This is very cool but it's not quite what I expected out of economic primitives.

I expected to see measures of the economic productivity generated as a result of artificial intelligence use.

Instead, what I'm seeing is measures of artificial intelligence use.

I don't really see how this is measuring the most important economic primitives. Nothing related to productivity at all actually. Everything about how and where and who... This is just demographics and usage statistics...

The title actually cringes me out a bit, it reads like early report titles in academia where young students (myself no doubt incl back when) try their hardest at making a title sound clever but in actuality only achieve obscuration of their own material.

> This also highlights the importance of model design and training. While Claude is able to respond in a highly sophisticated manner, it tends to do so only when users input sophisticated prompts.

If the output of the model depends on the intelligence of the person picking outputs out of its training corpus, is the model intelligent?

This is kind of what I don't quite understand when people talk about the models being intelligent. There's a huge blindspot, which is that the prompt entirely determines the output.

Skimmed, some notes for a more 'bear' case:

* value seems highly concentrated in a sliver of tasks - the top ten accounting for 32%, suggesting a fat long-tail where it may be less useful/relevant.

* productivity drops to a more modest 1-1.2% productivity gain once you account for humans correcting AI failure. 1% is still plenty good, especially given the historical malaise here of only like 2% growth but it's not like industrial revolution good.

* reliability wall - 70% success rate is still problematic and we're getting down to 50% with just 2+ hours of task duration or about "15 years" of schooling in terms of complexity for API. For web-based multi-turn it's a bit better but I'd imagine that would at least partly due to task-selection bias.

> These “primitives”—simple, foundational measures of how Claude is used, which we generate by asking Claude specific questions about anonymized Claude.ai and first-party (1P) API transcripts

I just skimmed but is there any manual verification / human statistical analysis done on this or we just taking Claude’s word for it?

I'm not an economist so can someone explain whether this stat is significant:

> a sustained increase of 1.0 percentage point per year for the next ten years would return US productivity growth to rates that prevailed in the late 1990s and early 2000s

What can it be compared to? Is it on the same level of productivity growth as computers? The internet? Sliced bread?

Every single AI economic analysis talks about travel planning but none of the AI labs have the primitives (transit routing, geocoding, etc.) in a semantic interface for the models to use.

> How is AI reshaping the economy?

oh I know this one!

it's created mountains of systemic risk for absolutely no payoff whatsoever!

I got halfway through then asked the LLM to extract any surprising or particularly interesting results, below. Verified each one manually.

1. Accelerated US Diffusion vs. Global Stagnation

The report estimates that usage parity across US states could be achieved in 2–5 years. This rate of diffusion is approximately 10 times faster than that of major 20th-century technologies (e.g., electricity or the automobile). However, this rapid convergence is not observed globally. International adoption remains strictly correlated with GDP per capita, with no evidence of lower-income countries "catching up." This suggests AI may currently exacerbate rather than narrow the digital divide between nations.

2. The Education Mirror Effect

A high correlation (r > 0.92) exists between the education level required to write a prompt and the education level of Claude's response. This implies that the model's sophisticated capabilities are only unlocked by users who already possess high formal education. Rather than acting as a cognitive equalizer, AI appears to function as a capital multiplier for existing high-skill workers.

3. Productivity "Haircut" via Reliability

While raw "speedup" data suggest significant time savings (e.g., a 12x speedup for college-level tasks), these gains are substantially offset by task failure. Adjusting aggregate productivity growth estimates for task reliability reduces the projected impact from 1.8 percentage points to approximately 1.0 percentage point of annual labor productivity growth. This 44% "haircut" highlights that current model unreliability is a primary bottleneck for macroeconomic impact.

4. Selective Deskilling of White-Collar Work

The report finds that AI disproportionately covers high-education tasks. In a task-displacement model, this leads to a net "deskilling" effect for several professions. For example, Technical Writers may lose their most complex analytical tasks to AI, leaving behind only routine illustrative or observational work. Conversely, occupations like Property Management may experience "upskilling" as AI handles bookkeeping, leaving managers to focus on high-stakes negotiations.

5. Multi-Turn "Task Horizon" Extension

The "task horizon"—the maximum task duration at which AI achieves a 50% success rate—varies wildly by interface. For single-turn API interactions, the horizon is 3.5 hours. For multi-turn Claude.ai conversations, it extends to 19 hours. This suggests that the iterative, human-in-the-loop chat interface is significantly more effective at managing complex, long-duration tasks than programmatic automation.

Critical Uncertainties:

Temporal limitations: The data was collected over a single week in November 2025 and predates the release of Opus 4.5.

Self-selection bias: Success rates reflect only the tasks users choose to bring to AI. If users avoid tasks they expect the AI to fail at, success rates are artificially inflated.

Measurement of "Education Years": The report uses a Ridge regression model to predict education requirements from task embeddings. If the training data (BLS occupation levels) is a lagging indicator of actual skill requirements, the deskilling/upskilling analysis may be skewed.

All of this performative bullshit coming out of Anthropic is slowly but surely making them my least favorite AI company.

We get it guys the very scary future is here any minute now and you’re the only ones taking it super seriously and responsibly and benevolently. That’s great. Now please just build the damn thing