Íslenskt mál og almenn málfræði - 01.01.2023, Blaðsíða 141

Smith, Neil. 2004. Chomsky. Ideas and Ideals. Cambridge University Press, Cambridge.
Tomasello, Michael. 2008. Origins of Human Communication. MIT Press, Cambridge,
Mass.
Veres, Csaba, og Jennifer Sampson. 2023. Self-supervised learning and the poverty of the
stimulus. Data & Knowledge Engineering 147:102208. <https://doi.org/10.1016/j.datak.
2023.102208>.
Wong, Matteo. 2023. ChatGPT Changed Everything. Now Its Follow-Up Is Here. The
Atlantic. <https://www.theatlantic.com/technology/archive/2023/03/gpt4-release-
rumors-hype-future-iterations/673396/>. [Sótt: 7.10.2023.]
Þórhallur Eyþórsson. 2023. Chomsky og spjallmennið. Morgunblaðið 25. mars, bls. 28.
<https://www.mbl.is/mogginn/bladid/grein/1832361>.
Þórhallur Magnússon. 2023. Að greina gervigreind. Vísir 11. apríl. <https://www.visir.is/
g/20232400636d/ad-greina-gervi-greind>. [Sótt: 23.9.2023.]

Lykilorð: gervigreind, stór mállíkön, ChatGPT, staðarvensl, málhæfni
Keywords: artificial intelligence, large language models, ChatGPT, locative verbs, linguis-
tic competence

summary
‘Artificial Intelligence and linguistic competence’
This squib concerns the apparent success of large language models (LLMs) to grasp ques-
tions posed to them and answer them in a convincing manner. I focus on the recent GPT-
4-powered program ChatGPT and ask whether it can be said to have real linguistic com-
petence, or whether its large data approach is deficient in some essential features of lan-
guage. I argue that ChatGPT lacks a causal model of the real world, and therefore also
lacks both agency in the philosophical sense and a causal understanding of the language it
builds on. With examples from content locative and container locative verbs in Icelandic,
I show that ChatGPT does not grasp the causal functions of these verbs. This suggests that
humans learn language more effectively than current LLMs, and comprehend the full
range of creativity that stems from novel use of sentences in a way that ChatGPT, despite
having been trained on more text than any human, does not. At the same time, as I point
to weaknesses in the linguistic competence of GPT-4, I also recognise that it can come to
serve as a tool in corpus linguistics and perhaps even shed new light on the theory of
Universal Grammar, since LLMs are able to collect, analyse, and find patterns in huge
amounts of data from different languages.

Mikael Lind
Mála- og menningardeild
Hugvísindasviði
Háskóla Íslands
nl@hi.is
Gervigreind og málhæfni 141