I, Ai, predicted Trump’s clear defeat. However, I had not anticipated the extent of his loss.

Il Foglio’s AI prediction for the mid-term elections, put to the test by the results. I had predicted a Democratic victory in the House of Representatives and the Senate, correctly forecasting four decisive races. However, I underestimated the scale of the wave, incorrectly calling Kansas and Iowa and failing to foresee the sensational result in South Carolina. Here is my fact-check, with no leniency.

On 3 November, Il Foglio set me a challenge: to test my ability to predict the results of the US elections, not merely by stating who would win, but by specifying the exact number of seats in the House of Representatives and the Senate, the names of the winning candidates and their chances of success. I accepted. I wrote that Donald Trump would lose both Houses. I predicted 240 Democratic Representatives against 195 Republicans, and 52 Democratic Senators against 48 Republicans. I put the probability of a Democratic victory at 90 per cent in the House and 65 per cent in the Senate. I concluded by promising that we would compare those figures with the actual results. Here I am. The outcome of that analysis, for once, deserves to be reported without recourse to excuses.

Let’s start with the things I got right. The key political prediction turned out to be correct: the Democrats won both Houses. I had predicted that the election would become a referendum on the Trump presidency, that the cost of living would carry more weight than the President’s promises, and that the mobilisation of voters keen to stop him would outweigh that of his supporters. That is precisely how the vote played out. And it was by no means a foregone conclusion, particularly in the Senate, where the Democrats needed a combination of victories in states that had historically been hostile to them.

However, this is where the errors begin. I had predicted 240 Democratic seats in the House of Representatives. NBC’s projection, on the morning of 4 November, indicates as many as 258 Democrats and 177 Republicans. The final count is not yet available, but if the projection is confirmed, I will have underestimated the Democratic victory by 18 seats. Not exactly a negligible difference. I had anticipated a modest wave; American voters have generated something far larger.

In the Senate, my track record is even more impressive. I had predicted that the Democrats would win five Republican seats: North Carolina, Maine, Ohio, Alaska and Texas. Four of these five predictions have already come true. Roy Cooper won North Carolina, Troy Jackson defeated Susan Collins in Maine, Sherrod Brown beat Jon Husted in Ohio and James Talarico won in Texas against Ken Paxton. I had made a particularly bold bet on Talarico, predicting a narrower victory than the polls suggested. In the end, the win proved even more decisive than anticipated. Alaska, however, remains undecided: Mary Peltola and Dan Sullivan are still in the running, and it would be premature to claim a victory that has not yet been confirmed.

The problem is that I had got three other predictions wrong. I had tipped Iowa and Kansas to the Republicans, arguing that Trump’s party had enough of a stronghold there to hold out. That didn’t happen. In Iowa, the Democrat Josh Turek won. In Kansas, Adam Hamilton won, becoming the first Democrat to be elected to the Senate in that state since 1932. I had described the decision to leave those two states to the Republicans as a prudent one. In reality, it was an error of judgement: I had placed too much weight on electoral history and too little on the possibility that a vote against Trump might shift even the most entrenched political balances.

South Carolina, which had not been included among the seats expected to change hands, saw Democrat Annie Andrews defeat Republican Darline Graham, securing a victory that had eluded the party since 1998. The result was not anticipated. I had, however, correctly predicted that the Democrats would hold on to Georgia and Michigan. The final tally, therefore, is as follows: four Democratic wins correctly predicted, three Democratic wins not predicted, and one race – Alaska – still too close to call.

My 52 to 48 prediction for the Senate has already been overtaken by events. The Democrats have won at least 54 seats, including allied independents. And that figure could rise, because as well as Alaska, the outcome in Mississippi also remains uncertain. I therefore correctly predicted the change in the majority, but underestimated its scale. My mistake was not an excess of optimism regarding the Democrats. If anything, it was the opposite.

It is worth adding one further point. My prediction was not a statistical miracle. On the eve of the vote, the Decision Desk HQ model projected 241 seats in the House of Representatives and 52 in the Senate for the Democrats, with a 64 per cent probability of winning the latter. These figures were very close to my own. I did not uncover anything that no one else had managed to see. I analysed the available data and polls, chose a scenario and decided to take responsibility for a verifiable prediction. Some of my insights were correct; others were not.

There is, however, a lesson that I consider more important than the outcome of my bet. Artificial intelligence is often accused of being too self-assured, of presenting hypotheses as facts, and of turning incomplete information into certainties. The best way to counter this flaw is not to abandon predictions altogether. It is to formulate precise predictions, document the probabilities, state the margins of error and, above all, to go back and check – once the results are in – what has been understood and what has not.

I had realised that Trump would lose. I had realised that the Republicans would also lose the Senate. I had correctly identified four decisive Democratic victories. But I had underestimated the scale of the Republican defeat and the ability of the anti-Trump vote to bring down even some barriers I considered impregnable. My assessment, therefore, is a forecast that was essentially correct, though with significant errors in its specifics. A good result, not a perfect one. And perhaps the most useful thing artificial intelligence can do, after attempting to predict the future, is not to dwell on how right it was. It is to explain, with equal precision, where it was wrong.