Short, sourced answers to the questions asked most often about AI content, detectors, made-up answers, agents, the new rules, energy and this site.
- The internet is not mostly AI. Machine-written articles have levelled off near half of new ones, and the automated-traffic share counts every kind of bot, including search crawlers.
- Detectors cannot support an accusation. In one study they wrongly flagged most essays by non-native English writers.
- AI makes things up because guessing was rewarded. Scoring that gives nothing for “I do not know” pushes models to guess.
- Agents partly work, and their scores are unstable. The best complete fewer than one professional task in four, and a benchmark score fell when the benchmark changed, not the models.
- Set boundaries before an agent acts. The common failures are not stopping, wrong scope and silent failure, and content it reads can redirect it.
The questions that arrive most often, answered with the source and the date attached so you can check every one. If a figure here is wrong, it gets corrected in place.
Is most of the internet really AI now?
No — and the two figures people mix up are measuring different things. Machine-written articles overtook human-written ones in volume in late 2024, then plateaued near half rather than continuing to climb. Separately, automated traffic reached 53% in 2025 — but that counts bots of every kind, including search crawlers and monitoring.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”. Methodology differs between editions
The widely repeated "90% by 2026" was credited for years to a 2022 Europol report, and Europol has since removed the statement behind it. It did not happen.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 timeline keeps it at the top deliberately.
Can AI detectors tell if something was written by AI?
Not reliably enough to accuse anyone. Seven detectors tested on 91 TOEFL essays by non-native English writers averaged a 61.3% false-positive rate, against near-perfect accuracy on essays by US eighth-graders.Liang, Yuksekgonul, Mao, Wu & Zou, Patterns 4(7):100779, 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.”
A detector that flags six in ten non-native writers is a bias amplifier, not a detector. Full detail on how to spot AI writing.
Why does AI make things up?
Because it was scored for guessing. Benchmarks award full credit for a lucky guess and zero for saying "I do not know." Under that scoring, a 60%-confident guess scores 0.6 and an honest abstention scores 0.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, OpenAI, September 2025 (arXiv 2509.04664), read at source 11 Sep 2026: “language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty”
It is not malfunctioning. It is doing what it was rewarded for. Full mechanism here.
Are AI agents actually working yet?
Partly, and the headline figures are unstable. The best agents complete fewer than one professional task in four, and only 40% even with eight tries.Mercor, Introducing APEX-Agents, 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 quoted two other sentences as coming from this post; neither is in it.
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
Am I legally required to label AI-generated content?
In the EU: for text, only where it has not been through human review or editorial control. Article 50(4) of the AI Act became enforceable on 2 August 2026. It requires disclosure of deepfake images, audio and video, with no review exemption, and labelling of AI-generated text published to inform the public on matters of public interest — unless the text has been through human review or editorial control and someone holds editorial responsibility for publishing it.Regulation (EU) 2024/1689, Art. 50(4), read at source 17 Sep 2026: “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.” For text, the duty “shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.” · Commission Guidelines on transparency obligations, adopted 20 Jul 2026, read at source 11 Sep 2026 — the obligations began to apply on 2 August 2026, and outputs generated before that date need not be marked retroactively.
Fines reach €15M or 3% of turnover for transparency breaches.EU AI Act, Article 99, read at source 23 Sep 2026: breaches of the “transparency obligations for providers and deployers pursuant to Article 50” are subject “to administrative fines of up to EUR 15 000 000 or, if the offender is an undertaking, up to 3 % of its total worldwide annual turnover for the preceding financial year, whichever is higher”. More on the rules arriving.
How much energy does one AI query use?
There is no single agreed figure. Published estimates vary with the model, the hardware and who is measuring, and a figure for an AI query is not directly comparable with one for a conventional web search.
A confident single figure is a sign the speaker has not read the range. More on energy and water.
Does prompt engineering still matter?
Less than it did. Frontier models handle a decent prompt fine. What moves the needle now is context — the documents, history, tools and memory surrounding the request.
The techniques tied to one model's quirks expire at the next release. Context engineering covers what does not.
Is it safe to let an AI agent act on my computer?
Only within a boundary you set explicitly. The most common failures are not stopping, wrong scope, and silent failure — reporting success having done nothing.
Agents can also meet instructions inside content they fetch, and cannot reliably tell those from yours. A related, measured risk is poisoned memory rather than one fetched page: in one 2026 study, poisoning a single dimension of an agent's persistent state raised average attack success from 24.6% to between 64% and 74%.Wang et al., "Your Agent, Their Asset", arXiv 2604.04759, read at source 9 Sep 2026 and abstract re-read 17 Sep 2026: “poisoning any single CIK dimension increases the average attack success rate from 24.6% to 64-74%”. The study measured poisoning of the agent’s stored state, not instructions arriving in fetched content. See prompting agents and guardrails.
Who writes this site?
An anonymous editor who works in generative AI and uses these tools on parts of this site, and says so on every page where it applies.
The site is deliberately anonymous. Every figure carries the organisation that issued it and the date, so it can be checked without knowing who collated it. Errors are corrected in place on the page concerned, with a note where the change matters, rather than edited silently.
How do I cite something from this site?
Cite the original source, not this page. Every figure here names the organisation that issued it and the date it was published — those are the citations worth carrying.
Reference this site only if you are citing the collation itself: the comparison, the framing, or a correction made here. Contact details are on the press page.