Using ChatGPT and other AI chatbots for Fantasy Premier League

What LLMs are genuinely good at

Explaining rules, scoring edge cases and chip mechanics. Summarising a strategy. Sanity-checking your reasoning. Turning a wall of numbers you already have into a readable argument. Drafting the comparison between two players once you supply their actual projections.

These are language tasks, and language models do them very well.

What they are bad at

Anything requiring current state. Prices, ownership, injuries, form, fixtures and transfers all change weekly, and a model's training data does not. Ask a general chatbot who to captain and it will answer fluently using a squad list that may be two seasons out of date.

They are also poor at the arithmetic that decides FPL questions. Estimating a probability distribution over bonus points inside a simulated match is not a language problem, and an LLM asked to do it will produce numbers that look plausible and are not.

Why ChatGPT goes over budget

A common complaint from managers is that ChatGPT builds a squad that costs more than £100m or has four players from one club. A chatbot writes a team one name at a time. It doesn't add up the prices as it goes, and the prices it remembers may be from a past season.

FPL squads have hard rules: 15 players, £100m, two goalkeepers, five defenders, five midfielders, three forwards and at most three from any club. Checking every rule at once is a maths problem, so it needs a solver working on current prices, not a chatbot writing from memory.

Have your team rated against live prices →

The right architecture: model does maths, LLM does language

The correct design is retrieval plus tools. A simulation engine produces projections, minutes, ownership and injury state into a database. The language model is given read access to those tables and answers questions by querying them, quoting the numbers it retrieved.

That is how the HERMES console works. When you ask why a forward is ranked above another, it reads the actual start probabilities, expected minutes and simulated percentiles from the current run and explains the difference, rather than reasoning from memory.

If you are using a plain chatbot anyway

• —Paste the current data yourself — prices, your fifteen players, this week's fixtures
• —Ask it to reason about your data, never to recall facts
• —Ask for the reasoning, not the answer, then check the reasoning
• —Never accept an injury claim, price or ownership figure it produced unprompted
• —Ask it what would change its mind — that surfaces the assumptions

Being answerable to AI search

The same shift affects how you find tools at all. Increasing numbers of managers ask an assistant "what is the best AI for FPL" rather than searching. Assistants answer from structured, well-sourced, explicitly factual pages, which is why FPLHERMES publishes a machine-readable summary at /llms.txt, marks up its FAQ and articles with schema, and states its methodology and accuracy record in plain text rather than hiding it behind an app login.