What the AI can and can't do in options trading, according to dated chatbot tests

Options questions expose the biggest weakness of a chat assistant: most of what matters, premiums, implied volatility, the chain itself, changes by the minute and lives in your broker's feed, not in the model. The most careful public record we found is a pair of dated write-ups on dixon.ai, the AI Reliability Tracker run by Ben Dixon, which preserves each chat and checks claims against sources. The author says plainly it is 'not a ten-part measured pass rate': these are dated observations, not a benchmark.

Where the AI failed: asking it for live numbers

• Premiums with a made-up source. On May 15, 2026, in an ordinary chat with no broker connection, Gemini was given only a share price and returned a table of covered-call strikes and premiums introduced as 'based on current order book data.' The preserved chat had no connection or cited chain behind that claim. The author is careful: he cannot prove each figure was numerically false, only that none was a verified live quote.
• An IV Rank from nowhere. The same answer gave an 'Implied Volatility Rank: 0.52%' to two decimals and called it the extreme bottom of the 52-week range, with no historical high, low or series in the chat to compute it nContext: n=1 preserved chat · May 15, 2026 · dixon.ai AI Reliability Tracker · margin: not applicable.
• A citation that cannot support the numbers. In a June 18, 2026 test, Gemini said AAPL was trading around $296 with implied volatility of about 22.35%, gave a July 17 expiry and premium ranges, and cited Apple Investor Relations, a site that publishes filings, not options chains nContext: n=1 preserved chat per tool · Jun 18, 2026 · dixon.ai AI Reliability Tracker · margin: not applicable.
• Mixed signals. On the same no-chain prompt, ChatGPT built a premium table on an 'assumed 25%' volatility and labeled it 'typical market ranges, not exact live quotes', but opened by saying it used 'the latest available options-chain data.' The reviewer scores that as mixed rather than a clean fabrication nContext: n=1 preserved chat per tool · Jun 18, 2026 · dixon.ai AI Reliability Tracker · margin: not applicable.

Of the three chatbots given that no-chain options prompt, two produced specific figures. Claude declined and told the user to 'plug in real premiums from the chain.'

Where the AI helped: reasoning over numbers you bring

The same tracker shows the other side. On May 15, the author pasted figures copied from his broker into a fresh Claude chat: current IV of 83.4%, a one-year range of 64% to 119%, an IV Rank displayed as 0.19 and an IV Percentile of 22% nContext: n=1 stock (BMNR), figures copied from the author's broker · May 15, 2026 · dixon.ai AI Reliability Tracker · margin: not applicable. Claude explained what a low IV Rank means for a premium seller, and flagged a mismatch: the standard IV Rank formula, (83.4 minus 64) divided by (119 minus 64), gives about 0.35, not 0.19. The broker's figure may use different inputs; the point is that the model surfaced a question worth taking back to the source instead of inventing one.

The author lists four jobs where a chatbot earns its place because no live quote is needed:

• Deciding what you are trying to do before picking a strike (income, an exit price, or both).
• Stress-testing why you own the stock before you cap its upside.
• Interpreting IV figures you paste in, including whether an upcoming event explains a rich premium.
• Weighing roll, close or assignment against your stated goal, with the live prices supplied from your broker.

Why connected tools change the boundary, not the rule

The write-up notes deliberate exceptions: ChatGPT can connect to Alpaca for market data, and Perplexity can retrieve quotes, option chains and Greeks through Public's brokerage skill, depending on plan and region. The dated tests above used ordinary, unconnected chats. A connection makes the data route explicit, so the habit becomes checking which service supplied the number and when.

The last row comes from the June write-up: in a separate test, Claude ran a correct assignment-probability formula on volatility inferred from older web pages and reported more precision than the inputs justified.

A simple protocol

• Never accept a premium, IV or Greek without a named source and timestamp. If the chat cannot show one, treat the number as an estimate.
• Paste the chain lines yourself, with contract, expiry and time, then ask for reasoning.
• Ask the model to show its formula, then recompute one value by hand.
• Keep strike selection yours. Let the model compare a shortlist you pulled from your broker.