No doubt some of the excitement about OpenAI’s chatGPT may have reached the group. It’s an open beta and you’re welcome to test it here: https://chat.openai.com/chat

I played around with it conducting the sort of information requests a researcher may want to do on secondary data. i.e. Assume you have a KM system and a user is looking to conduct a literature review and summarise the results (a fairly common task).

chatGPT and other adversarial networks work (very, very summarised almost to the point of uselessness) by taking a corpus of work and breaking it down into a set of weights and measures. The adversarial part is where those weights and measures are reassembled and evaluated by a classifier to see if it reaches a specified level of coherence. So, in response to a prompt it should return text that is relatively indistinguishable from an existing document of a similar nature.

Build up this process over billions of documents and you get a system that can accept any natural language question and return a natural language response. This should be factually accurate, but note that the classifier does not explicitly test for this. It only tests for readability and internal coherence.

Here’s an attempt at something relatively simple:

However, a Google search for France’s GDP growth rate in 2000 returns this:

The statement about raising the retirement age from 60 to 62 is true, but the events described are in 2010, and there is no Economist article on this https://www.bbc.com/news/10326002. Similarly, the ecotax https://www.france24.com/en/20131116-french-truck-drivers-block-highways-protest-ecotax ... from 2013.

I’ve tested this on a live project I’m working on (200 documents to summarise around a set of due diligence questions) and … entirely made-up references.

You can think of this as cargo cult research. It has the form and authoritative voice of a research summary, but is no different from cutting out words from different reports and sticking them together to make new reports.

This shouldn’t be a surprise as it works a little like image generators (Dall-E, Stable Diffusion) which assemble diverse images from random components. That means it understands the form of a factual response, but is unable to produce the actual facts and references directly since they don’t exist in that form but as a network of weights and measures.

I'm not sure throwing more processing power, or better-optimised algorithms at it is the answer. You need an entirely different approach to summarise and interpret factual information, starting with a classifier that refuses to accept references that aren’t real.

We are now in a period of great risk for KM as these systems will see general use, and will generate realistic-sounding reports which take tremendous time and expertise to evaluate and refute. As it is, parts of the tech community are already banning the use of chatGPT since it produces such low quality, but realistic-sounding, responses (e.g. https://meta.stackoverflow.com/questions/421831/temporary-policy-chatgpt-is-banned).

I’ll play around with it a little more, but I think it’s best use – and killer app right now – is to fix autocorrect when typing on my mobile.

>--------------------<

Gavin Chait is a data scientist and development economist at Whythawk.

uk.linkedin.com/in/gavinchait | twitter.com/GavinChait | gavinchait.com