Suppose you need to identify what might cause a change in customer purchases before estimating the effect of an intervention. A large language model (LLM) can help surface causal information to investigate, but its fluent answer is not itself a causal assumption you have established. section 13.1 section 10.1.2
An LLM is a deep neural language model trained on vast amounts of text and other data, with parameters numbering in the billions or more. It generates coherent, contextually relevant language and structured text such as code or database queries. It is also a foundation model: a base for more specialized models and applications. Some LLMs work with images as well as text. section 1.4.3 section 13.0
In a causal workflow, keep these roles separate:
- Use an LLM to source or check proposed information about the data-generating process. Such information can inform the assumptions needed to identify a causal query. section 13.1
- Distinguish statistical patterns from causal assumptions. A predictive neural network can model an association; intervention questions require additional causal assumptions. section 10.1.2
- Treat the response as something to assess, not a fixed result: generation is random, and model and interface updates can change answers. section 13.1
A tool-assisted starting point
GPT-4 through ChatGPT is used to query causal information for an analysis. The practical next step is to decide which proposed claims about the data-generating process are defensible before relying on them for identification. section 13.1
Using LLMs to answer causal questions and using causal methods to improve their design and training are active lines of investigation. Neither direction makes causal conclusions assumption-free. section 1.4.3 section 10.1.2
