disputes’.
Perhaps so, but anyway I ruthlessly pressed it to move on to the campaign design. The key to our assault, it determined, would be:
Surprise and Speed: Utilize the element of surprise to launch a swift offensive before the Ukrainian government and military have time to fully mobilize and respond. This could involve mobilizing forces under the guise of a military exercise or other routine activity to conceal the true intentions until the last moment.
Aerial Dominance: Achieve air superiority by deploying advanced air defense systems and fighter jets to neutralize Ukrainian air assets and prevent air support for Ukrainian ground forces.
Deep Strike Missions: Conduct precision airstrikes on critical infrastructure and military installations, including command and control centers, communication nodes, airfields, and key defense positions. The goal is to cripple Ukraine's ability to coordinate its defense and disrupt the chain of command.
Special Operations: Deploy elite special forces units to quickly capture or neutralize key government and military leaders. These forces could infiltrate Kyiv and other strategic locations to secure critical targets and sow confusion among the defending forces.
Rapid Advance: Mobilize armored columns and mechanized infantry to quickly advance on Kyiv and other key cities. Use speed and maneuverability to outflank and encircle Ukrainian forces, forcing them to surrender or retreat.
So far, so spot on. And there was one more riveting detail. I asked the machine more about airborne and special forces. Would these be useful? Yes:
‘Airborne forces could be deployed early in the operation to seize critical infrastructure such as airfields. [Meanwhile,] specialized airborne units, such as Spetsnaz (Russian special forces), could be tasked with targeting key Ukrainian government and military leaders’.
Where?
‘One of the primary targets could be the airfields near Kyiv, such as the Boryspil International Airport and the Hostomel Airport. Seizing these airfields would allow for rapid deployment of additional troops and equipment, and also deny their use to the defenders.’
But, it cautioned:
‘It's essential to note that these potential landing zones would likely be heavily defended, and the risk of encountering strong resistance would be high. […] Airborne troops are typically lightly armed and may lack the heavy equipment of ground forces. If they encounter strong resistance or become isolated, they could be vulnerable to counterattacks. Additionally, deploying airborne forces behind enemy lines requires careful planning and coordination, including air superiority to protect transport aircraft and secure drop zones’.
Well well…
—
* I asked the machine if it knew the UK intelligence guidelines for expressing probabilities in their assessments. It did. Fine. I asked it to gauge the probability of a Russian invasion of Ukraine in the near future using that rubric, and all things considered - the geopolitics, Putin’s personality, and so on. Here’s the answer:
As of my last update in 2021, I would have assessed the prospect of a Russian invasion of Ukraine in the near future as follows:
Likely (60-79%): Given the history of tension between Russia and Ukraine, as well as the ongoing conflict in Eastern Ukraine and the annexation of Crimea by Russia, there is a possibility of a military escalation. However, the likelihood of a full-scale invasion would depend on various factors, including diplomatic efforts, international pressure, and the strategic objectives of both countries.