ChatGPT does not know today's exchange rate: why AI gets rates, forward points and premiums wrong

The scene has become routine: before calling the bank, you ask an assistant "what is EUR/USD today?" or "how much does a six-month FX option cost?". The answer arrives in two seconds, well written, quoted to the fourth decimal place.

The problem is not that the answer is sometimes wrong. It is that it looks exactly the same when it is wrong as when it is right.

Why a language model does not know the rate

A language model is trained on a body of text frozen at a given date; vendors in fact publish, for each model, the cut-off date of its training data. An exchange rate changes constantly. Without a tool connected to a market source, the assistant reads no rate at all: it produces the most plausible number given what it has learned.

Why doesn't it simply say "I don't know"? A study published by OpenAI researchers in 2025 explains it: the way models are trained and evaluated rewards guessing over acknowledging uncertainty. A model that offers a plausible number scores better on average than one that abstains.

Even connected, there is no single "rate of the day"

The question itself is ambiguous. There are several rates of the day, all legitimate, all different:

Forward points: a calculation, not a number to look up

For a forward, the error changes in nature. Forward points are not information you find somewhere: they are calculated from the spot rate and both currencies' interest rates at the exact maturity. This is covered interest parity.

An assistant that "gives" six-month forward points has therefore either invented them or calculated them with rates it does not state — often policy rates, which are not six-month market rates. The result can differ markedly from the forward actually available.

An option premium: the most missing input

An option needs one more input. The reference model for FX options, Garman and Kohlhagen's, requires the spot rate, the strike, the maturity, both currencies' rates — and volatility.

Volatility is the problem: not past volatility, but the implied volatility quoted in the options market for that maturity and strike. No general-purpose assistant has access to it. A premium "estimated" with an assumed volatility is not an approximation of the right price: it is the price of a different option.

What a market number must carry to be usable

Where the assistant genuinely helps

• Explaining a formula: how forward points are calculated, what implied volatility measures.
• Building the spreadsheet in which you will do the calculation with your own sourced data.
• Reviewing reasoning and flagging an inconsistency — an inverted quote, a wrong day count.
• Drafting the memo that presents figures you have checked elsewhere.

The split is always the same: the assistant for method and text, a time-stamped market source for the numbers.

Sources

• Kalai, A. T., Nachum, O., Vempala, S. and Zhang, E. (2025), "Why Language Models Hallucinate", OpenAI, arXiv:2509.04664.
• Anthropic, "Models overview": training-data cut-off date published for each model.
• European Central Bank, "Euro foreign exchange reference rates": published around 16:00 CET, market conditions around 14:15 CET, use for transactions strongly discouraged.
• Borio, C., McCauley, R., McGuire, P. and Sushko, V. (2016), "Covered interest parity lost: understanding the cross-currency basis", BIS Quarterly Review, September 2016.
• Garman, M. and Kohlhagen, S. (1983), "Foreign currency option values", Journal of International Money and Finance, 2(3), pp. 231-237.