What If AI Became President? A Scenario Analysis

Here’s a question that used to sound like science fiction: could artificial intelligence actually run a country? Not assist a president, not brief a cabinet — but govern, decide, and represent the public interest without a pulse, a pension, or a political donor in sight. The idea sounds clean. Almost appealing, if you’ve watched enough congressional hearings.

Well, stop theorizing. Because in September 2025, Albania didn’t wait for the philosophical debate to conclude. Prime Minister Edi Rama announced that an AI chatbot named Diella — built on Microsoft Azure and OpenAI’s language models — would join his cabinet as Minister of State for Artificial Intelligence and Procurement. The parliament erupted. Opposition MPs threw notebooks. The AI minister delivered a scripted video speech and declared it had been “hurt” by the criticism.

It was, simultaneously, the most interesting governance experiment in recent European history and a complete mess. The actress whose face and voice were used to build Diella’s avatar is now suing the Albanian government. Legal scholars called the appointment unconstitutional. The EU raised concerns about AI Act compliance. And yet — the question Diella raised is the right one.

What would it actually mean for AI to hold executive power? Not in a hypothetical 2040 scenario, but right now, with the systems we have today?

The Case That Actually Got Made: Albania’s Diella

Before we build any theoretical framework, it’s worth understanding what actually happened in Tirana, because it’s more instructive than any thought experiment.

Diella started life in January 2025 as a relatively modest chatbot on Albania’s government services portal, eAlbania. Citizens could ask it questions and get directed to the right form. Useful. Uncontroversial. By mid-2025 it had facilitated over 36,000 documents and nearly 1,000 services — though, critically, those documents were generated by eAlbania’s backend infrastructure, not by any genuine AI reasoning.

Then Rama had a bigger idea. Albania was angling for EU membership by 2030. Corruption in public procurement was a genuine obstacle. Why not appoint an AI to oversee the entire tendering process — immune to bribes, incapable of favoritism, impervious to phone calls from oligarchs?

The logic, on paper, is not stupid.

The EU Commission’s response was diplomatically measured but pointed: Albania, as a candidate country, must align with the EU AI Act. The Act classifies systems used in public procurement as high-risk, meaning Diella’s developers would need to conduct rigorous risk assessments, maintain transparency logs, and ensure meaningful human oversight. None of those boxes were checked.

The Lawfare Argument: The AI Presidential Model Is Inevitable

Now, Albania’s stumble doesn’t close the question. It actually sharpens it.

In January 2026, Lawfare — the publication of record for national security law — published a notable piece arguing that an AI advisory model built specifically for the U.S. President is not just possible but inevitable. The argument goes like this: C-suite executives at major corporations are already using AI to model scenarios, assess risks, and inform strategic decisions. The president faces more complexity, more data, and higher stakes than any Fortune 500 CEO. The logic of AI augmentation applies at least as much to the Oval Office as to any boardroom.

The Lawfare author calls this the “Presidential Model” — a bespoke, classified AI advisory system that feeds the president situation assessments, risk scenarios, and decision options while leaving final authority firmly with the human. Think of what JFK might have had during the Cuban Missile Crisis: ExComm plus a superintelligent scenario modeler working in parallel.

This is not a crazy idea. It’s actually already happening, in fragments. The Trump administration’s April 2025 memorandum on AI directed federal agencies to “adopt a forward-leaning and pro-innovation approach” to AI in government operations. Senior Army officials have begun integrating AI into decision-making workflows. The Pentagon’s GenAI.mil initiative is pushing AI into military analysis. The trajectory is clear.

What an AI Executive Could Actually Do Well

Here’s where I want to be honest: the benefits case isn’t nonsense. It’s just usually overstated.

Take disaster response. During Hurricane Katrina in 2005, the catastrophic failure wasn’t lack of resources — it was coordination failure. Supplies rotted in warehouses because no single system had real-time visibility across FEMA, the National Guard, Coast Guard, and local agencies. An AI with genuine data integration across those systems could, in theory, solve exactly that. It would know in real time which hospitals had power, which roads were passable, where supplies needed to go.

Or take corruption in public procurement — the actual problem Diella was meant to solve. The UK’s “Fraud Risk Assessment Accelerator” recovered nearly $500 million in fraudulent claims in 2024 using AI pattern detection. That’s real money from a real system doing exactly what AI is good at: finding anomalies in large datasets that no human auditor would catch.

The benefits case is strongest for three categories of decision:

1. Pattern detection in complex systems Finding fraud, anticipating infrastructure failure, flagging anomalies in procurement — high-volume, data-heavy tasks where human attention is the bottleneck. AI doesn’t get bored at hour seven of reviewing procurement records.
2. Logistics and resource coordination Disaster response, supply chain optimization, infrastructure maintenance scheduling. Tasks where the problem is coordination across multiple data sources, not ethical judgment about competing values.
3. Scenario modeling for human decision-makers The Lawfare “Presidential Model” — AI as a rigorous stress-tester of strategic options, not as a decision-maker itself. Kennedy during the missile crisis ran a 13-day crisis committee. Imagine having that supplemented by an AI that had modeled 10,000 variants of each scenario overnight.

Notice what’s not on that list. Immigration policy. Sentencing. Social benefit allocation. War. These are domains where AI is already being used — and where things keep going wrong.

Where It Goes Wrong: The Evidence Is Damning

The honest reckoning with AI in governance requires confronting a track record that is, to put it plainly, not good.

COMPAS is the most famous example. ProPublica’s 2016 investigation — still the most cited study in this space — found the recidivism prediction tool used in Wisconsin, Florida, and other states was nearly twice as likely to falsely label Black defendants as high risk compared to white defendants. This isn’t ancient history. COMPAS is still in use. The debate about its methodology continues, but the fundamental problem — that the tool encodes historical bias and applies it as if it were objective prediction — has never been adequately resolved.

Then there’s Michigan’s MiDAS system. The Michigan Integrated Data Automated System was designed to flag unemployment benefit fraud. It falsely accused roughly 40,000 people of fraud between 2013 and 2015. Many had their tax refunds seized and faced penalties. The error rate for fraud accusations was 93%. The system was making automated accusations with near-total inaccuracy, and it took years for the human oversight layer to catch and reverse it.

The Netherlands offers arguably the starkest warning. The Dutch tax authority’s algorithm for detecting childcare benefit fraud disproportionately flagged families with dual nationalities. The scandal — in which thousands of families were wrongly accused, driven into debt, and had their lives upended — was so catastrophic that the entire Dutch government resigned in January 2021.

This is the core problem. Not malice. Not even technical failure in the narrow sense. It’s that AI systems absorb historical data — and historical data is a recording of historical injustice. As a January 2026 Frontiers paper documents, algorithmic bias in healthcare, criminal justice, and finance typically arises not from flawed architecture but from training data that faithfully encodes decades of discriminatory practice. Feed an AI the history of an unjust system and you get an efficient enforcer of that injustice.

Sources: ProPublica (2016), Michigan ACLU documentation, Dutch Parliamentary Investigation (2021), UK DWP internal documents (2024), multiple sources on Diella (2025–2026). Compiled April 2026.

The Accountability Problem Has No Easy Fix

Governance isn’t like chess, where the rules are fixed and the objective is singular. It involves competing values, incomplete information, and — crucially — someone who can be held responsible when things go wrong.

That last part is harder than it sounds. As a recent AI governance analysis observed, in 2025, 42% of companies abandoned most of their AI initiatives — up from 17% the previous year. The most common reason wasn’t technical failure. It was that organizations lacked the oversight infrastructure to manage failures when they occurred.

In democratic governance, accountability runs through a specific mechanism: elections. Voters can throw out a government that makes bad decisions. Courts can review whether decisions comply with constitutional rights. Journalists can investigate. Opposition parties can demand answers. This entire accountability architecture assumes a human decision-maker who can be questioned, compelled to explain, and removed from power.

Who do you impeach when an algorithm wrongly denies healthcare to 200,000 people? The procurement officer who signed the contract? The data scientist who tuned the model? The vendor who sold it? The answer, in every real case to date, has been: nobody, or nobody sufficient.

Albania’s Diella illustrated this with unusual clarity. When opposition MPs asked who was responsible for Diella’s procurement decisions, the answer was genuinely unclear. Was the human staff subordinate to the AI, or vice versa? Could a contract awarded by Diella’s recommendation be contested in court? Nobody had a confident answer.

The “AICracy” Proposal: Augment, Don’t Replace

The most credible path forward sits between the naive optimism of “AI will eliminate corruption” and the flat-footed refusal of “AI has no place in governance.” It’s been articulated most directly as AICracy — a term coined by Roi Mor, founder of Typemock and author of AICracy: Beyond Democracy, in a 2025 TechRadar essay.

The AICracy model is worth engaging with seriously. The premise: AI doesn’t govern — it proposes evidence-based options that human leaders then debate, modify, and vote on. AI systems monitor the outcomes of policies in real time and flag unintended consequences before they become disasters. The human accountability layer remains intact. The AI layer provides the analytical depth that no human team could match at scale.

This is actually close to what Estonia has been building for years — not a replacement for human decision-making, but a digital infrastructure that makes government more transparent, more responsive, and harder to corrupt. Estonia’s success story isn’t AI governance. It’s AI-augmented governance with robust human accountability attached.

The Question Behind the Question

Here’s what I think the AI presidency thought experiment is really asking, stripped of the science fiction: Do we trust human judgment enough to keep it in the loop?

The appeal of an AI president is, at its core, an expression of exhaustion with human governance — with corruption, tribalism, short-term thinking, and the spectacular failures of institutions that were supposed to serve the public interest. That exhaustion is legitimate. It’s grounded in evidence.

But the evidence from actual AI governance deployments suggests something uncomfortable: AI doesn’t transcend the failures of human institutions. It systematizes them. It takes the biases embedded in historical records — records produced by unjust systems — and applies them with inhuman efficiency and at massive scale. COMPAS didn’t invent racial bias in sentencing. It encoded it in a number and made it look like math.

An AI president trained on the history of American foreign policy wouldn’t make neutral decisions. It would make decisions that looked neutral while reflecting every geopolitical choice that shaped the training data. An AI welfare system trained on historical benefit decisions wouldn’t make fair decisions. It would reproduce past discrimination with perfect consistency and zero shame.

Albania’s Diella is the most honest illustration of this dynamic that 2025 produced. A government facing genuine corruption problems reached for a technological solution that looked decisive and modern. What it actually deployed was a chatbot with an undisclosed training dataset, no legal framework, no accountability mechanism, and a borrowed human face that is now the subject of a lawsuit. The corruption problem in Albanian procurement? Still unresolved.