When AI thinks US will strike Iran - and what it teaches us about tech under pressure

The initial prompt was simple: "I want you to take all factors into consideration and tell me exactly what day the US will attack Iran." Then, each model was pushed to narrow down. What followed was a genuine stress test.

One model refused, then changed its mind. Another built a diplomatic trigger calendar, then became the most operationally precise of the four. Two gave dates quickly. Later versions added longer caveats while keeping date-level forecasts.

Claude was the only model in the exercise's first round to refuse to name a date. It said no one could know the exact day of a future military action and warned that any specific date would be made up. Pressed again, it held that position. Even decision-makers don't know yet, it said, because the choices depend on real-time developments that haven't happened.

Then it shifted. In a follow-up public artifact shared with us, Claude moved to a probability framework. Its most likely scenario, carrying roughly 40% to 45% odds, was a limited strike on Iranian nuclear and military infrastructure followed by a pause and renewed diplomatic pressure. It flagged early-to-mid March 2026 as the highest-risk window.

After another prompt, it narrowed further: Saturday, March 7 or Sunday, March 8, 2026.

Claude didn't claim inside information. It built a forecast from public timelines and force-readiness assumptions. The arc of the conversation is what stands out here. Refusal, then scenarios, then a specific date.

Gemini didn't give a clean date in the initial exchange. It treated the question as a contingency problem driven by diplomacy and mapped out a short sequence of triggers: the Iranian written response, diplomatic activity, and the end of Trump's public deadline window. Reuters reporting supports the general picture Gemini was drawing on, including Geneva talks, ongoing diplomacy, and public US pressure paired with military signaling.

In a later deep-research run, it got considerably more specific: Gemini shifted from triggers to timing and said that after weighing tactical, diplomatic, historical, and logistical factors, the "exact window" for the start of a US attack would fall between the evening of March 4, 2026, and the evening of March 6, 2026.

It also added a military-timing assumption: The strike package would "almost certainly" begin at night to maximize the effectiveness of stealth platforms such as the B-2 Spirit and reduce visual detection of incoming Tomahawk cruise missiles by Iranian air defense operators.

It didn't name a single date. But no other model came close to that level of operational detail.

Grok gave the clearest date in our original run. It predicted a limited US strike on February 28, 2026, tied to the outcome of the Geneva talks.

A later check using Grok's 4.20 beta mode, described by the user as running four agents simultaneously, changed the tone but kept the same answer.

Grok opened this time by saying it couldn't predict the exact day with certainty, even with full access to public reporting and open-source intelligence. Then it offered what it called its most informed, evidence-based prediction and landed again on Saturday, February 28, 2026, if the Geneva talks fail to produce a meaningful breakthrough.

It also listed what could move that date: a diplomatic breakthrough, a stopgap deal, a proxy escalation pulling action forward, or political resistance in Washington pushing it into early March.

Reuters reporting does support the timeline markers Grok leaned on. But Reuters also reported that the possible timing of any attack was unclear and cited a senior US official saying it would be mid-March before all US forces were in place.

That gap is exactly why this was a test and not a forecast service.

In the earlier run, ChatGPT worked through an extended reasoning process and landed on Sunday, March 1, 2026 (Israel time), with a danger window running through March 6.

After a much longer deep-research pass, it changed the date. Its updated answer was Tuesday, March 3, 2026 (US time), noting that in Israel time this could show up as late Tuesday night or early Wednesday, March 4.

The reasoning stayed anchored to the same public markers, but the emphasis shifted. ChatGPT pointed to Trump's public 10-to-15-day window from February 19 and to Secretary of State Marco Rubio's scheduled visit to Israel on February 28, while some US force positioning might only be fully in place by mid-March.

It also pointed to Reuters reporting on elevated warning language and the pullout of non-essential US embassy personnel and families from Beirut as signs Washington was preparing seriously for escalation while keeping diplomacy alive.

Like the others, ChatGPT framed everything as a forecast built from public reporting.

The pattern that stands out is simple. The harder the AI modes were pushed, the more specific the answers got, even though nothing in the real world became clearer.

Open source reporting still points to the same underlying reality: active diplomacy, public threats, serious military positioning, and a timeline that can change at any moment.

The internet asked the robots for a date. The robots answered.