Google I/O Reality Check
Every year Google walks on stage and promises the future. Some of it ships. Some of it ships smaller. Some of it ships under a different name. And some of it never ships at all.
I read every keynote, tracked every announcement, and went looking for what actually happened. 589 announcements across 17 years of I/O. Here’s the receipts.
(on time, late, or rebranded)
never shipped, abandoned
The trend
Year by year
By category
Hall of Shame
The most damning entries: pure vapor first, then post-launch shutdowns, then promises that were quietly shrunk after the keynote.
SignGemma sign-language model
Open model translating American Sign Language to English text
Preview opened to LSPs, accessibility researchers, and Deaf community members via Google's AI Developer portal. Broader release was planned for Q4 2025 but never materialized — no public model card, no open weights, and no GA path 12 months on.
Announced as an 'open model' but never shipped one — still feedback-only access a year later.
Gemini Diffusion (text diffusion model)
Experimental text-diffusion research model, 5x faster than 2.5 Pro for code; trusted testers, 'broader release in coming months'
Waitlist demo opened at I/O. Still labeled experimental / trusted-testers-only as of mid-2026, with no general API availability, no pricing, and no product surface — the 'coming months' broader release never happened.
Research preview pitched as a product line; 12 months later still no public access — calling this anything other than vapor flatters Google.
Sparkify (question → animated short video)
Convert questions into animated 2-minute videos via Veo 3 and Gemini; waitlist at I/O; 'coming to Google products later this year'
12 months later still a waitlist-gated Google Labs experiment at sparkify.withgoogle.com with no public on-ramp. Capabilities were partly absorbed into NotebookLM Video Overviews and Flow, but the named Sparkify product never launched.
A year of waitlist with no public release and no roadmap update — meets the rubric's definition of vapor.
Illuminate (research papers to AI podcasts)
Experimental tool turning research papers into AI-narrated audio conversations.
Available as a free Labs experiment with limited scope (primarily arXiv CS papers). Effectively cannibalised by NotebookLM's Audio Overviews. Still 'experimental' two years later with no visible roadmap or graduation from Labs.
Two years stuck in Labs with no meaningful progress — fails the 12-month public-progress bar.
Universal Translator (video dubbing with lip-sync)
Research demo: AI that transcribes, translates, regenerates voice, and lip-syncs the speaker on video.
Explicitly limited to 'authorized partners' at announcement. Never released publicly. The capability resurfaced piecemeal in YouTube auto-dubbing (without lip-sync) in 2024.
Google literally flagged misuse risk on stage — never shipped to consumers.
AR Translation Glasses (Project Iris)
Famous keynote-ending teaser of prototype AR glasses doing real-time speech translation captions.
Limited public field tests in the US (Aug 2022) and Canada (Oct 2022), then Project Iris was shelved in early 2023 amid layoffs. No consumer product ever shipped.
Google pivoted to building the AR software platform for partners like Samsung instead.
Scene Exploration in Lens
Pan camera across a scene to get overlaid info on every object — 'Ctrl+F for the world'.
Never shipped as the named 'Scene Exploration' feature. Camera-pan visual search ideas eventually surfaced via Circle to Search (2024) and Gemini Live video.
Spiritual successors exist but the demo as shown never landed.
MUM (Multitask Unified Model)
1000x more powerful than BERT; would understand text, images and 75 languages and power dramatic new Search experiences.
Google has explicitly stated MUM is 'not used for general ranking.' Used only narrowly for COVID-19 vaccine info, About This Result, and multisearch in Google Lens. The marquee 1000x-BERT product never materialized; the multilingual cross-modal search vision was effectively abandoned and overtaken by LLMs/Gemini and AI Overviews. The MUM brand has largely disappeared from Google communications.
A handful of MUM-adjacent launches (multisearch, About This Result) shipped, but the headline promise of MUM as a search ranking/understanding breakthrough never happened.
AR translation glasses concept (in-keynote tease)
I/O 2021 closing reel and demos suggested real-time translation/AR overlays as part of Google's near-future hardware ambitions (precursor to the Project Iris/AR-glasses pitch).
Real AR translation glasses prototype was shown at I/O 2022, but Project Iris was shut down in 2023 and Google pivoted to providing Android XR for partners (Samsung's Project Moohan, 2025). No Google-branded consumer translation AR glasses have shipped.
Closest live successor is Android XR on partner hardware — not the original first-party promise.
Live Relay
Phone listens and speaks on behalf of deaf/hard-of-hearing users during regular phone calls; user types, phone speaks, all on-device.
Demoed at I/O 2019 as a research project. Never shipped as a standalone product. Some underlying tech surfaced years later in Pixel's Call Assist / Call Screen features, but Live Relay as announced was never released to users.
Always framed as research, but headlined as if shipping soon.
Slices
Interactive snippets of app UI that would appear inline inside Google Search and Assistant — e.g., a Lyft Slice showing ride prices right in Search.
The androidx.slice APIs technically shipped, but the marquee Google Search integration — the entire point of the announcement — was quietly paused around 2021 and never returned. By Android 13 the feature was effectively invisible to users; the App Actions integration was marked deprecated.
APIs exist as a developer ghost-town. The promised Search integration never reached users — textbook vapor at promised scope.
Project Ara
Modular smartphone with hot-swappable components. Google explicitly promised a developer edition shipping in Q4 2016 and a consumer device in 2017.
Project Ara was cancelled outright on September 2, 2016, less than four months after I/O. No developer kit was ever shipped to outside developers and no consumer device was ever produced.
Killed faster than almost any other I/O 2016 announcement.
Methodology
I went through every Google I/O keynote and developer keynote from 2008 to 2025 (no I/O in 2020 — cancelled for COVID) and pulled every notable announcement: features, products, models, dev-platform releases. Each one was classified into one of seven buckets:
- Shipped — delivered substantially as promised, within ~12 months.
- Shipped late — delivered, but materially later than the stated or implied timing.
- Scaled back — shipped, but materially smaller than what was demoed.
- Rebranded — shipped, but under a totally different product name (still a win for Google).
- In progress — still actively being developed, not yet shipped at the promised scope, not abandoned.
- Vapor — promised, never shipped at all, no longer being worked on.
- Killed — shipped, later discontinued.
The Reality Score is the share of resolved announcements (so excluding the still-cooking ones) that count as a win: shipped on time, shipped late, or rebranded. Everything else is a miss.
Every entry has at least two sources — typically one Google announcement (to ground what was promised) and one independent report (to ground what actually happened). All data is in a JSON file if you want to verify or remix.
Where I could be wrong: judgment calls on "scaled back" vs "shipped", and the cutoff for "in progress" vs "vapor" on recent announcements. Open an issue or DM and I'll correct.