AI Observatory / 20 September 2026
AI has moved from a research question into everyday tools and public decisions. Follow what systems could do, how people used them, and how institutions responded.
Explore the timeline ↘01 / Selected history
Twenty-three sourced moments, with six larger turning points. Select a point to see what happened, why it matters, and what it cannot prove. The spacing groups eras; it does not measure the pace of change.
This is a selected history, not a ranking of importance or a complete account. Read the method note.
02 / AI weather
Four readings taken at different times and places. They describe capability, actual use, sentiment, and institutional response. Our interpretation is dated 20 September 2026; the underlying observations are dated below.
Measured agents handle some longer bounded tasks. Chatbots have broad but uneven U.S. use. Concern has grown, while institutional rules are still taking shape.
METR Time Horizon 1.1 estimates the human duration of tasks an agent can complete with 50% success.
Some measured task horizons have lengthened, but the task suite and intervals bound the comparison. Mainly software, ML, and cybersecurity tasks; human task duration is not agent runtime or job automation.
U.S. adults reporting their own AI chatbot use in a February 2026 Pew survey.
Direct chatbot use is widespread, though far from universal. Self-reported U.S. chatbot use is narrower than all AI use and does not measure global use.
U.S. adult views of increased AI use in daily life, surveyed in June 2026.
Public concern is part of the setting in which tools and rules develop. The percentages do not measure global sentiment or trust in a specific tool; remaining share reflects nonresponse or rounding.
Selected institutional responses from 2019 to 2026.
Institutions have adopted principles, recommendations, frameworks, and staged legal duties. A rule’s existence says little about compliance, enforcement, or effects; these instruments are not equivalent.
These readings are a fixed editorial snapshot, checked 20 September 2026. They are not a live index or a prediction.
03 / Conditional readings
Each connection has a mechanism, an assumption, and a way it might fail. These are questions to investigate, not forecasts.
Signal. ChatGPT made a conversational answer interface public in 2022. In a U.S. teen survey fielded in 2024 and published in 2025, 26% reported using ChatGPT for schoolwork, up from 13% in 2023.
Conditional reading. If students use fluent answers while evidence checking remains difficult, teaching verification may need to become an explicit classroom practice.
What could change this. Tools that expose evidence well, or teaching methods that integrate checking, may make this a richer learning aid.
Follow Conversation becomes a public interface, More U.S. teens report schoolwork use.
Signal. A study of customer support agents at one firm found about a 14% average gain in issues resolved per hour with a generative AI assistant, with larger gains for less experienced workers. ILO estimates occupational exposure, which is not displacement.
Conditional reading. If entry tasks shift to AI, employers may need new ways to teach oversight and judgment. Assistance could also improve feedback for learners.
What could change this. The effect varies by task and workplace. Neither occupational exposure nor one field study predicts job loss.
Follow Assistance changes one workplace, Work exposure is estimated.
Signal. AlphaFold’s 2021 method and public protein structure database made predictions easier for scientists to inspect and use.
Conditional reading. If useful predictions help select experiments, more questions may become practical to ask. The value depends on access and laboratory validation.
What could change this. Error outside validated settings, or bottlenecks in experiments, may limit the benefit. A predicted structure is not an experimental discovery.
Field note / Method
Solid describes a documented event or finding in its source’s scope. Our reading connects events into an editorial interpretation. Could mean is conditional speculation. The timeline tracks and eras are editorial groupings, not measured quantities.
Events were chosen for explanatory range across capability, use, and response. The selection is incomplete and currently favors English-language, U.S., and European sources. Sources and limits appear beside each event and reading; the dated measurements are an editorial snapshot.
Return to Stimarium ↗