From claim to proof.
How many tasks have reached each step. A task climbs only when new evidence of the next kind is published.
How stages work- 1Claimed
- 2Launched
- 3Independently tested
- 4Tested with real people
- 5Regulator-cleared
- 6Widely used with a verified record
By subject.
Furthest along first within each subject. Open a task for every piece of evidence behind its stage.
Health4 tasks
Spot breast cancer on a mammogram
Stage 5 of 6Regulator-cleared
Transpara has been FDA-cleared since 2018, most recently in Aug 2026
Check a mole for skin cancer
Stage 5 of 6Regulator-cleared
EU Class III CE mark lets DERM clear benign cases alone (company)
Answer my health questions
Stage 4 of 6Tested with real people
Randomized trial: chatbot users made no better health decisions
Be my therapist
Stage 4 of 6Tested with real people
Pilot trial: therapy chatbot did no better than ChatGPT on depression
Money and legal4 tasks
Do my taxes
Stage 3 of 6Independently tested
Tax-prep chatbots were unhelpful or wrong as much as half the time
Give me financial advice
Stage 3 of 6Independently tested
Which? test: chatbot money answers were 58–67% accurate
Do legal research
Stage 3 of 6Independently tested
Leading legal AI tools gave wrong information over 17% of the time
2,199 court decisions with AI-invented content
Write my will or legal documents
Stage 1 of 6Claimed
DoNotPay marketed itself as 'the world's first robot lawyer'
Home and daily life5 tasks
Drive me across town
Stage 5 of 6Regulator-cleared
Driverless Waymo rides approved for more of the Bay Area
Plan and book my trip
Stage 3 of 6Independently tested
2024 benchmark: GPT-4 made a valid travel plan 0.6% of the time
Shop and buy for me
Stage 2 of 6Launched
ChatGPT let US users buy from Etsy sellers inside the chat (Sep 2025)
Translate a conversation live
Stage 2 of 6Launched
Apple launched Live Translation on AirPods in beta (Sep 2025)
Do my household chores
Stage 1 of 6Claimed
1X says its NEO home robot can do chores; pre-orders at $20,000
School2 tasks
Work4 tasks
Answer customer-service chats and calls
Stage 4 of 6Tested with real people
Field study of 5,179 support agents: AI assistant raised output 14%
Make scientific discoveries
Stage 3 of 6Independently tested
METR: AI-era maths results up somewhat; no clear algorithm speed-up
Tell me what's in the news
Stage 3 of 6Independently tested
Almost half of AI assistants' news answers had a significant issue
Read and understand a long document
Stage 3 of 6Independently tested
2024 book test: GPT-4o 55.8% on claims people found easy
Coding2 tasks
Agents2 tasks
Creative1 task
Suggest a task.
We add tasks people ask about once there is evidence to place them. Every one of the 6 steps has a written rule.
Suggest a taskVideos on what AI can do.
Talks, interviews and explainers, each with our short summary. They play in YouTube's own player.
All videos on what AI can doAre AI benchmarks doomed?
Epoch AI researchers Greg Burnham and Tom Adamczewski discuss whether AI benchmarks are in trouble as they go out of date quickly and cost more to build. They cover the gap between benchmark scores and real-world performance and whether a benchmark for AGI could exist.
AXRP · Jan 2, 202647 - David Rein on METR Time Horizons
METR researcher David Rein discusses how METR measures AI 'time horizons', the length of tasks models can complete, and why it focuses on software engineering. The conversation also covers whether AI progress is speeding up and the open questions with the method.
Google DeepMind · Nov 18, 2025Gemini 3: Turn a research paper into an interactive website
Google DeepMind's launch demo shows Gemini 3 in Google AI Studio reading a materials science research paper and writing code for an interactive guide with 3D visuals.
Made by the company itself