Models remember more than you think
A model trained on scraped web text can memorize fragments of it verbatim, and under the right prompt it can spit them back out — names, addresses, phone numbers, snippets of private documents. This is privacy leakage, and it has two faces: data that leaked in during training, and data you paste now that may be logged, used for future training, or simply seen by a third-party provider. Both deserve care.
Mitigations exist on both sides. On the training side, techniques like differential privacy and aggressive de-duplication limit how much any single record can be memorized. On the usage side, the discipline is operational: do not paste secrets, health records, or other people's personal data into a tool whose data-handling you have not checked.
Differential privacy bounds how much any single training record can change the model's output — the formal guarantee behind the privacy mitigation.
Whose words, whose pictures?
The same scraping raises a hard, unsettled legal question. Models are trained on copyrighted books, articles, code, and images, generally without the creators' explicit permission, and they can produce output that closely resembles protected work. Copyright and LLMs is now the subject of major lawsuits with no settled answer about what counts as fair use, who owns model output, or whether creators are owed compensation. The terrain is shifting, so for any commercial use, the safe posture is to assume the questions are open rather than to assume you are in the clear.
The cost is real and physical
It is easy to forget that a chat reply is the tip of a very physical iceberg. Training a frontier model consumes enormous electricity and water for cooling, and serving it to millions of users adds a steady running cost on top. This environmental impact is concentrated in data centres whose footprint is large and growing. None of this means you should never use these tools — but it does argue against treating an LLM as a free, weightless oracle to be queried for trivia a calculator or a search box would answer more cheaply.
Log-log plot of loss decreasing as a power law against model scale and compute.
Cheap text, expensive truth
The deepest societal risk is one of scale. These models make it nearly free to generate fluent, tailored, human-sounding text, which means misinformation at scale: spam, propaganda, fake reviews, and persuasive falsehoods can now be produced faster than anyone can fact-check them, and combined with synthetic images and audio they fuel deepfakes that erode our shared sense of what is real. The threat is not one spectacular lie but a rising tide that makes the whole information environment harder to trust.
There is a quieter, personal version of this too: overreliance and automation bias, our tendency to trust a confident machine more than we should and to stop checking. As the output gets more fluent, the temptation to outsource our judgement grows — and skills we stop exercising tend to fade. The risk is not only being deceived by others' fakes, but slowly ceding our own discernment.
Using them well anyway
None of this is a counsel of despair. The same honesty that names these failures points to responsible deployment: match the tool to tasks where its strengths (drafting, transforming, explaining) outweigh its weaknesses, and keep a human in the loop wherever a mistake would actually cost something. Verify specific claims, protect private and copyrighted data, disclose when content is AI-generated, and reserve the heaviest models for jobs that genuinely need them.
The MLOps lifecycle: data to training to deployment to monitoring, with a drift-detection feedback loop back to retraining.