Understanding AI
What it does, and what it doesn't.
A plain account of how today's systems roughly work, what they are genuinely useful for, and where they fail — with, for each limit, something you can actually do about it.
Roughly how it works
Prediction, at a scale that surprised everyone.
Documented
Today's most visible AI systems — the ones behind chat assistants, coding tools and image generators — are built by training very large statistical models on very large collections of data, so that they can produce a plausible continuation: the next words of a text, the next pixels of an image, the next step of a task.
That description sounds reductive, and it is. What surprised nearly everyone, including the people building them, is how much apparently sophisticated behaviour falls out of doing this at sufficient scale: explanation, translation, code, argument, style.
Open question
Whether that constitutes understanding is not settled, and the disagreement is not merely semantic — serious researchers hold quite different positions. This site will not pretend the question is closed in either direction.
Our reading
Two consequences follow from the mechanism, and they matter more than most feature announcements. First, confident errors are structural rather than incidental: a system optimised to produce plausible output will sometimes produce output that is plausible and wrong, and it has no separate faculty that notices. Second — and this one is routinely overstated in both directions — we cannot simply read a model's reasoning off the page.
When a system explains its answer, that explanation is itself generated text. It may correspond to the process that produced the answer; it may not. Interpretability research is making real progress on looking inside these models, but "the system told us why" and "we observed why" are different claims, and only one of them is currently cheap.
What it is genuinely good for
Useful, unevenly.
The gap between the demonstrations and the daily reality is wide in both directions — some promised capabilities disappoint, and some unglamorous ones quietly save people hours every week. The italic line under each use is this project's own assessment, not a measured result.
Learning and explaining
Rephrasing something until it clicks, working through a problem step by step, getting a first map of an unfamiliar field.
Widely used. Reliable for orientation, unreliable as a final authority — it is a study partner, not a textbook.
Writing and creating
Drafting, restructuring, changing register, getting past a blank page, generating variations to react against.
Strongest where a human judges the output. The taste still has to come from somewhere.
Searching and summarising
Condensing long documents, comparing sources, pulling out what matters from a pile of text.
Genuinely useful, and genuinely risky: summaries drop things, and you cannot see what was dropped.
Communicating across languages
Translation and rephrasing that let people read and be read outside their own language.
One of the clearest documented benefits — and one where errors are hardest for the reader to catch.
Accessibility
Description, captioning, reading and writing support, simplification for different needs and abilities.
Often overlooked, frequently life-changing for the people who rely on it.
Lightening repetitive work
The formatting, the sorting, the boilerplate, the tenth version of the same message.
Where the time actually goes. Less impressive than the demos, more valuable than them.
Where it fails
Limits and risks, with what actually helps.
Listing risks is easy and slightly useless on its own. Each of these comes with something a person or an organisation can actually do — not a solution, but a direction. None of them eliminates the risk.
Confident errors
These systems produce plausible text, and plausible is not the same as true. Errors arrive fluent, well-structured and without hesitation — which is exactly what makes them hard to catch.
Treat output as a draft to verify, not an answer to trust. Ask for sources and check them. Be most suspicious when it sounds most certain.
Privacy and confidentiality
What you type may be transmitted, stored, and in some configurations used to improve a service. Policies differ sharply between providers and between plans.
Read the actual policy for the specific product and tier you use. Assume anything pasted may persist somewhere. Prefer local or contractually restricted options for sensitive material.
Dependence and skill loss
A capability you always delegate is a capability you slowly stop having. This is not automatic — it depends on how the tool is used — but it is a real pattern worth watching in yourself.
Keep doing the thing sometimes without help. Use assistance to go further, not to skip the part where you learn.
Persuasion and manipulation
Fluent, personalised, tireless text is a powerful instrument. It can inform; it can also nudge, flatter, or wear someone down — at a scale and cost that were previously impossible.
Be wary of systems designed to maximise engagement or attachment. Ask who benefits from the conversation continuing.
Concentration of power
Training and serving frontier systems requires capital, hardware and expertise available to relatively few organisations. That concentration shapes what gets built and who gets to object.
This is the strongest practical argument for keeping open and independent alternatives alive — see the ecosystem.
Work and livelihoods
Some tasks are being automated, some jobs are being reshaped, some are being created. Which effect dominates, and for whom, is genuinely unsettled and varies enormously by sector.
Beware of confident predictions in either direction. Watch what is actually happening in a specific trade rather than in general.
Social effects
Information ecosystems, synthetic media, relationships with systems that simulate warmth, and what it does to public trust when anything can be fabricated convincingly.
Slow, collective questions with no technical fix. They belong in the horizons as much as in engineering.
Environmental cost
Training and running these systems consumes energy, water and hardware. The scale of the impact depends heavily on the model, the infrastructure and the energy mix, and published figures vary widely.
Ask for specifics rather than accepting either dismissal or alarm. Efficiency has improved substantially; total usage has also grown.
On sources. The claims on this page are deliberately kept general, because the specific ones — how much energy a given model uses, how many jobs a technology displaced, how a particular provider handles your data — change quickly and vary by provider, version and country.
Rather than cite figures that would be stale or wrong by the time you read them, this page points you at the right question to ask. For anything that matters to a decision you are making, check the primary source: the provider's own current documentation and terms, or the published research itself. Where this project could not verify something, it has been left out rather than approximated.