Why I don’t think ChatGPT is going to replace search anytime soon

There’s been a lot of breathless coverage of ChatGPT in the past week. One comment I keep seeing on social media is that it’s going to replace Google and other search engines. I don’t think that’s going to happen anytime soon, and here’s why.

Why ChatGPT isn’t about to replace search

1. It's already out of date.

Language models only “know” what they’ve seen in training. In ChatGPT’s case, its training data extends to September 2021. It doesn't know there's a war going on in Ukraine. And with current methods, it wouldn’t be feasible to keep such a model up to date, for example with daily retraining. Frequent retraining may become feasible in the future, but would probably have a massive carbon footprint.

2. It’s unable to cite its sources.

This is mainly because metadata is not part of the training data. So, you might think, why not include metadata in the training corpus by converting it to text and appending it to the associated document? It’s not clear whether that would work, especially with long documents. There’s actually some anecdotal evidence to suggest it wouldn’t work: in a highly publicized example, GitHub Copilot spat out some code it memorized verbatim, and then spat out the wrong license for that code, even though the correct license appeared at the beginning of the same file in the training data.

3. It boils complex, controversial topics down to a single answer.

As a trivial example, I asked it “Should you put eggs in the refrigerator?”, and it answered "Yes", along with a detailed rationale. The answer didn’t even hint at the fact that this is a U.S.-centric view and the rest of the world doesn’t think refrigerating eggs is necessary. This example also shows that minority viewpoints will tend to be silenced in such a system. ("Minority" here meaning less well-represented on the internet.)

What looks more feasible instead

Models like ChatGPT can be integrated with search in a different way: through an IR-QA (information retrieval-question answering) system. In this alternative setup, the system runs a search, then the language model ingests the top result(s) and gives a natural-language answer. Optionally, an extractive question answering model may be used to extract the most relevant passage from the retrieved documents before passing the output to a language model. Separating information retrieval from answer generation in this way allows the system to access fresh, up-to-date information without constant retraining.

This setup addresses points 1 and 2 above, but not point 3. Its output would still be reduced to a single answer, and maybe even a less nuanced answer than you would get from a model like ChatGPT, since the answer would most likely be derived from a single information source. In this kind of system, it would be important to provide a way for users to access the raw search results too, so they have the option to explore the range of possible answers for themselves.

This is not my idea. IR-QA is a very active area of research, and many different ways of implementing it are being developed.

UPDATE (2/22/23): Microsoft’s new AI-enabled Bing uses a version of the approach described above, where search results are fed to a language model.