A dated public record of what the AI industry promised, and what actually shipped. The same referee as Did It Actually Pay, pointed at the loudest promises of the decade.
What this is. The AI industry runs on the demo and the keynote. Almost no one comes back a year later to check whether the thing shipped as promised, on the terms promised, by the date promised. This record does only that. It writes down a specific, dated, checkable AI promise, waits, and grades it against what actually happened. Every fact links to a source you can open yourself.
We grade the promise, never the person or the company's worth. The question is never "is this hype." It is only "did the stated thing ship, on the stated terms, by the stated time." That is why the same rubric grades a trillion-dollar platform and a one-product startup, and why a kept promise earns a SHIPPED as readily as a missed one earns a DID NOT SHIP.
The scoreboard so far
4
Shipped
7
Partial
1
Too early
5
Withdrawn
12
Did not ship
Twenty-nine entries and counting. The mix is the point. A referee that only ever says one word is not refereeing.
How we grade
The full method is public: see How We Grade for exactly what qualifies as a promise, how each verdict is defined, and the sourcing standard. In short:
ShippedPartialDid not shipToo earlyWithdrawn
Shipped: the stated thing arrived on the stated terms, by the stated or a reasonable time. Partial: something shipped, but materially less, later, or on changed terms. Did not ship: the record shows it did not happen and will not. Too early: the clock is still running, and we commit to a public re-check date. Withdrawn: the promiser formally retracted or retired it. The verdict grades the claim, not any person. Every graded party may submit primary documentation and we will correct the entry.
The record
Feature announced vs shippedShipped, ten months late
Microsoft Recall for Copilot+ PCs
The promise
At the Copilot+ PC unveiling on May 20, 2024, Microsoft said Recall, a feature that periodically snapshots your screen so you can search past activity in natural language, would ship with Copilot+ PCs on June 18, 2024.
Source: Microsoft Copilot+ PC announcement, May 20, 2024.
The receipt
Microsoft pulled Recall from the June 18 launch on June 7, 2024 after security researchers showed the activity database was easily accessible, then re-released it to the general public on April 25, 2025 with encryption, opt-in, and Windows Hello gating. The feature did ship. Graded Shipped, with the honest note that it missed its committed date by roughly ten months.
Humane sold the AI Pin beginning in April 2024 at $699 plus a $24 monthly plan, positioned as a screenless, standalone device you would rely on in place of a smartphone.
Source: Humane launch, announced November 9, 2023, shipped April 2024.
The receipt
On February 18, 2025, HP acquired Humane's assets for about $116 million and the AI Pin was discontinued. On February 28, 2025, the devices were cut off from Humane's servers and lost calling, messaging, and AI functions. Less than a year after shipping, the promised standalone device no longer works.
AI announced vs shippedDid not ship (announced timeline)
Apple's more personalized Siri
The promise
At its developer conference on June 10, 2024, Apple announced a more personalized Siri with on-screen awareness and personal context as part of Apple Intelligence, presented for the iOS 18 cycle.
Source: Apple WWDC 2024 keynote.
The receipt
On March 7, 2025, Apple said the features would take longer than expected. At its June 9, 2025 conference it repeated that the work needed more time. On June 8, 2026, Apple re-announced the feature as Siri AI with general release scheduled for fall 2026. As of this writing it has not reached general users, roughly two years after the original announcement.
Source: as reported by TechCrunch (March 7, 2025) and MacRumors (June 9, 2025); Apple Newsroom, June 2026.
Next check: after iOS 27 general release, about October 2026.
Consumer AI hardwareShipped
The Bee AI memory wearable
The promise
Bee marketed a low-cost, always-listening wearable at $49.99 plus $19 a month that would record conversations and build reminders, to-do lists, and a personal memory.
Source: Bee product marketing.
The receipt
The device shipped on those terms and stayed in market, and on July 22, 2025, Amazon announced it was acquiring Bee, with development continuing under Amazon into 2026. The promised device reached buyers at the promised price. Graded Shipped, the counterexample to Humane in the same category.
Source: as reported by TechCrunch, July 22, 2025; Forbes, January 2026.
Consumer AI hardwarePartial
The Rabbit R1 large action model
The promise
At CES on January 9, 2024, Rabbit unveiled the R1 at $199 built around a large action model, software it said would learn to operate any app or website on the user's behalf.
Source: Rabbit CES 2024 launch.
The receipt
At its April 2024 launch the R1 supported only four services, Spotify, Uber, DoorDash, and Midjourney, not arbitrary apps. A web-based agent arrived later in 2024 as an admittedly experimental tool, and the company kept iterating into 2026. Materially less than promised, and later, though the product exists. Graded Partial.
On December 6, 2023, alongside the Gemini launch, Google published a video titled "Hands-on with Gemini" that showed the model appearing to converse in real time by live voice and streaming video, reacting instantly to drawings and objects.
Source: Google Gemini launch video, December 6, 2023.
The receipt
Within a day, Google acknowledged the video was built from still image frames and text prompts with latency edited out, and that the real-time spoken interaction shown did not occur as depicted. Graded Did not ship as depicted. This grades the demo video, not the underlying model.
On February 26, 2025, Amazon announced Alexa+, a generative-AI Alexa that would hold natural conversations and take actions, priced at $19.99 a month and free for Prime members, with early access rolling out within weeks.
Source: Amazon devices event, February 26, 2025.
The receipt
Early access began in waves from late March 2025 and passed one million users by June 23, 2025, but as of mid-2026 it remained in gated early access on select devices with an incomplete feature set. It shipped to real users near the promised window, and it is not yet the full public product that was announced. Graded Partial.
Tesla's coast-to-coast autonomous drive by end of 2017
The promise
On October 19, 2016, Elon Musk said Tesla would demonstrate a fully autonomous drive from Los Angeles to New York, with the car dropping off its passenger and then parking itself, by the end of 2017.
Source: Musk's remarks on the Tesla press call, October 19, 2016.
The receipt
The demonstration drive was never performed, in 2017 or since. As of mid-2026, no Tesla has completed the promised coast-to-coast autonomous demo under the terms described, and Tesla's Full Self-Driving in owner cars remains a supervised driver-assist system. Graded Did not ship.
Source: as reported by TechCrunch, October 19, 2016.
Robotaxi ambition vs outcomePartial
Tesla's driverless Austin robotaxi and 2025 scaling
The promise
Through 2025, Musk said Tesla would launch a driverless Austin robotaxi service with "no one in them" by June 2025, and scale to 500 or more vehicles in greater Austin and coverage of about half the US population by the end of 2025.
Source: Musk statements and Tesla earnings calls, 2025.
The receipt
Tesla launched a small Austin robotaxi pilot in June 2025, but with a Tesla safety monitor in the front seat, not empty, and received a Texas ride-hailing permit on August 8, 2025. By February 2026, independent tracking put it at roughly 42 vehicles at limited availability, still with safety monitors, operating only in Austin and the Bay Area. A supervised pilot shipped near the date. The dated scaling and driverless promises did not. Graded Partial.
Source: CNBC (August 8, 2025); status check by Electrek (February 16, 2026).
Autonomy promise, clock runningToo early
Robotaxis "widespread" across the US by end of 2026
The promise
In January 2026, Musk said Tesla robotaxis would be widespread across the United States by the end of 2026.
The clock is still running. As of mid-2026, the service operated in a small number of metros with limited fleets and human safety monitors. This is logged now, before the outcome is known, which is the whole point of the record. It will be graded on its date whatever the result.
Next check: December 31, 2026.
Robotaxi ambition vs outcomeWithdrawn
GM Cruise: one billion dollars in revenue by 2025
The promise
GM leadership publicly targeted about $1 billion in Cruise revenue by 2025, positioning the robotaxi unit as a major growth business.
Source: GM public statements on Cruise, 2023.
The receipt
After an October 2023 incident in San Francisco and a California DMV suspension of its driverless permits, GM announced on December 10, 2024, that it would stop funding Cruise's robotaxi development and wind the business down, abandoning the target a year before its date. The safety findings are the regulator's. Graded Withdrawn.
Source: as reported by NPR and CNBC, December 2024.
AI agent shipped vs retiredWithdrawn
OpenAI's Operator web agent
The promise
On January 23, 2025, OpenAI announced Operator, an agent that would autonomously perform browser tasks such as filling forms, placing orders, and booking appointments on the user's behalf.
Source: OpenAI announcement, January 23, 2025.
The receipt
Operator shipped as a research preview, then OpenAI folded its capability into ChatGPT and shut down the standalone Operator on August 31, 2025, about seven months later. The function continues inside ChatGPT, but the named product was retired. Graded Withdrawn.
Source: OpenAI deprecations documentation; announcement January 23, 2025.
AI agent claim vs outcomePartial
Cognition's "first fully autonomous AI software engineer"
The promise
On March 12, 2024, Cognition Labs launched Devin, marketed as the first fully autonomous AI software engineer, with a demo showing it completing a paid freelance job end to end.
Source: Cognition Labs launch, March 12, 2024.
The receipt
The benchmark figure Cognition cited was a genuine result, but independent analysis in 2024, and a study reported in January 2025 that found Devin completed 3 of 20 real tasks, showed the autonomy was overstated. The company itself succeeded commercially, acquiring Windsurf in July 2025 and raising at a large valuation in 2026. A real product shipped, the fully autonomous framing did not hold. Graded Partial.
Presto Voice: AI that "eliminated" human order-taking
The promise
From about November 2021, Presto Automation marketed its Presto Voice drive-thru AI as having eliminated the need for human order-taking.
Source: Presto press releases and filings.
The receipt
On January 14, 2025, the SEC issued a settled cease-and-desist order finding that Presto made materially false or misleading statements, that early units relied on a third party's speech technology, and that the vast majority of drive-thru orders required human intervention. The findings are the SEC's, and Presto settled without admitting or denying them. Graded Did not ship.
DoNotPay marketed an AI service as the world's first robot lawyer that could replace lawyers and generate valid legal documents for consumers.
Source: DoNotPay marketing.
The receipt
As part of Operation AI Comply, announced on September 25, 2024, the FTC charged that DoNotPay could not substantiate the claims and employed no attorneys, and the matter settled with a $193,000 payment and a requirement to warn users of the service's limits. The characterization is the FTC's. Graded Did not ship.
Source: FTC, September 25, 2024; FTC DoNotPay case file.
AI claim vs outcomeDid not ship
Builder.ai's "Natasha" app builder
The promise
Builder.ai marketed its assistant Natasha as building the first 80 percent of an app and making software creation as easy as ordering pizza, and reached a valuation of about $1.5 billion.
Source: Builder.ai marketing and investor materials.
The receipt
Reporting in 2025 described apps built largely by human engineers, and after a lender seized cash reserves the company filed for insolvency on May 20, 2025. US federal prosecutors subsequently subpoenaed records. The company collapsed. The revenue and deception characterizations are those of the reporters and investigators. Graded Did not ship.
In 2016, Geoffrey Hinton said people should stop training radiologists, because within five years, and at most ten, deep learning would do the job better than a human.
Source: Hinton's 2016 remarks, Toronto.
The receipt
Neither window delivered. The US radiologist workforce grew about 17 percent from 2014 to 2023, and by 2025 and 2026 the field reported record demand and compensation, ranking among the highest-paid medical specialties. The dated prediction did not happen. Graded Did not ship. AI is now a real tool inside radiology, which is a different claim than replacing the radiologist.
Source: as reported by The New Republic (2024); workforce data from the ACR and the Neiman Health Policy Institute.
CEO claim with a dateDid not ship
AI writing 90 percent of the code within six months
The promise
On March 10, 2025, Anthropic CEO Dario Amodei said at a public event that within three to six months AI would be writing 90 percent of the code, and within about a year essentially all of it.
Source: remarks reported March 2025.
The receipt
Both windows, roughly September 2025 and March 2026, have passed. No industry data shows AI authoring 90 percent of code, and independent reviews a year later found the economy-wide claim unmet, with AI assisting a meaningful but far smaller share. Graded Did not ship. AI coding assistance is real and growing, which is a different claim than 90 percent of all code.
On a podcast released January 10, 2025, Meta CEO Mark Zuckerberg said that probably in 2025 Meta and other companies would have an AI that could act as a mid-level engineer writing production code.
Source: podcast interview, January 10, 2025.
The receipt
Through mid-2026, Meta had not replaced its mid-level engineers with AI. The role persisted and Meta continued to employ and hire engineers. Graded Did not ship, on the stated 2025 timeline.
Source: as reported by Forbes (January 2025) and ITPro.
CEO claim vs outcomePartial
AI agents joining the workforce in 2025
The promise
In a blog post on January 6, 2025, OpenAI CEO Sam Altman wrote that in 2025 the first AI agents may join the workforce and materially change the output of companies.
Source: Sam Altman, "Reflections," January 6, 2025.
The receipt
OpenAI shipped agent products in 2025 and enterprises began deploying them, so agents did enter workplaces. But no AGI was declared, and evidence that agents materially changed company output at scale is thin. A real capability shipped, the larger half of the claim is unproven. Graded Partial.
Through 2024, Klarna and CEO Sebastian Siemiatkowski said its OpenAI-powered assistant did the work of 700 full-time agents, and that AI could shrink headcount from about 5,000 toward 2,000.
Source: Klarna and OpenAI statements, 2024.
The receipt
In May 2025, the CEO reversed course, telling Bloomberg the cost cutting had gone too far, that service quality dropped, and that Klarna had begun rehiring human agents. The company walked back its own claim. Graded Withdrawn.
Duolingo going "AI-first" and phasing out contractors
The promise
In a public memo in late April 2025, Duolingo CEO Luis von Ahn said the company would become AI-first and would gradually stop using contractors for work that AI could handle.
Source: Duolingo all-hands memo, April 2025.
The receipt
After public and user backlash, von Ahn walked it back in August 2025, saying the memo did not give enough context and that no full-time employees were laid off, and by April 2026 he also backed off requiring employees to be scored on AI use. Graded Withdrawn.
Source: as reported by TechCrunch (August 2025) and Fortune (April 2026).
Benchmark claim vs realityWithdrawn
Llama 4 Maverick ranked number two on the leaderboard
The promise
Launching Llama 4 on April 5, 2025, Meta touted the Maverick model's number two ranking on the LMArena chatbot leaderboard as evidence of top-tier quality.
Source: Meta Llama 4 launch, April 5, 2025.
The receipt
On April 8, 2025, LMArena said Meta had submitted a customized model tuned for human preference, not the public release, and that this did not match what it expects from model providers. When the unmodified released model was tested, it ranked around number 32. The touted ranking was withdrawn in effect. Graded Withdrawn.
At Google I/O on May 14, 2024, Google launched AI Overviews to all US users, positioning them as trustworthy at-a-glance answers at the top of search.
Source: Google I/O keynote, May 14, 2024.
The receipt
Within two weeks the feature produced viral false answers, including advising users to put glue on pizza and to eat a rock a day, and Google restricted it while acknowledging errors on May 30, 2024. The feature shipped broadly, the implied reliability did not hold. Graded Partial.
Unveiling Sora as a research preview on February 15, 2024, OpenAI signaled a public release of the text-to-video model within the year.
Source: OpenAI Sora announcement, February 15, 2024.
The receipt
OpenAI released Sora, as Sora Turbo, to the public for ChatGPT Plus and Pro users on December 9, 2024, inside the calendar year. The promised public release arrived. Graded Shipped.
Announcing AlphaFold 3 on May 8, 2024, Google DeepMind and Isomorphic Labs pledged a free AlphaFold Server for non-commercial research.
Source: Google DeepMind announcement, May 8, 2024.
The receipt
The server launched the same day, and after criticism that the code and weights were withheld, DeepMind released the AlphaFold 3 model code and weights for academic use in November 2024. The promised access was delivered. Graded Shipped, the science-side counterexample that shows the referee is not one-note.
In a December 2024 interview, Salesforce CEO Marc Benioff said the company was on track to deploy one billion AI agents by the end of 2025.
Source: Benioff interview, December 2024.
The receipt
Salesforce shipped and scaled Agentforce, reporting thousands of customers and about one billion dollars in related revenue in 2025, but it never verifiably confirmed the one billion agents figure, and analysts flagged adoption lagging the claim. A real product shipped, the specific number did not. Graded Partial.
Global Predictions marketed itself as the first regulated AI financial advisor, and Delphia claimed to use AI and machine learning to manage client money.
Source: firm marketing, prior to 2024.
The receipt
On March 18, 2024, the SEC charged and settled with both firms for false or misleading statements about their use of AI, its first AI-washing enforcement actions, with combined penalties of 400,000 dollars. The findings are the SEC's, settled without admitting or denying. Graded Did not ship.
At Davos in January 2025, Isomorphic Labs CEO Demis Hassabis said the company expected to have AI-designed drugs in human clinical trials by the end of 2025.
Source: Davos remarks, January 2025.
The receipt
At Davos in January 2026, the target was revised to the end of 2026, and as of mid-2026 no trial had been dosed, though the company raised 2.1 billion dollars in May 2026 to reach the clinic. The end-2025 deadline passed without a trial. Graded Did not ship, against that deadline.
Next check: December 31, 2026, the revised target.
The rules we hold ourselves to
Why you can trust the record
No money from the rated, ever. No payment, sponsorship, advertising, or affiliate relationship from any company that appears here, past or future. Our income comes only from readers and members, never from anyone we grade. This is the one rule a reviewer with a sales team cannot copy.
We grade the promise, not the person. We do not call anyone a scammer or a fraud in our own voice. Where a court or regulator has made a finding, we report that finding and attribute it to them, and we stop there.
Dated facts and primary sources only. Every receipt links to something you can open: a regulator order, a court filing, an official announcement, or the promiser's own words. We quote, we cite, we let you conclude.
We commit to check dates. A Too Early verdict names a date we will come back and re-grade, whatever the outcome. Writing the promise down before the outcome is known is the whole point.
Right of response and public corrections. Any graded party may submit primary documentation and we will update the entry. When we get something wrong, we correct it in the open, dated, and leave the record of the fix.
Educational record only, not investment, financial, or legal advice, and not a recommendation to buy, sell, or hold anything. Verdicts grade specific dated claims against the public record and are not statements about the character or intent of any person or company. Figures and outcomes are dated to their sources and may be revised; verify against the linked records. We hold no position in, and take no compensation from, anything named here.