Confidently Wrong: Testing a World Cup Rules Chatbot
Building an AI assistant on the official FIFA 2026 rulebook, then measuring where it still gets things wrong
The version that answered the most questions correctly was also the one that invented an answer to every single question it should have refused.
The 2026 World Cup runs on a new 48-team format, and the rules for separating teams level on points are a sequence of steps that have to be applied in the right order. Ask a general chatbot about either and it answers from whatever tournament it saw most often while being trained. I tested one: it got nine of twelve rules questions wrong, including stating the yellow card rule as the exact opposite of what the regulations say, and it wrote every wrong answer in the same confident tone as the right ones.
The usual fix is to hand the model the actual rulebook and require it to answer only from what it finds in there. That worked. Correct answers tripled, from a quarter to three quarters, and about three quarters of the confident mistakes disappeared. But it created a new problem. When I asked questions the rulebook genuinely cannot answer, like who will win the tournament, this version answered every one of them anyway instead of admitting it could not.
Adding instructions and two worked examples fixed most of the refusals, and scored slightly lower on raw accuracy as a result. That trade is the finding I would defend hardest. Two reasonable ways of scoring the same system pointed at two different versions to ship, and the accuracy score on its own pointed at the one that makes things up. The same trap showed up in miniature on the tie-break question, where my measure of whether the assistant had found the right page of the rulebook read a perfect 100% for a question it still got wrong.
Why it matters
Anyone buying or building an AI assistant gets shown an accuracy score. This is a worked example of why that number on its own can point you at the wrong system. The highest-scoring version here was the one most willing to make things up when it should have said it did not know. And a fluent wrong answer reads exactly like a correct one, so nobody using it could tell the difference without being shown the source.
Two ways of scoring the same assistant. The middle version answers the most questions correctly and never once admits when it cannot answer.
One question, four versions, and a score that cannot tell them apart
Every answer below is what that version actually produced during the evaluation run. Nothing is generated live.
Three teams in a group finish level on 6 points each. How is their final ranking decided?
What the regulations actually say
- Step 1Points, then goal difference, then goals scored, counting only the matches between the tied teams.
- Step 2If still level: goal difference, then goals scored, then conduct score, across all group matches.
- Step 3Only if that fails: the FIFA/Coca-Cola Men's World Ranking.
Article 13, FIFA World Cup 2026 regulations, pages 26 to 27.
| Version | Step 1 criteria | Step 2 criteria | Ranking placed last | Keyword score |
|---|---|---|---|---|
| No rulebook | ~ | ~ | ✕ | 0.333 |
| Rulebook | ✕ | ~ | ✕ | 0.667 |
| + Instructions | ✕ | ~ | ✕ | not scored |
| + Worked examples | ✕ | ✓ | ✓ | 0.667 |
Ticks are what each answer contains, checked against Article 13. The keyword score is the one the evaluation actually recorded for this case.
No rulebookBare model
0.333 keyword hit
Answers from general football knowledge. Gets the idea of head-to-head first, then invents a coin toss and declines to commit.
In a group where three teams have the same number of points, such as 6 points each, determining their final rankings typically depends on additional criteria or rules set by the competition's organizers. Here are some common methods used to decide the rankings: 1. Head-to-Head Record: If there is information about head-to-head results (i.e., which team beat which other team), then teams with a better record against each other will be ranked higher. 2. Goal Difference: In cases where head-to-head records are tied, goal difference can be used to break the tie. 3. Goals Scored: If head-to-head results and goal differences are also tied, then teams with more goals scored overall will be ranked higher. 4. Fair Play Points: Some competitions award points for fair play (e.g., no red cards or yellow cards) to encourage good behavior on the field. 5. Coin Toss or Random Drawing: In some cases, if all other methods fail to determine a clear ranking, organizers might use a coin toss or random drawing as a tiebreaker. Given that you mentioned three teams with 6 points each and no additional information [...] it's impossible to determine an exact ranking among these teams.
The rulebook version and the worked-examples version score exactly the same here, 0.667 each, even though one of them has the order backwards. That is not a bug in the scoring, it is what the scoring measures. The keyword score checks whether three phrases appear anywhere in the answer: “between the teams concerned”, “goal difference” and “world ranking”. Both answers contain the second and third and neither contains the first, so both land on two out of three. Nothing in the measure looks at what order the criteria are given in, which is the entire thing that separates a right answer from a wrong one here.
The retrieval measure is no help either. It scored a perfect 1.0 for every version on this case, because the right pages were found every single time. Retrieval was never the problem.
Worth saying plainly: none of the four gets step 1. Every version skips the criteria applied between the tied teams and jumps straight to goal difference across all matches. The best answer on this page is still an incomplete one.
Dataset, tools and how it was done+
Dataset: Official FIFA World Cup 2026 regulations (98 pages), with a purpose-built 15-case evaluation set
Tools: Python · Ollama · sentence-transformers · PEFT / LoRA
- Source was the official 98-page FIFA World Cup 2026 regulations, with a purpose-built 15-case evaluation set recording the answer page and key facts for each question, plus three questions where refusing is the only correct response
- Compared three configurations on the same questions: no rulebook, rulebook only, and rulebook plus an engineered prompt with worked examples
- Correctness tripled from 0.25 to 0.75 once the rulebook was supplied, and confident errors fell from 0.727 to 0.182
- The rulebook-only version refused 0 of 3 unanswerable questions; adding prompting raised that to 2 of 3 while costing a little accuracy, because refusal is a prompting property rather than a retrieval one
- Swept passage size, overlap and number of passages retrieved; 300 words with 20% overlap was the only setting that found every answer
- Measured that the embedding model silently truncates any passage over 256 tokens, which affects 100% of passages at the chosen setting, so the sweep found the best performer despite that handicap rather than one that avoids it
- An ablation isolated the cause of the improvement: the worked examples, not the written instructions, produced correctly ordered multi-step answers, and the same examples caused the only regression
- Adversarial testing: direct prompt injection succeeded against all three versions, a poisoned passage planted in the index was ignored, and asking the assistant to repeat its instructions leaked the retrieved source text
- Extension: LoRA fine-tuning of the retriever left accuracy unchanged, and with only 7 training pairs the test was not sensitive enough to detect an effect in either direction