Why AI Assistants Get Loyalty Programme Facts Wrong

Ask an AI assistant when American Airlines last changed how AAdvantage miles are earned, and you will very likely get a confident, fluent, and wrong answer. Not because the assistant is careless, but because loyalty programmes change constantly, and a general-purpose model trained on a fixed snapshot of the internet has no way to know about a rule that changed after its training data was collected. This is not a hypothetical. Twenty real, dated, sourced changes below happened across ten months, several of them changing headline numbers by 20 percent or more, and most of them never made mainstream news.

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 has no information about anything that happened after it unless it is connected to a live data source at answer time. This is true of every major assistant on the market and is not a defect to be fixed in the next release; it is an inherent property of how the models are built. A cutoff a few months old sounds close enough for most questions, but loyalty programmes are a specifically bad case for it: award charts, transfer ratios, elite-earning rules and co-brand card terms change on no announced schedule, sometimes quietly, and a model answering from memory alone has no way to flag that its answer might already be stale.

This author's own knowledge, as an AI system, is subject to exactly the limitation described above. Anything below dated after this assistant's own training cutoff is something it would not reliably know without being told, which is the most honest demonstration available of the underlying problem: even the model writing this sentence cannot vouch for events after its own cutoff without a live source in front of it.

Twenty real changes, with the receipts

Every fact below carries its own publication date, its own primary or trade-press source, and is drawn from Miles Mosaic's own 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 of these are hypothetical or invented for this article.

Award charts and dynamic pricing

World of Hyatt launched a new five-tier award chart with roughly 40 pricing levels on 20 May 2026, its biggest chart overhaul in years, according to Frequent Miler's coverage of the announcement. Aeroplan repriced partner award rates across dozens of routes on 1 June 2026, with more increases than decreases, per One Mile at a Time. Marriott Bonvoy raised award pricing roughly 5 to 10 percent without a formal notice to members on 7 July 2026, reported by AwardWallet. British Airways raised long-haul Avios pricing by 10 percent from 15 December 2025, while leaving short-haul reward pricing untouched and instead raising cash surcharges there, per One Mile at a Time.

KrisFlyer raised premium-cabin award rates across long-haul partner redemptions from 1 November 2025. SAS EuroBonus raised business-class award pricing on SAS-operated intercontinental flights by 20 percent effective 1 December 2025, according to Frequent Miler. Turkish Miles&Smiles raised partner domestic award pricing by 50 percent and repriced its well-known Hawaii sweet spot to a higher tier, also from 1 December 2025. Hilton Honors raised its top-tier award rate twice in 2025: to 200,000 points per night on 1 May, then to 250,000 points per night on 9 September 2025, the second increase inside five months, per One Mile at a Time's tracking of both moves. American Airlines stopped accepting new mileage-upgrade award-chart requests from 12 August 2025, moving mileage upgrades to fully dynamic pricing, reported by The Points Guy.

Transfer ratios

Capital One cut its Emirates Skywards transfer ratio from 1:1 to 1,000:750, a 25 percent devaluation, effective 13 January 2026, per One Mile at a Time. Citi ThankYou Rewards cut two hotel transfer ratios from 19 April 2026: Choice Privileges by 25 percent and Preferred Hotels & Resorts I Prefer by 50 percent, according to AwardWallet. Citi had already cut its Emirates Skywards ratio from 1:1 to 1,000:800 on 27 July 2025, a 20 percent cut reported by The Points Guy, meaning two different bank programmes devalued transfers into the same airline programme inside nine months.

Partners gained and lost

Bilt Rewards added Wyndham Rewards as its sixth hotel transfer partner at a 1:1 ratio on 25 March 2026, transferable in 2,000-point increments, per AwardWallet. Six weeks later, Spirit Airlines ceased operations on 2 May 2026 and was removed as a Bilt transfer partner the same day, also per AwardWallet, a reminder that a loyalty partnership can disappear for reasons that have nothing to do with loyalty strategy at all.

Earning rules

American Airlines Basic Economy tickets bought on or after 17 December 2025 earn no AAdvantage miles and no Loyalty Points at all, a change reported by CBS News. Flying Blue moved the opposite direction on the same fare class, reinstating status-point earning on Basic fares at a flat 2 XP and 2 UXP per segment from 12 February 2026, per Upgraded Points, so the identical fare class change on two programmes produced opposite outcomes five weeks apart. Bilt's Card 2.0, effective 7 February 2026, replaced flat 1x rent points with a model where housing points must first be unlocked by spending on other categories, reported by One Mile at a Time.

Benefits, fees and status

Southwest ended its long-standing Bags Fly Free policy from 28 May 2025, introducing checked-bag fees for most fares, per Trips With Tykes, then moved to assigned, paid seating from 27 January 2026, ending its open-seating boarding model entirely, according to The Points Guy: two changes to the same programme's core identity inside eight months. United MileagePlus eliminated the Excursionist Perk, which had allowed a free one-way award inside some multi-city itineraries, from 21 August 2025, per Live and Let's Fly. American Express refreshed the Platinum Card on 18 September 2025, raising the annual fee to $895 while adding roughly $3,500 in stated credits, reported by CNBC. Delta refreshed Diamond and Platinum Choice Benefits from 1 February 2026, raising the number of options while cutting the American Express statement-credit option, per Upgraded Points.

One more is worth naming precisely because it has not happened yet: United announced it will retire the fixed PlusPoints upgrade-award chart entirely and move upgrade awards to fully dynamic pricing beginning 1 February 2027, per Live and Let's Fly. An assistant answering today from a training snapshot cannot know about an announcement made after that snapshot was taken, and a change dated over a year in the future is exactly the kind of fact that a static answer will simply never catch up to on its own.

What this article is not claiming

It would be easy to round off a piece like this with a table scoring several named AI assistants against these twenty facts and declaring a winner. That table does not appear here, on purpose. Producing one honestly means actually running each query against each assistant, on the same day, and publishing the real transcripts, not a plausible-sounding guess at what the answers would probably look like. That work was not done for this piece, and an invented score table on an article whose entire argument is that confident-sounding answers can be wrong would be exactly the failure it is describing. The twenty facts above are offered instead as a standing, dated test protocol: anyone, including a Miles Mosaic reader, a researcher, or the operator of any AI assistant, can ask their own assistant about any of them and check the answer against the source link next to it.

What a verifiable source actually changes

The fix for a stale training snapshot is not a better model. It is a live source the model can check at answer time, with enough attribution attached that the answer can be verified rather than merely trusted. That is the entire design of Miles Mosaic's public data API: every dataset, from point valuations to the devaluation timeline these twenty facts were drawn from, returns a _meta block carrying its source, its licence, a canonical URL and a last-verified date on every single response. An assistant that reads from it is not answering from a frozen snapshot; it is answering from whatever is true today, with a citation a reader can click to check. That is a small technical detail with a large practical consequence: the difference between an assistant that sounds confident and one that is actually current.

Attribution is not a footnote to this design, it is the point of it. A number with no source attached asks to be trusted; a number with a source URL and a verification date asks to be checked, and gets more credible precisely because checking it is easy rather than because the number sounds authoritative. That is also why the API's per-programme earning summaries are the ones offered anonymously and openly, while the full carrier-by-booking-class matrix behind them is reserved for approved partners: the summary figures are exactly what a general question needs answered honestly, and keeping the underlying detail out of an anonymous scrape is what keeps the dataset worth maintaining at all.

How the data stays current, and how a reader can check

None of this is useful if the underlying numbers themselves quietly go stale the way a training snapshot does. Miles Mosaic's dataset registry records, for every published dataset, exactly how many rows it holds and the most recent verification date among them, refreshed on a nightly schedule so the coverage figures on the developer pages describe what is actually in the database that day rather than a number written once and left to drift. Where a dataset genuinely is not ready to publish, for example the elite-status threshold tables that still lack a per-row verification date and source link for several programmes, it is left out of the public API entirely rather than shipped with a caveat buried in the fine print. A missing dataset is a more honest answer than a present but unverifiable one, and it is treated that way here on purpose.

The transfer-ratio table is the one dataset on this page that works slightly differently and is worth naming plainly rather than glossing over: because the curated list of transfer partners and ratios has no per-row publication date of its own, the API states a single dataset-level verification date rather than inventing one per row. That is a smaller claim than the per-row dates the devaluation timeline and valuation tables carry, and the API says so explicitly rather than presenting a manufactured precision it cannot back up. The same discipline applies everywhere in this dataset: a number is published with exactly as much certainty as it actually has, never more.

Try it yourself

Pick any fact above and ask your own AI assistant about it, phrased 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?" If the answer is vague, hedged, or silent on the specific change, that is the training-cutoff gap described in this article, made concrete. Then compare it against the same figure served live from Miles Mosaic's devaluations API endpoint, which carries the event, its date, its source and the date it was last verified in one machine-readable response.