Most broken AI promises were never promises

We set out to build a ledger of broken AI promises. The first three we investigated — the most notorious ones we could think of — were not broken promises at all. Two were never promised. One was delivered on time and simply did not work.

What we were trying to build

The plan was ordinary: a public ledger of things AI companies said they would do, set against what actually happened, with an archived copy of every original claim so readers could check our reasoning rather than take it on trust. Sortable, filterable, one row per promise.

We applied one test to every candidate row. A promise had to be:

1. A specific thing — a named product, model, or capability.
2. By a stated time — a date, a quarter, or a phrase like “in about six months” that can be checked against a calendar.
3. In the vendor’s own words — a statement we can quote and archive, not a characterisation in someone’s coverage.

That test turned out to be the whole story.

The three that dissolved

Kimi K3 was never promised under a Modified MIT licence

The widely-reported version: Moonshot AI said K3 would ship under the same Modified MIT licence as K2, then shipped a restrictive bespoke licence instead. We went looking for the statement. There isn’t one. The K3 repository carries a bespoke licence from its initial commit, with no licence-replacement commit in its history, and Moonshot’s own GitHub README says both code and weights are released under the Kimi K3 License. No Moonshot statement promising Modified MIT exists.

What happened is that outlets extrapolated from the previous generation. Including us. Our own K3 deep dive asserted Modified MIT four times and framed it as a vendor commitment — “committed to”, “the promised Modified MIT terms”. It now carries a dated correction. The phrase that made it look like a broken promise was written on this site, not by Moonshot.

Meta never committed to releasing Llama 4 Behemoth

Behemoth is the model everyone remembers as promised and undelivered. Meta announced it in April 2025, it never shipped, and by 2026 the company’s frontier work had moved to closed weights. The Hugging Face organisation still lists only Scout and Maverick.

But Meta’s own announcement says: “While we’re not yet releasing Llama 4 Behemoth as it is still training, we’re excited to share more technical details about our approach.” That is a status report. There is no release date, and no commitment to release at all. The expectation was real; the promise was not.

Rabbit shipped its Large Action Model on the date it named

This is the sharpest case. The r1 is the emblem of AI over-promising: a device that was going to “replace apps and control all services with a single sentence”, and then did not.

That CES 2024 pitch carried no date, so it fails the test. The one dated commitment Rabbit made — that the web-based Large Action Model agent would arrive on 1 October 2024 — was met. LAM Playground shipped that month. What failed was quality: it struggled with CAPTCHAs and unintended behaviour, and by September 2024 roughly 5,000 of about 100,000 buyers were using the device at any given moment.

Recording that as a broken promise would mean scoring the vendor on whether the product was good, which is a different question and one a promise ledger is not built to answer.

What survived

Four rows cleared the test. This is the evidence table — every claim is dated and quoted, and every outcome was checked against a primary surface on 8 August 2026.

Verdicts are editorial judgements on delivery against a stated commitment. They are not quality assessments, and they are not weighted by how much each promise mattered.

Two things about the shape of that table are worth saying out loud.

First, after all of that, exactly one outright broken promise survives. We began from the most notorious failures we could recall, applied a test that any of them should have passed, and finished with a single clean miss, one partial, and two kept. That is not a defence of the industry — it is a statement about the difference between what is remembered and what is on the record.

It is worth dwelling on the one that survived, because it is stronger than a missed deadline. Musk named a model and a six-month window in August 2025. When that window expired in February 2026 he was asked directly whether Grok 3 would be open-sourced, and answered “Yes”. Six months after that, the weights still do not exist. The commitment was made, restated at its due date, and remains unmet — which is precisely the shape a promise ledger is built to record, and precisely the shape almost nothing else in this research turned out to have.

Second, the same vendor appears as both the one broken promise and one of the two kept ones. xAI makes dated public commitments more often than most companies, which means it both delivers against them and misses them in public. A ledger like this therefore rewards vagueness: a company that never names a date can never be recorded as late, and the quietest vendor scores best by saying nothing. That is a real limitation of the format, and we would rather state it than let the table imply otherwise.

Why so few

Vendor communication comes in two shapes, and neither can be adjudicated.

The first is announcement as delivery. The thing ships the day it is described. There is no interval during which a promise is outstanding, so there is nothing to check later. Most model launches work this way.

The second is directional sentiment. “We believe in open models.” “We were on the wrong side of history.” These are real positions and they shape expectations, but they name no deliverable and no date, so no outcome can falsify them.

The ledgerable shape — a specific thing, by a stated time — is the exception, not the rule. That is why Musk’s six-month Grok 3 commitment stands out: it named a model and a timeframe, which almost nothing else does.

Where we were wrong

What this does not show

How to check us

1. Every promise above is quoted with its date. Search the quote.
2. Every outcome was verified on a primary surface — the vendor’s own model repository or product page — not on reporting about it.
3. Where the claim and the outcome come from different evidence classes, the table says so.
4. If you find a dated, quotable commitment we missed, we want it. That is the hard part of this work, and the list above is certainly incomplete.

We are continuing to collect rows against the same three-part test. If the ledger grows enough to stand on its own, it will be published as one. For now the honest result is the pattern: the gap between “everyone knows they promised it” and “they said it, on this date” is wide, and it is where nearly all of the story turned out to be.