The useful result
A one-line result for your own task, written as a count out of 5 with a review step, and a way to read hours-of-work claims.
- Read an AI "task horizon" or "hours of work" claim with three questions.
- Test one task of your own five times and score it before looking at the answers.
- Write your result as a count, such as "3 of 5", instead of "AI can do it".
You may have seen charts that say AI can now do tasks that take a human expert hours. The chart behind many of these claims comes from METR, a research group, and it measures something narrower than the headline.
METR defines the 50%-time horizon as the task length, measured by how long a human expert takes, at which an AI agent is predicted to succeed half the time. It does not say the AI works alone for that many hours, and it does not say it succeeds every time. METR's own page warns that measurements above 16 hours are unreliable with its current task suite, and says the page is no longer actively updated (last updated 8 September 2026). It lists GPT-5 at around 2 hours 17 minutes.
METR's January 2026 update (Time Horizon 1.1) estimates a doubling time of about 130.8 days for models since 2023, with a range of 107 to 161 days, and about 196.5 days for the whole stitched trend. Its suite has 228 tasks, 31 of them longer than 8 hours, and human baseline times were measured for only 5 of those 31. This lesson gives you three questions for reading any "hours of work" claim and a five-run test for your own task.
Before you start
- One task you do that takes about an hour and has a clear way to check the result.
- Any free AI chat assistant.
Review evidence
Documentation-based review · 2026-11-02
Sources opened and checked on 2 November 2026. Figures are dated and describe what the sources reported, not rules. Prompts and examples are illustrative; no model output or outcome is claimed.
Read the claim with three questions
Find a claim about AI doing hours of work. Copy it exactly. Ask three questions. First: success how often? A 50%-time horizon means about half the time. Second: whose time? It is how long a human expert takes, not how long the AI runs. Third: how was it measured? METR's Time Horizon 1.1 uses 228 tasks and measured human baselines for only 5 of its 31 tasks longer than 8 hours, and METR warns above 16 hours is unreliable. Then paste the claim into an AI chat and ask which of these details it leaves out. Do not ask the AI whether the claim is true.
Below is a claim I read about AI and "hours of work". Do NOT say whether it is true. Do NOT search. Using only my text, list what the claim leaves out under these headings: 1) How often the AI succeeds. 2) Whose time the hours refer to. 3) What kind of tasks were used. 4) How many tasks or people were measured. 5) Who made the claim and how they could know. Quote the words in the claim that relate to each heading, or write "not stated".
CLAIM:
[PASTE THE CLAIM]Example / expected result
Illustrative example; not a recorded model result.
Claim: "Models now complete more than a third of day-long research tasks with no human help."
- Success: "more than a third" (of tasks, not of attempts per task).
- Whose time: "day-long" (not stated whose day).
- Kind of tasks: "research tasks" (no detail).
- How many measured: not stated.
- Who and how: not stated in the claim.
Check your result
Every quoted phrase is really in the claim you pasted. Anything the claim does not say is marked "not stated", not guessed.
Test your own task five times
Choose one task you know well that takes you about an hour, with a result you can check, such as summarising a document you know or drafting a table. Before you run anything, write your success rules: what makes a result fully good, partly good or a fail. Then run the same prompt five times, each in a fresh chat, and score each result against your rules. Keep your rules fixed. If you change them after seeing a result, you are fooling yourself. Use only material you are allowed to share with an AI tool.
I will test an AI on one task five times. Turn my success rules into a scoring table with 5 empty rows (Run 1 to Run 5) and these columns: Full success / Partly / Fail / What went wrong. Do NOT score anything and do NOT do the task. Repeat my rules at the top of the table in my own words.
TASK:
[ONE SENTENCE]
MY RULES:
Full success = [..]
Partly = [..]
Fail = [..]Example / expected result
Illustrative example; not a recorded model result.
Task: Summarise a 6-page report in 5 bullets. Full success = all 5 bullets match the report and no invented figure. Partly = 4 match, none invented. Fail = any invented figure or a missing main point. | Run | Full | Partly | Fail | What went wrong | | 1 | | | | |
Check your result
You wrote the success rules before the first run, and you scored all five runs against the same rules.
Write the result as a count, and add a review step
Count your results. Write one sentence in this form: "On my task, [task], this tool fully succeeded X of 5 times, partly Y times and failed Z times." Do not turn it into "AI can do this". Then decide the review step you will use every time, for example checking every figure against the source. Compare your count with the claim you read in step 1: the claim may be true for its own tasks and say nothing about yours. Repeat the five-run test in three months to see your own trend.
From the scoring table below, write exactly two sentences for a colleague. Sentence 1 must state the task and the count in the form "fully succeeded X of 5 times, partly Y, failed Z". Sentence 2 must name the most common thing that went wrong. Do not use the words "can do", "reliable" or "ready". Do not add any number that is not in the table.
TABLE:
[PASTE YOUR FILLED TABLE]Example / expected result
Illustrative example; not a recorded model result.
On my task of summarising a 6-page report in 5 bullets, the tool fully succeeded 3 of 5 times, partly 1 time and failed 1 time. The most common problem was a figure that was not in the report.
Check your result
The sentence contains the count and the task, no claim like "can do", and no number that is not in your table.
When it goes wrong
A task-horizon chart measures a set of tasks chosen by researchers, not your work. Five runs are a small sample, so your count is a rough signal, not a measurement. Do not paste confidential work into a chat tool. If you change your success rules after seeing results, the test no longer means anything. The AI can mis-score or invent details, so score the runs yourself.
Make it your own
Take one headline about AI and "hours of work" and run step 1 on it. Write down the two details it leaves out.
Useful next resources
Sources and review
- METR: time horizons page (last updated 8 September 2026)
- METR: Time Horizon 1.1 (29 January 2026)
- The Next Web: OpenAI DevDay keynote and the AI research intern claim (29 September 2026)
Last checked Nov 2, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical exercises based on checked sources. Examples are illustrations, not recorded model results.
