What AI is reading

A robot committed to self-improvement. Generated on Hotpot.ai with the prompt “a robot in a hat reading a book” with the charcoal style, which I hoped would rely more on old, out-of-copyright art.

Readers, you will likely be aware that since the artificially intelligent chatbot ChatGPT was first released in late 2022, it has been pretty extensively tested by users curious to see what an AI can do. A plethora of articles have sprung up (and continue to appear) in various sectors on how you can use ChatGPT to help you in your line of work or course of study. And it seems a lot of people are just using the chatbot to kind of mess around, too, which is fine, because as far as the parent company OpenAI are concerned, messing around with the bot is the whole point; any interaction with the service may be reviewed and used to further train the system.

OpenAI say as much in “What is ChatGPT?”, an article posted in their help pages. Basically an FAQ, the article goes through fifteen commonly asked questions about ChatGPT, from the practical — how much does it cost? — to the philosophical — why does it seem so real and lifelike?

But have a look at question 4: “Can I trust that the AI is telling me the truth?” In answer to that we are told that the bot may sometimes give factually inaccurate answers, because “ChatGPT is not connected to the internet.”

Yes, it’s true. ChatGPT is not online. Hey, that’s the theme of this blog!

Get Jacqueline Barlow’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

What does this mean? Obviously the bot is online. You have to go online to consult it. But when you ask the bot a question, it is not searching the internet for the answer. It is searching a large digital archive of information harvested from the internet, pre-2022. The archive is static and not currently updated. If you think of the bot like that robot in the picture up there (which you definitely should), and the archive like a big giant dusty archival collection on shelves and in boxes (which you also totally should), then the robot is very politely standing at a desk (or “website”) waiting to be asked a question, and when asked, it consults the collection and comes back to you with a reply. But it doesn’t cite its sources, and you cannot go into the stacks yourself.

Unless the robot has memorised the collection, in which case it is like Hap from Storybots. Image capture by juliet_delta on Reddit, here.

Is this how they all work — all the AIs? In short, no, and it is far beyond the scope of this article to list the myriad uses artificial intelligence can be put to and is used for every day. (There’s An AI For That claims to be the largest AI aggregator; go browse if you’re curious!) Let’s narrow it down to the AI chatbots that aim to reproduce natural human conversation while helping you with a multiplicity of tasks. Taking a small selection:

1. Google Bard, according to itself (I asked), draws data from Google Search, Google’s internal knowledge base, and other public datasets (this does chime with what’s in their FAQ, so this probably isn’t an example of an AI going off the rails). The “internal knowledge base” is presumably not online, being internal and all.
2. Claude, developed by Anthropic, cannot access the internet. However, they are working with partners including DuckDuckGo to “integrate Claude with sources of solid, real-time information” (more here).
3. Chatsonic, which openly describes itself as a ChatGPT alternative, claims GPT-4 capabilities, but also offers a Chrome extension delivering “content suggestions from anywhere on the internet.”
4. Perplexity is described as a conversational search engine, drawing solely from the web.

Turns out there are a lot of ways to be a chatty AI. But is there a best way? It is ridiculous to suggest that being connected to the internet would make an AI more accurate. (Just ask Perplexity! It provided a very serviceable, referenced reply to the question “how accurate is online information”, advising careful evaluation of online sources.) The closed online archives accessed by ChatGPT are probably highly curated and structured (though you’d have to take their word for it), so does that make it better than one that is muddied by unfiltered yet current online information? Maybe that depends on the question. And of course, it still makes mistakes. Aside from the funny ones, here is a more rigourous analysis of ChatGPT’s failures: https://arxiv.org/abs/2302.03494. Like any tool, you need to know what it is, how it’s made and what it’s for — and be aware of its limitations — to be able to use it effectively.

Don’t put all your eggs in one basket. There are so many baskets. Photo credit George Chernilevsky, shared under a Creative Commons Attribution-Share Alike 4.0 International license

So if anyone ever asks you, in a pub quiz or something, what ChatGPT has in common with your old family photos from the eighties, now you know.

*

I have not waded into the ethics of artificial intelligence — for example, whether AI image generators like DALL-E and Midjourney are stealing from artists. Recent lawsuits have alleged they do; here is one report on those. Hotpot.ai, which I used to create my header image, gives a lot of credit to a lot of entities, and has clear terms of use, but these don’t dispel the ethical questions.

For more on where ChatGPT’s data is coming from, try ChatGPT — Show me the data sources.

This work is licensed under a Creative Commons Attribution 4.0 International License.