“This is the weirdness of the moment. The general public has mostly not interacted with [AI]. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.” – Andrej Karpathy – AI guru, Anthropic

The defining problem is not simply whether artificial intelligence works. It is that different groups are forming radically different judgements about what works, how quickly capability is improving, and what the technology means for their own work. A casual user encounters an unreliable conversational interface, while a specialist may be integrating increasingly capable systems into coding, research, analysis, or operational workflows. Both observations can be accurate at the same time, which makes public debate unusually vulnerable to confusion.^1,2

That split is partly a matter of exposure. Adoption is growing quickly, but it is still uneven across populations, occupations, organisations, and use cases. A nationally representative US survey found that nearly 40 percent of people aged 18 to 64 had used generative AI by late 2024, while 23 percent of employed respondents had used it for work during the previous week and 9 percent used it every working day.³ Later tracking reported that 37,4 percent of employed respondents had adopted the technology for work by August 2025, alongside 54,6 percent of the wider working-age population using it at work or home.²

These figures describe adoption, not competence or impact. A person who asks a chatbot for a recipe, a student who requests an explanation, a marketer who drafts copy, and a software engineer who uses an agent to inspect a repository are counted within the same broad category. Their experiences are not comparable. The first may judge the system by fluency and occasional factual errors; the last may judge it by the number of useful changes it can make across a large codebase. Segmentation by task is therefore more informative than a single user percentage.

Why the productivity experience diverges

Workplace gains depend on the relationship between a model, a task, and the surrounding process. When the task is well specified, digitally accessible, repetitive, and easy to verify, AI can remove substantial friction. When success depends on tacit knowledge, ambiguous objectives, accountability, or long chains of dependent decisions, the same system may create more checking work than it removes. The relevant question is not whether a model is generally intelligent, but whether it can reliably complete a particular part of a workflow at an acceptable cost.

Available evidence supports both modest and substantial accounts of impact. In late 2024, respondents to the Real-Time Population Survey reported time savings equivalent to 1,4 percent of total work hours, while users saved considerably more during the hours in which they actually used the technology.³ A Federal Reserve analysis estimated that the implied aggregate productivity effect could reach 1,1 percent, while noting that the gain was concentrated among users and particular occupations.⁴ By August 2025, another analysis estimated aggregate time savings at 1,7 percent of total worked hours and suggested a possible productivity increase of up to 1,3 percent since the arrival of ChatGPT.²

Those estimates should not be mistaken for settled measurements of long-term economic output. They rely partly on self-reported time savings, and saved time can be absorbed by additional work, higher quality standards, organisational delay, or new verification requirements. A worker may complete a draft more quickly but spend the difference reviewing errors. A company may gain capacity without reducing headcount. Productivity can rise at the task level while remaining difficult to observe in national statistics.

The capability curve and the interface problem

The most consequential source of divergence is the distance between publicly available interfaces and the strongest systems used by advanced practitioners. Consumer products often expose a general-purpose assistant designed to be safe, accessible, and responsive across many contexts. Specialist users may instead work with better models, larger context windows, tool access, custom data, evaluation harnesses, persistent memory, and multi-step agents. The resulting difference is not merely a matter of expertise. It is a difference in the system being evaluated.

Technical progress has also become uneven across capabilities. The Stanford AI Index reported that performance on SWE-bench rose from 4,4 percent in 2023 to 71,7 percent in 2024, while noting that difficult benchmarks still expose substantial weaknesses.⁵ Its subsequent reporting described performance approaching 100 percent on SWE-bench Verified in 2025, although autonomous software engineering and computer-use tasks remained areas where systems could fall below human baselines.⁶ The UK AI Security Institute likewise reported rapid improvement in tested frontier capabilities, including a rise in success on apprentice-level cyber tasks from 9 percent in late 2023 to 50 percent in its later testing.⁷

Such benchmark movement can feel vertical to people whose work is close to the frontier, particularly when several improvements arrive together: better reasoning, stronger tool use, longer context, lower costs, and more reliable execution. Yet benchmark scores do not automatically translate into dependable autonomy. A system can solve a constrained test while failing in an unfamiliar environment, misunderstanding a requirement, or producing an answer that is difficult to audit. The apparent acceleration is therefore real in some dimensions, but its practical value remains conditional.

Three realities operating together

The public reality is shaped by visibility and disappointment. People encounter generated text, images, or answers in products that may be limited by access, poor prompting, weak integration, or cautious product design. Errors are memorable, especially when they concern basic facts. Because most people do not routinely compare models or construct evaluation pipelines, they may reasonably infer from these encounters that progress has been incremental or cosmetic.

The professional middle experiences a more modest uplift. Many workers use AI for brainstorming, summarisation, translation, drafting, meeting notes, or routine analysis. These applications can save minutes without changing the structure of a job. They may improve convenience rather than create a dramatic change in output. Organisational policies, data restrictions, procurement processes, and uncertainty about responsibility further limit the effects. Adoption can therefore be widespread while transformation remains shallow.

A smaller frontier-facing group experiences a different phenomenon. These users repeatedly test systems against difficult tasks, connect them to tools, and redesign processes around their strengths and limitations. Their learning compounds because each successful experiment reveals another use case. They may move from asking for an answer to assigning a bounded objective, providing context, checking intermediate results, and letting the system execute part of the workflow. The gain comes from the combination of model capability and workflow redesign, not from the chat interface alone.

Objections and limits

There are good reasons to resist dramatic interpretations. Capability demonstrations can overrepresent favourable tasks, omit failed attempts, and blur the difference between assistance and independent responsibility. Frontier systems remain prone to hallucination, brittle reasoning, security failures, and inconsistent behaviour. Access is also unequal: expensive models, specialised tools, high-quality data, and technical expertise are not distributed evenly. A small group may therefore experience a genuine transformation without that experience representing the median worker or consumer.

There is an equally serious risk in treating ordinary experience as a reliable forecast of the frontier. Current interfaces are not fixed descriptions of future use. Costs can fall, models can become more dependable, and tools can make complex capabilities easier to deploy. The gap between expert and general users may narrow, but it could also widen if the most valuable advances require engineering skill, proprietary data, or organisational authority. Policymakers and business leaders therefore need measures that distinguish access, frequency, task completion, quality, supervision, and economic value rather than relying on a single adoption headline.

The strategic implication is that AI change will not arrive as one uniform wave. It will appear first where tasks are measurable, digital, and easy to decompose, then spread through organisations that can redesign roles and controls around the technology. Public scepticism, moderate workplace gains, and intense frontier optimism can coexist because they describe different layers of the same transition. The important question is where a system sits in a workflow, what evidence supports its reliability, and who bears the cost when its confidence exceeds its competence.

 

References

1. The State of Generative AI Adoption in 2025 | St. Louis Fed – https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025

2. The Generative AI Adoption Tracker – pw.hks.harvard.edu – 2025-11-13 – https://pw.hks.harvard.edu/post/the-generative-ai-adoption-tracker

3. Frontier AI Trends Report by The AI Security Institute (AISI) – https://www.aisi.gov.uk/frontier-ai-trends-report

4. The Impact of Generative AI on Work Productivity – 2025-10-07 – https://www.frbsf.org/research-and-insights/publications/system-research-st-louis-fed/2025/10/impact-generative-ai-work-productivity/

5. Karpathy Exposes AI’s Hidden Truth: 2 Billion People Deceived Globally, 20 Million Trapped in Agent Vertigo – 2026-10-12 – https://eu.36kr.com/en/p/4022171722944643

6. Technical Performance – https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_2_technical.pdf

7. Generative AI, Productivity and the Future of Work – https://www.stlouisfed.org/open-vault/2025/oct/generative-ai-productivity-future-work

8. The 2026 AI Index Report | Stanford HAI – https://hai.stanford.edu/ai-index/2026-ai-index-report

9. Technical Performance – https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance

10. 5 key findings from our first Frontier AI Trends Report – https://www.aisi.gov.uk/blog/5-key-findings-from-our-first-frontier-ai-trends-report

11. hai.stanford.edu › ai-index › 2025-ai-index-reportThe 2025 AI Index Report – Stanford HAI – https://hai.stanford.edu/ai-index/2025-ai-index-report

12. The Rapid Adoption of Generative AI | NBER – 2024-09-23 – https://www.nber.org/papers/w32966

13. AI’s Great Flattening: What Happens when Everyone Is … – https://www.wisdomtree.com/us/insights/blog/ais-great-flattening-what-happens-when-everyone-is-state-of-the-art

14. The Impact of Generative AI on Work Productivity | St. Louis Fed – https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity

15. Generative AI in the Workplace: A Systematic Review of … – https://rsisinternational.org/journals/ijriss/uploads/vol10-iss3-pg4426-4437-202604_pdf.pdf