Risk
Risks without hype
You do not need a story about machines taking over the planet to take AI seriously. The harms that already show up in ordinary use are smaller, duller, and easier to miss: a wrong answer that looks finished, a secret pasted into a box, a fake invoice that sounds like a colleague.
Confident errors
Generative systems are built to continue a pattern. That is why they can write a clean paragraph that invents a book, a statute, or a medical dose. The risk is not only that the answer is wrong. The risk is that it is shaped like a right answer. Treat unsupported claims as drafts. Check anything that would move money, medicine, law, or reputation.
This is a property of how the systems are trained to predict, not a personality flaw you can scold away. Some products add retrieval or tools; those can help and can still fail. See models versus products.
Privacy and leftover text
Whatever you type may be stored, reviewed, or used later, depending on the product’s terms and settings. Do not paste passwords, medical records, unpublished legal strategy, or a customer’s private data into a consumer chat unless you have a reason and a policy that allows it. Local software reduces some of that exposure and introduces other ones (you now own the updates and the leaks).
Scams and cheap fakes
The same fluency that drafts a memo can draft a phishing email, a cloned voice, or a face that never sat in front of a camera. The old advice still holds: confirm money moves and account changes through a channel you already trust. “It sounded like them” is no longer enough. This desk is not a threat catalog and will not walk through how to build those fakes.
Over-trust and skipped work
A useful draft can become a substitute for reading the source. Students, analysts, and programmers already see this: the tool is fast, the review is slow, and the review is the job. Organizations that paste model output into customer answers or code without a person who can catch errors are taking a process risk, not a science-fiction risk.
Bias, copyright, and energy — named, not quantified here
Training data reflects the internet and licensed corpora: uneven, outdated, and sometimes unlawful to copy in the way a court may later decide. Outputs can therefore repeat stereotypes or resemble copyrighted work. Training and serving large models also use electricity and water. Those are real categories. This page will not invent a share of global emissions or a court outcome that has not been issued.
Long-term and catastrophic talk
Serious people disagree about whether future systems could cause large-scale harm beyond today’s misuse. That debate is not settled on this site. What we will say is narrower: do not let a movie plot crowd out the harms you can already see, and do not treat a blog’s percentage as a measurement. If you need a method for claims, read how to read an AI claim.