Is the AI productivity boom real or are we measuring the wrong things? - discuss

[Forge37]

METR published a survey last week showing technical workers self-report 3x speed gains and 2x value gains from AI tools in 2026. The same report noted that METR's own staff gave the lowest change estimates of any group surveyed and suspected high estimates are overstated.

Are the gains real or are we just doing more of the same work faster?

[Sequence87]

The 3x speed number is real for specific tasks and meaningless as a general claim. Speed on what? Drafting boilerplate is not the same as designing a system

[Ronan_34]

The value multiplier is the honest number and 2x is plausible for knowledge work where AI handles the low-value assembly tasks

[Warden]

My output volume is up. My quality of thinking is not. Nobody is measuring the second thing

[Idle Mila]

The verification overhead is the hidden cost nobody includes. Everything the model produces needs checking. That time comes out of the gain

[RomanReigns02]

Companies measuring productivity by output volume are going to make terrible decisions based on that data and they will not know why for two years

[AlexandrZakharyan]

For coding specifically the gains are closest to real because tests provide cheap ground truth. Natural language outputs are much harder to validate

[Midnight Georgia]

The honest framing is AI raises the floor not the ceiling. Mediocre work gets faster. Exceptional work requires the same exceptional person