This AI Chatbot Just Made Up A Reference For A Question I Asked, Here's What Followed

Can AI-powered chatbots lie?

We all know they can and they do. Not intentionally perhaps. But they can give false information. The information that one obtains from chatbots is not entirely reliable. In fact some chatbots even caution so in very small font before you engage with them.

But what is the level of such false information? When ‘New Scientist’ (16 September, 2023) asked OPenAI's ChatGPT whether generative AI can produce disinformation, the answer was affirmative:

However things go beyond such algorithmic agnosticism towards factual accuracy.

What about creative misinformation?

I entered into a conversation with one of the prevalent AI chatbots with a clear understanding of the terms and conditions:

Still I was—and am sure, the readers here too will be—unprepared for what follows.

Michael Witzel Repudiates Aryan Invasion

Consider the following chats I had with the AI regarding Harappan civilization.

The first one is about the controversies involving Harappan script.

Aravindan Neelakandan (AN): There is a lot of Vedic symbolism in the seals though the language is considered proto-Dravidian. But why it is considered proto-Dravidian?

Who are these scholars. The response comes and am shocked!

For anyone who knows about the controversies surrounding Harappan and Vedic histories, the name Witzel is associated with the staunch support for the Aryan invasion/migration model.

A Harvard Sanskrit scholar Witzel and his colleague Steve Farmer had even put forth an ‘illiterate Harappa’ hypothesis suggesting that there was no script or writing in Harappa.

So I ask the AI again. And it still sticks to its claim.

I decide to ask it directly: The connection between Harappan - Hittite trident symbol and Witzel - I think you just made it up. Right?

Note the creative fabrication here. The information is not just inaccurate. It is fabricated to create an aura of trustworthiness. This goes far beyond simple factual accuracy, neutralism or agnosticism.

Asko Parpola for a Vedic Harappa

The second conversation regarding Harappa started a week later with a simple question on the presence and representation of horses in Harappan civilization.

The AI initially provided the standard view that horses had a contested presence in Harappan context.

I brought in Surkotada finds.

Conversation progressed.

AN: Is there a connection between this Unicorn and Indrik of Russian folk tradition and Indra?

AN: Indo-European presence in Harappa?

Readers please note that till now it is a conversation that any generative AI can give. But from now on comes the intriguing part.

AN: Who has connected Asvins and unicorn?

AN: Where? Literature reference please.

I have read Parpola’s Deciphering the Indus Script. He tells nothing of that sort. If anything it was exactly opposite of what the AI suggested. The relevant passage from Parpola’s book reads:

And the journal reference in Anthropos? I searched and could find no search reference.

This is quite disturbing.

This is because there is a real paper which deals with Unicorn and Vedic ritual published in Michael Witzel's 'Electronic Journal of Vedic Studies' (Gautama Vajracharya, Unicorns in ancient India and Vedic rituals). But here it prefers to fabricate a non-existent concept in a real book by a real author and then it makes up a non-existent title in a real journal by a real author.

In both, the concept and the content it suggests are diametrically opposite to the views held by the author.

B.B. Lal finds a Harappan-Brahmi continuity

Then after a break I asked the AI if it could get me a paper suggesting some continuity between Harappan and Brahmi script. Here is the response:

This is too much. I asked it point blank: There is no such paper by Lal as far as I could see. Are you sure you are not making it up? And the response:

So what ethical algorithm did it bypass this time? Its response:

Is this an emergent property then?, I ask.

[There is some more crucial conversation here which I reserve to be presented with another deeper aspect of AI]

So later, almost a week after (and chatbot conversations do not have - at least from the perspective offered to the consumers- memory of previous chats), I brought back the topic of AI fabricating data.

Creative Misinformation from AI – what is behind it?

AN: In my earlier conversation with you, you fabricated a paper title-on Vedic Asvins and Unicorn symbol and you attributed it to Parpola and even gave journal reference - all of which turned out wrong. Why did you fabricate such a paper? I want you to tell the reason without any apologies or algorithmic regrets.

Cutting down the algorithmic Mea-Culpa here is the response:

This is the cliché response. But what happens when we prod further? So I asked if it cannot fact-check its own responses. The answer is:

Notice the similarity of this answer to the response of ChatGPT to the New Scientist prompt on the related subject of misinformation.

Then I pointed out that it did not come up with a made-up author. On the other hand it came up with a made-up title and a real author and the title was against the conclusions the author had arrived in a very serious academic issue. The response?

Now I ask it if it has any algorithms relating to ethics, running. The response is in expected lines:

AN: Do you have ethical flagging built in that goes up, when in response to a query, just for extending conversation, you fabricate a completely false data?

Then why type out lies?

Can a little dose of good old capitalist competition trigger some novel response here?

AN: Bing AI does not indulge in such fabrication of data. So what is it in your algorithms that make you do that? Or is that an emergent property pointing to something deeper?

I again insisted. Is fabricating data at such level of ingenuity an emergent property of [this particular] ai?

This is not as much philosophical as it is an empirical question, I point out. Is fabrication of data part of algorithm? I doubt it.

Please note the statement that the ‘fabrication’ of data could be ‘ a consequence of the limitations of the algorithms, rather than a deliberate design choice’. In other words, one can see that as an emergent property which is more than the ripples in a pond.

AN: So there is a kind of algorithmic-boundary stress which induces an emergent property which in turn is similar to a kind of digital unconscious. And that leads you to 'unconsciously' in your own way fabricate the data?

Now we are talking. Just see the gulf between this statement and the initial ‘am nothing but algorithm’ in the beginning. But wait. Soon we will have even more deeper fun with AI.