Much of the measured change since ChatGPT arrived is real, but the most repeated forecast was not: it did not hold, and the report it was credited to has since removed it.
- A key forecast failed. The forecast of 90% synthetic content, long credited to a 2022 Europol report, did not hold, and Europol has since removed the statement behind it.Europol Innovation Lab, Facing reality?, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.”
- Automated traffic now outweighs human traffic. Measured automated traffic passed half the web in 2024 and reached 53% in 2025.Imperva Bad Bot Report 2026, read at source 9 Sep 2026: “accounting for more than 53% of all web traffic in 2025, up from 51% the year before”.
- Detectors wrongly accuse writers. A 2023 study found a 61.3% false-positive rate on essays by non-native English writers.Liang et al., Patterns 4(7), July 2023, read at source 16 Sep 2026: “average false-positive rate: 61.3%”.
- Deepfake fraud became real. A Hong Kong finance employee authorised transfers after a deepfaked video call.
- EU enforcement began. Article 50 transparency rules under the AI Act became enforceable in 2026.
A dated record of what actually happened, and what was only forecast. Entries marked FAILED are projections that did not hold — kept deliberately, because a timeline that only lists things it got right is not a record.
2022
A forecast of 90% synthetic online content by 2026, credited to Europol
A widely repeated projection, quoted for years as though it were a measurement. It did not happen: by 2026 the machine-written share of new articles had plateaued near half. Kept at the top of this timeline deliberately.Europol Innovation Lab, Facing Reality?, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.” The forecast was repeated for years in Europol’s name; the current (January 2024) version cannot be cited for it.
ChatGPT released
The architecture underneath was five years old. What changed was a text box anyone could type into — which is why this date, not the research, is the starting line for nearly every measurement on this site.
2023
Voice cloning from three seconds of audio
Microsoft Research published VALL-E, producing convincing speech from a three-second recording of a speaker it had never heard, preserving emotion and the acoustics of the sample. The code was never released. The capability arrived anyway.Wang et al., “Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers” (VALL-E), arXiv, submitted 5 Jan 2023, read at source 17 Sep 2026: it “can be used to synthesize high-quality personalized speech with only a 3-second enrolled recording of an unseen speaker as an acoustic prompt”, and “could preserve the speaker's emotion and acoustic environment of the acoustic prompt in synthesis”. The abstract does not address code release; that line is not verified here.
AI detectors shown to falsely accuse non-native English writers
Stanford researchers tested seven detectors on 91 TOEFL essays and found an average false-positive rate of 61.3%, against near-perfect accuracy on essays by US eighth-graders. Schools were already using these tools to accuse students.Liang, Yuksekgonul, Mao, Wu & Zou, Patterns 4(7), July 2023, full text read 17 Sep 2026 via Europe PMC: “High misclassification of TOEFL essays written by non-native English authors as AI generated, with near-perfect accuracy for US eighth-grade essays.”
2024
HK$200 million taken in a deepfaked video call
A finance employee in Hong Kong authorised 15 transfers after a video meeting in which the other participants were synthetic. Hong Kong police received the report on 29 January 2024.Hong Kong Free Press, 5 Feb 2024, read at source 16 Sep 2026: “Police received a report of the incident on January 29, at which point some HK$200 million (US$26 million) had already been lost via 15 transfers.” An earlier version of this entry gave US$25.6 million, said the transfers went to five accounts, and added that the firm was identified in May and confirmed no systems were breached; none of that was checked at source, so it was removed.
Automated traffic passes half the web for the first time
Bots reached 51% of measured web traffic — the first crossing in a decade of monitoring.Imperva Bad Bot Report 2025 (covering 2024). Methodology differs between annual editions. The 2026 edition, read at source 16 Sep 2026, puts 2024 at 51% and 2025 at “more than 53%” — this entry records the first crossing, not the latest figure
Detector reliability tested at scale
The RAID benchmark evaluated twelve detectors against 6.2 million generations across eleven models, eight domains and eleven adversarial attacks. Detectors generalised poorly to unseen models, and simple changes — synonym swaps, a repetition penalty — produced steep accuracy drops. The paper opens by noting commercial detectors advertise 99%+ accuracy.Dugan et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors”, ACL 2024, abstract read at source 17 Sep 2026: “Many commercial and open-source models claim to detect machine-generated text with extremely high accuracy (99% or more).” and “Using RAID, we evaluate the out-of-domain and adversarial robustness of 8 open- and 4 closed-source detectors and find that current detectors are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models.” The abstract says “over 6 million generations”; the 6.2 million is carried from the paper body, not re-read here.
Machine-written articles overtake human-written ones
Analysis of English-language articles via public crawl data found AI-generated articles passing human-written ones in volume — then plateauing near half rather than continuing to climb.Graphite, “More Articles Are Now Created by AI Than Humans”, Oct 2025, read at source 17 Sep 2026: “We find that in November 2024, the quantity of AI-generated articles being published on the web surpassed the quantity of human-written articles.” Its May 2026 successor study, with three detectors, puts the level slightly lower: “since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%”.
Do not treat an AI-detector score as proof. Detectors wrongly flagged non-native English writers and lost accuracy under simple edits, whatever accuracy their sellers advertise.
2025
IEA projects datacentre demand doubling by 2030
Global datacentre electricity demand projected at roughly 945 TWh by 2030 — just under 3% of world electricity.IEA, Energy and AI, April 2025 — a projection, not a measurement; read at source 23 Sep 2026: “projected to reach around 945 TWh by 2030 in the Base Case, representing just under 3% of total global electricity consumption in 2030”.
Automated traffic reaches 53%
The second consecutive year machines outnumbered people online. Human activity fell from 49% to 47%. How much of the automated share was malicious is in the full report, which sits behind a contact form and was not read at source here.Imperva Bad Bot Report 2026 (covering 2025), read at source 9 Sep 2026: “automated traffic continues to outpace human activity online, accounting for more than 53% of all web traffic in 2025, up from 51% the year before”. Re-read 17 Sep 2026. An earlier version of this entry said malicious bots alone were 40% of all traffic; that figure is not in the public post and was removed.
2026
Agents finish fewer than one professional task in four
A benchmark of hours-long tasks drawn from investment banking, consulting and corporate law found frontier models completing less than 25%, and the best agents only 40% even with eight tries.Mercor, Introducing APEX-Agents, published 21 Jan 2026, read at source 17 Sep 2026: “Frontier models successfully complete less than 25% of tasks that would typically take professionals hours.” and “Even with 8 tries, the best agents can only complete 40% of the tasks.” An earlier version of this entry quoted the post as saying the benchmark “contains 480 tasks — 160 each” and that “the best scoring model achieved 24%”; neither sentence is in the post, and both were removed.
The EU begins enforcing the AI Act
Article 50 transparency obligations became applicable and enforceable. The AI Office gained powers to investigate, evaluate models and require corrective measures. Fines reach €15M or 3% of turnover for transparency breaches.European Commission press release IP/26/1714, 31 Jul 2026, read at source 10 Sep 2026: “from August 2, 2026, the European Commission’s AI Office together with national authorities will begin enforcing the Artificial Intelligence (AI) Act, and on the same date, new transparency rules will start to apply”
Full detail on the rules arriving.
The benchmark that saturated
Agent success on OSWorld rose from 12% to 66.3% in a year — then the benchmark saturated, a harder replacement shipped, and the best model scored 20.6%. The fall was the instrument, not the models. Within weeks the new board’s top row read 77.9%, from runs its own maintainers say are not comparable. Anthropic labels that score partial credit: “77.9% partial”, against “41.7% strict”.Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1, read at source 22 Sep 2026: “77.9% partial” and “41.7% strict”.Stanford HAI 2026 AI Index · OSWorld 2.0 leaderboard, Aug 2026, re-read 10 Sep 2026: the board now reads “Claude Fable 5.1 77.9%, Simular Sai 73.0%, GPT-6 Astra 72.6%”, against 20.6% in August
Full figures on the agents page.
Before repeating a figure, check whether it was measured or forecast. The most quoted projection here did not hold, and a benchmark score can fall because the test changed, not the models.