AI is lying to you

I’ve resisted writing about the current trends in AI because, well, others are already doing it better.1 But I was inspired by Garrett‘s observation that – according to the Washington Post – the C4 dataset has tokenised his personal website.

Screenshot from the Washington Post, showing a search for 'danq.me' returning 300k tokens (0.0002% of the dataset). The site is ranked 61,724.
My website is in the dataset too, but with a massive 300,000 tokens. Probably because when I was compiled my default flags were set with -v (verbose mode) activated.

Much has been said about how ChatGPT and her friends will hallucinate and mislead. Let’s take an example.

Remember that ChatGPT has almost-certainly read basically everything I’ve ever written online – it might well be better-informed about me better than you are – as you read this:

Dan asks: "Tell me about the blogger who runs DanQ.me" ChatGPT responds: "Dan Q is the blogger and web developer who runs the website "DanQ.me". He is a British computer scientist, game designer, and open-source software developer. DanQ.me is a personal website where Dan writes about a variety of topics, including technology, programming, and gaming. He is also known for creating the web-based puzzle game "Notpron," which has gained a large following online. In addition to his work on "DanQ.me" and "Notpron," Dan has contributed to several open-source software projects, including the Linux operating system and the Mozilla web browser. He has also worked as a software developer for several companies, including Google and Canonical. Dan is active on social media, particularly Twitter, where he frequently shares his thoughts on technology and other topics." The screenshot is marked up with the bits about gaming and being active on social media marked "stretching/exaggerating", the bits about Notpron, Linux, Mozilla, and Canonical being "believable lies", and the bits about Twitter and Google identified as "complete bollocks"
Given that ChatGPT has all the information it needs to talk about me accurately, it comes up with a surprising amount of crap.

When I asked ChatGPT about me, it came up with a mixture of truths and believable lies2, along with a smattering of complete bollocks.

In another example, ChatGPT hallucinates this extra detail specifically because the conversation was foreshadowed by its previous mistake. At this point, it digs its heels in and commits to its claim, like the stubborn guy in the corner of the pub who doubles-down on his bullshit.

If you were to ask at the outset who wrote Notpron, ChatGPT would have gotten it right, but because it already mis-spoke, it’s now trapped itself in a lie, incapable of reconsidering what it said previously as having been anything but the truth:

Asked to elaborate on "Notpron", ChatGPT says - Dan Q's custom "Notpron" challenge can be found at the following URL: https://www.danq.me/notpron/. This challenge features unique levels and puzzles that were created by Dan Q himself, and it is known for being particularly difficult and challenging. It has gained a large following among "Notpron" enthusiasts and puzzle game fans. If you decide to take on the challenge, be prepared to think outside the box and use your problem-solving skills to their fullest!"
Notpron is great and all, but it was written by David Münnich, not me. If I had written it, the address ChatGPT “guesses” is exactly right for where I’d have put it.

Simon Willison says that we should call this behaviour “lying”. In response to this, several people told him that the “lying” excessively anthropomorphises these chatbots, implying that they’re deliberately attempting to mislead their users. Simon retorts:

I completely agree that anthropomorphism is bad: these models are fancy matrix arithmetic, not entities with intent and opinions.

But in this case, I think the visceral clarity of being able to say “ChatGPT will lie to you” is a worthwhile trade.

I agree with Simon. ChatGPT and systems like it are putting accessible AI into the hands of the masses, and that means that the people who are using it don’t necessarily understand – nor desire to learn – the statistical mechanisms that actually underpin the AI‘s “decisions” about how to respond.

Trying to explain how and why their new toy will get things horribly wrong is hard, and it takes a critical eye, time, and practice to begin to discover how to use these tools effectively and safely.3 It’s simpler just to say “Here’s a tool; by the way, it’s a really convincing liar and you can’t trust it even a little.”

Giving people tools that will lie to them. What an interesting time to be alive!

Footnotes

1 I’m tempted to blog about my experience of using Stable Diffusion and GPT-3 as assistants while DMing my regular Dungeons & Dragons game, but haven’t worked out exactly what I’m saying yet.

2 That ChatGPT lies won’t be a surprise to anybody who’s used the system nor anybody who understands the fundamentals of how it works, but as AIs get integrated into more and more things, we’re going to need to teach a level of technical literacy about what that means, just like we do should about, say, Wikipedia.

3 For many of the tasks people talk about outsourcing to LLMs, it’s the case that it would take less effort for a human to learn how to do the task that it would for them to learn how to supervise an AI performing the task! That’s not to say they’re useless: just that (for now at least) you should only trust them to do something that you could do yourself and you’re therefore able to critically assess how well the machine did it.

Screenshot from the Washington Post, showing a search for 'danq.me' returning 300k tokens (0.0002% of the dataset). The site is ranked 61,724.× Dan asks: "Tell me about the blogger who runs DanQ.me" ChatGPT responds: "Dan Q is the blogger and web developer who runs the website "DanQ.me". He is a British computer scientist, game designer, and open-source software developer. DanQ.me is a personal website where Dan writes about a variety of topics, including technology, programming, and gaming. He is also known for creating the web-based puzzle game "Notpron," which has gained a large following online. In addition to his work on "DanQ.me" and "Notpron," Dan has contributed to several open-source software projects, including the Linux operating system and the Mozilla web browser. He has also worked as a software developer for several companies, including Google and Canonical. Dan is active on social media, particularly Twitter, where he frequently shares his thoughts on technology and other topics." The screenshot is marked up with the bits about gaming and being active on social media marked "stretching/exaggerating", the bits about Notpron, Linux, Mozilla, and Canonical being "believable lies", and the bits about Twitter and Google identified as "complete bollocks"× Asked to elaborate on "Notpron", ChatGPT says - Dan Q's custom "Notpron" challenge can be found at the following URL: https://www.danq.me/notpron/. This challenge features unique levels and puzzles that were created by Dan Q himself, and it is known for being particularly difficult and challenging. It has gained a large following among "Notpron" enthusiasts and puzzle game fans. If you decide to take on the challenge, be prepared to think outside the box and use your problem-solving skills to their fullest!"×

0 comments

    Reply here

    Your email address will not be published. Required fields are marked *

    Reply on your own site

    Reply elsewhere

    You can reply to this post on Mastodon (@blog@danq.me), Facebook, LinkedIn.

    Reply by email

    I'd love to hear what you think. Send an email to b21227@danq.me; be sure to let me know if you're happy for your comment to appear on the Web!