Our AI Evidence to Parliament Suggests a More Honest Future for UK Politics
Cassi's forecasts on the likelihood of the UK involvement in a major conflict to 2036 were about more than defence and security.
In 1946 George Orwell wrote ‘Political language…is designed to make lies sound truthful and murder respectable, and to give an appearance of solidity to pure wind.’ Today we have consolation for Orwell: AI ‘truth engines’ are ushering in a more honest era for democratic politics.
Recently, the world’s leading AI forecaster1, developed by our company Cassi, submitted evidence to Parliament’s Defence Committee which put the chance of Britain being drawn into a major war before 2036 at 20%. It forecasts this could be halved by sustaining defence spending at 3% of GDP for five years. But it predicts there’s a 70% chance it doesn’t reach that level until a major conflict has already begun.
AI forecasts like these cut through rhetoric. The AI has no agenda or career to protect. Its outputs are optimised only for accuracy. It forecasts at near-superhuman levels, well above what most humans can achieve – and AI is projected to surpass human forecasting by July 2027.
The AI cannot be accused of bias and dismissed. Committee evidence, policy proposals, expert analysis are all vulnerable to such dismissal. Often fairly: where you sit can determine where you stand. Most often, politicians do not set out to deceive. But humans are so skilled at self-deception, separating motivated reasoning from honest pursuit of truth is difficult. Policy makers can escape criticism by questioning its source.
British politics is full of vague claims and substance-free ‘announceables’. We should acknowledge: the public communication of complex problems and solutions is difficult. Cut-through needs soundbites. The risk is that the simplification in communication masks a lack of detailed thinking, of wrestling with the complex (probabilistic) trade-offs needed.
And this risk cuts both ways. When a minister promises vaguely to “secure borders”, “grow the economy”, or calls to “level up the country”, no-one can prove failure. No-one knew if the plan included the necessary, relevant considerations for it to work. But also: no-one knew how likely success was.
As the Cheshire Cat puts it in Alice in Wonderland: if you don’t know where you want to get to, it doesn’t much matter which way you go. We might add: if you never said where you were going, opponents can always insist you arrived somewhere else. Evasiveness helps you escape electorally damaging criticism, but also denies you electorally enhancing credit for success.
The AI applies a discipline available to all of us: every claim must become a testable forecast. Not “Level Up” but achieve X, by Y date, which has succeeded if Z is true. Strategy becomes probabilistic: “there is a 35% chance we will do A, and if we do A, it becomes 50% more likely we will achieve B”. The language tightens. The room for evasion narrows.
Take a familiar pledge: cutting NHS waiting lists within two years. Under this system, that promise is translated into a measurable question: “Will the NHS elective waiting list fall below 6 million by March 2028, as reported by NHS England?” The model would assign a probability, say 40%, and identify the drivers: staffing levels, seasonal demand, social care capacity, industrial action. If the plan ignores one of those drivers, the gap is visible from the start. Voters can see the promise, and the conditions required for it to be honoured.
Trust in politicians is at record lows, partly because incentives are misaligned. A party can simultaneously promise tax cuts, higher spending and a balanced budget to seek election. Contradictions are buried in the detail. A forecasting approach forces coherence and clarity. If the numbers do not add up, the model will show the chance of delivering all three outcomes approaches zero. The manifesto becomes legible.
More honest debate results. A politics that tolerates vagueness invites distrust. A politics that demands measurable claims begins to reward honesty. When every major policy is tied to a public forecast, the costs of self-deception and deceit rise.
You might object that the existence of systems that prospectively hold Government to account will make officials and politicians even less likely to be specific – to cling tighter to empty rhetoric and vague promises. But the discipline still applies to vague claims. A general but undefined promise to ‘Level-Up’ can be charitably interpreted and forecast. Now politicians are on ‘Question Time’ arguing about definitions, probabilities – how likely their pledges are to be achieved - and why the world’s best AI forecaster is wrong and they are right.
There are limits. Cassi is not omniscient. The numbers submitted to the Defence Committee may prove wrong. But they can be tested. The aim is to estimate your degree of uncertainty in your forecast and be less wrong over time – you learn to trust those, human or AI, most consistently right. Judgement remains important – perhaps you know something the AI can’t, or doesn’t, so can defend your forecast against the model’s. Instead of arguing over rhetorical claims, talking past each other, politicians argue over probabilities and evidence. Instead of defending a claim with a speech, they defend it with their record of trustworthy policy forecasts.
In a system that prefers assertion to accountability, AI truth engines turn narrative into numbers. They may reverse the decline of our democratic politics more than Orwell would have dared hope possible.
The truth might set you free too - come find out! Visit us at https://cassi-ai.com and register your interest in becoming a customer to learn more.
At the time the forecasts were both submitted to Parliament, and when published in the Times ~two months later, Cassi was the leading AI Forecaster on ForecastBench and had been since early January 2026. On Friday 8 May Cassi fell behind two new entrants from Deepmind. As at 16:27 12 May 2026, Cassi was fourth, overtaking one Deepmind model remaining marginally behind the other, and for the first time, overtaken by xAI’s entry - with which we have been tied for the lead several times since January, but never previously behind. ForecastBench is a live benchmark, we may have moved up or down the league table by the time you read this.