The evidence desk

AI moves quickly.
Responsibility moves slowly.

What AI bosses overstate is often the jump from capability to dependable responsibility. This is my interpretation of the evidence, not proof of bad faith.

Sources checked 20 October 2026Dated archive, not a live feed

01 / Current research

Measure the work behind the headline.

01

/ METR AI Usage Survey

Faster is not the same as more valuable

The claim
In a survey of 349 technical workers, respondents reported a median 1.4 to 2 times as much value and 3 times the speed when using AI tools.
What it does not prove
This was a convenience sample based on self-reports. It does not establish causal, measured gains across a business.
My operating rule
Count useful outputs accepted after review, not generated volume or speed alone.
02

/ METR Uplift Update

Even the anti-hype needs updating

The claim
METR revisited its widely repeated early 2025 result in which experienced developers were 19 percent slower with AI tools.
What it does not prove
Later results were affected by selection bias. METR says greater acceleration is likely, but cannot quantify it reliably. The earlier figure is not a current universal result.
My operating rule
Keep the measurement date, task population and uncertainty attached to every productivity number.
03

/ Anthropic Economic Index

A completed task is not a completed company

The claim
Anthropic analysed selected tasks in November 2025 conversations and estimated task success with Claude-classified interactions and human feedback loops.
What it does not prove
Those estimates are not independently audited business outcomes. A task can be completed while the surrounding commercial process still fails.
My operating rule
Trace accepted work through its handoffs, exceptions and final human decision.

02 / Claims archive

Keep the condition attached to the prediction.

/ Sam Altman, Reflections

Joining a workflow is not owning its consequences

The claim. Sam Altman wrote that he believed the first AI agents might join the workforce in 2025 and materially change company output.

The limit. That was a conditional forecast, not evidence that agents could own a company or its consequences.

Operating rule. Measure accepted work and human interventions instead of assigning anthropomorphic job titles.

/ Dario Amodei, Machines of Loving Grace

A scenario is not a timetable

The claim. Dario Amodei wrote that powerful AI could arrive as early as 2026, while allowing that it could take longer, and explored decades of biological progress compressed into years after that point.

The limit. The essay presents an explicitly speculative scenario. Its qualifications matter as much as its most dramatic numbers.

Operating rule. Treat forecasts as planning inputs, never as proof that a capability is dependable today.