Generative language models are currently the most prominent application in applied computational linguistics, and they raise an old question: what does it take to understand language? A language model is trained on huge corpora to predict the next token given the previous ones. At every step, it computes a probability distribution and samples from it, and repeated many times, this produces fluent text. What the model learns is distributional semantics, the idea that meaning can be approximated by the contexts in which forms occur. The stochastic parrot argument holds that this yields form without meaning, since the model has no grounding in perception, no causal model of the world, and no communicative intention. The opposing view points to emergent abilities, consistent semantic representations, transfer to new domains, and pragmatic competence in dialogue, and calls this functional understanding. The disagreement depends on what is meant by understanding, and it revisits the whole stack of levels from below: does the model command syntax, semantics and pragmatics, or only imitate their surface?
We organise the ethical discussion around four ideas: Training data raises questions of intellectual property, since large corpora contain copyrighted work without explicit permission or attribution. Energy use raises questions of environmental cost, both from training and from ongoing inference. Unequal, largely English-centric and Western-centric training data raises questions of bias, which a model’s dialogic nature can then reinforce over the course of a conversation. The annotation and moderation work behind alignment raises questions of labour, since low-paid workers around the world classify emotional or harmful content so that a system can appear cooperative. A related weakness is sycophancy, the tendency to confirm a user’s views even when they are false, which shows how a system optimised for approval can reinforce error.
Finally, the question of whether machine-written text can be detected returns to the same stack from the outside. Perplexity-based detectors look at how predictable a text is, linguistic analyses look for uniform vocabulary and sentence length, and stylometry looks for the loss of individual idiolect. Each works at a different level of granularity, and each is unreliable, especially for short texts, which shows how closely generated language now resembles human writing.