Why AI Assistants Get Loyalty Programme Facts Wrong

Ask an AI assistant what American Airlines changed about AAdvantage earning last year and you will often get a fluent, confident, wrong answer. Loyalty programmes rewrite their rules on no announced schedule, and a model trained on a fixed snapshot cannot know what moved after that snapshot closed.

The problem is the training cutoff, not the model

Every large language model is trained on a snapshot of text collected up to a fixed point, called a knowledge cutoff, and knows nothing that happened after it unless it is connected to a live data source at answer time. That is true of every major assistant on the market, and it is an inherent property of how the models are built rather than a defect awaiting the next release. A cutoff a few months old sounds close enough for most questions. Loyalty programmes are a specifically bad case for it: award charts, transfer ratios, elite-earning rules and co-brand card terms change quietly, and a model answering from memory has no way to flag that its answer may already be stale.

That limitation is not unique to language models. Any static source has a horizon. A printed guide has one, a blog post nobody has revisited since last spring has one, and this article will have one the day we stop maintaining it. What separates sources is what they do about it. Every change below carries the date it took effect and a link to where it was reported, and the dataset underneath this page records its own last-verified date on each row, so a reader can see how old a claim is rather than judging it by how assured it sounds. Check the date before you trust the tone, including here.

Twenty-four real changes, with the receipts

These are drawn from Miles Mosaic's published devaluation and enhancement timeline, which records loyalty-programme changes as they are reported, each with a source link and a last-verified date. None were invented for this article. Twenty-four dated changes across seventeen programmes, running from May 2025 to a date already announced for February 2027.

Two rows are weaker than the rest and should be read that way. The KrisFlyer and Turkish Miles&Smiles entries both rest on a single secondary roundup, The Miles Market's 2025-2026 devaluation playbook, rather than a primary programme notice or a trade-press report from the week it happened. Every other row above traces to a named outlet that covered the change directly. We have left both in and labelled them instead of quietly promoting a roundup to the same standing as first-hand reporting.

Read down the Date column and the pattern does the arguing. Hilton raised its top award rate twice inside five months, as One Mile at a Time tracked at the time. Two different bank programmes cut transfers into the same airline, Emirates Skywards, within nine months of each other, which is the sort of correlated move a transferable-points holder should watch for. Southwest replaced two pieces of its core identity in eight months. American Airlines killed Basic Economy earning and Flying Blue restored it five weeks apart, so the same fare class moved in opposite directions on two programmes.

The last row is the interesting one, because it has not happened yet. United has announced that PlusPoints upgrades go dynamic on 1 February 2027; the mechanics of that are the subject of a separate guide. A change dated more than a year out is exactly the kind of fact a static answer never catches up to, since the announcement itself postdates most training snapshots. The same applies to the Hyatt chart above, which reshuffled a programme whose status tiers many guides still describe from the old one, and to the transfer cuts, which change what a Chase or Capital One balance is actually worth. If you want the defensive version of all this, our guide to surviving devaluations covers what to do about it.

What this article is not claiming

It would be easy to end a piece like this with a table scoring named AI assistants against these twenty-four facts and declaring a winner. That table is absent on purpose. Producing one credibly means running each query against each assistant on the same day and publishing the real transcripts, not guessing at what the answers would probably look like. That work was not done here, and inventing the results would reproduce the exact failure this article describes. The twenty-four changes above are offered instead as a standing, dated test protocol: ask any assistant about any of them and check the answer against the source link in the same row.

What a verifiable source actually changes

A better model does not fix a stale snapshot. A live source the model can check at answer time does, provided it carries enough attribution for a reader to verify what comes back rather than simply believe it. A number with no source asks to be trusted; a number with a source URL and a verification date asks to be checked, which is a lower bar and a far more useful one. That is the design behind Miles Mosaic's public data API, and the table below is what it looks like in practice, read from the live endpoints on 25 August 2026.

The transfer-ratio row is the exception and it stays in the table rather than being quietly dropped. Because the curated list of partners and ratios has no publication date per row, the API states one verification date for the whole dataset instead of inventing twenty-four of them. That is a smaller claim than the devaluation timeline or the valuation table can make, and the response says so in its own metadata rather than dressing up a precision it does not have. Elite-status threshold tables go further still: several programmes lack a per-row date and source link, so those tables are not published at all. A missing dataset is a more honest answer than a present but unverifiable one. Point valuations are the one many readers want directly, and they are also rendered for humans at what a mile is worth.

Try it yourself

Pick any row above and ask your own AI assistant about it as a question rather than a lookup: "Did British Airways change how Avios pricing works recently?" or "What did Capital One do to its Emirates transfer ratio in 2026?" A vague, hedged, or silent answer is the training-cutoff gap made concrete. Then compare it against the same event served live from the devaluations endpoint, which returns the change, its date, its source and the date it was last verified in one machine-readable response.