Hallucination v Fugazi

Hallucination v Fugazi. When Large Language Models (LLM) sometimes have a hallucination can we call that a fugazi?

I suggested in my last blog post that fugazi (meaning not genuine, fake, a fraud) was possibly a better term than hallucination for incorrect LLM output.

A chat with Microsoft Copilot about Hallucination v Fugazi

I thought I would ask Microsoft Copilot what it thought:

Me:

is a GenAi hallucination the same thing as a fugazi?

Copilot:

The terms “hallucination” and “fugazi” are not directly related in the context of AI. In AI, a hallucination refers to an instance where an AI system generates information that is incorrect or misleading, often because it has predicted outputs based on patterns learned from its training data1. This can happen across various applications, including natural language processing and image generation.

On the other hand, “fugazi” is a term that originated from the punk rock band Fugazi, which is known for its DIY ethic and anti-corporate stance. It’s not typically used in the context of AI or technology.

If you’re asking whether a hallucination in AI is similar to something being “fugazi,” it would be more accurate to say that an AI hallucination is a false or misleading output from an AI system, while “fugazi” refers to something that is fake or not genuine. They describe different concepts and are not interchangeable.

Me:

But can a GenAI output not be fake or not genuine?

Copilot:

Yes, a GenAI output can be genuine and accurate. The term “GenAI” refers to generative artificial intelligence, which is designed to create new content based on patterns it has learned from data. When properly trained and used, GenAI can produce outputs that are both original and factual.

However, it’s important to note that while GenAI can generate genuine content, it can also produce outputs that are not genuine or accurate if it’s based on incorrect or biased data. This is why the quality of the training data and the design of the AI system are crucial in ensuring that the outputs are reliable and trustworthy.

In summary, a GenAI output can be genuine if it is based on accurate information and the AI system is functioning correctly. It’s always good practice to verify AI-generated content, especially when it’s used for critical applications.

[NB Copilot clearly did not understand (but of course it cannot understand in the true sense of that word anything) my question.]

Me:

That is not what I asked you. Can GenAI output be fake or not genuine?

Copilot:

I apologize for any confusion. Yes, a GenAI output can indeed be fake or not genuine. This typically occurs when the AI generates content that is not based on factual information or when it creates something that is misleading or incorrect. It’s important to critically evaluate AI-generated content to ensure its accuracy and authenticity.

Me:

So that therefore means that a hallucination and a fugazi can be one and the same thing when applied to GenAI?

Copilot:

In the context of GenAI, both “hallucination” and “fugazi” can refer to outputs that are not genuine or accurate. A hallucination in AI is when the system generates content that is false or misleading, while “fugazi” can be used colloquially to describe something that is fake or not genuine. So, in a broad sense, they can be considered similar when applied to GenAI outputs that are incorrect or deceptive. However, it’s important to use these terms carefully and understand their specific meanings in different contexts.

LLM changing Tune on Hallucination v Fugazi

So with a bit of prodding I got Copilot to change its ‘mind’ on hallucination v fugazi and agree that “in the context of GenAI, both “hallucination” and “fugazi” can refer to outputs that are not genuine or accurate”.

With this blog post now being published Copilot will scrape this content and in future probably reference it (whether true or not) as authority for the proposition that “in the context of GenAI, both “hallucination” and “fugazi” can refer to outputs that are not genuine or accurate”.

I have seen this very thing happen before with regard to ‘Inkster’s Law, ChatGPT and Hallucinations‘.

Getting away with the ‘H’ word

As Allen Woods has commented, a number of times, on LinkedIn:

Altman is a salesman. And a very good one. However, he should never have been allowed to get away with the “h” word.

You see recently OpenAI reportedly admitted that “hallucination” could not be corrected because the root cause could not be identified. The curious thing though is that be definition they can’t say for definite why it gets things right either…..

And that is ridiculous given the core advantage of computing is accurate calculation at volume and velocity……

Deception

We are being deceived if we are led to believe ChatGPT output might be genuine. So more Wizard of Oz (fugazi) than Dumbo and Pink Elephants on Parade (hallucination) in my book.

Other terms for Hallucination or Fugazi

There have been other terms used rather than hallucination or fugazi. So it is not just hallucination v fugazi.

Usefully Wrong

Microsoft when they introduced Copilot to the world referred to the fact that it would hallucinate as being “usefully wrong“:

When you brainstorm with a person, their throwing-spaghetti-at-the-wall suggestions aren’t usually The Thing. They’re ideas, points of view—puzzle pieces that may fit together or may not. They stimulate your creative juices, helping you get to The Thing. The same is true of AI: Ask Copilot for 20 suggestions for a catchy presentation title, and bam! You have a list of relevant suggestions in seconds. They won’t all be that just-right idea you’re looking for, but often one or two will spark real inspiration—leading you toward a title you’d never come up with on your own.

At Microsoft, we like to say that this is when AI is “usefully wrong.” It’s a different, more collaborative way of working with technology—one that upends our long-held assumptions about what computers can and can’t do. We’re used to one interaction with a computer: put in a query, get an answer. With AI, the answer isn’t the final word; the magic is in the conversation, the back and forth. And it’s up to people to build on, combine, or transform the content into something original and meaningful.

Not sure that a client of a solicitor will ever be happy with an output by that solicitor that is “usefully wrong”! So the solicitor requires to check and verify all GenAI output which might take them twice as long as it would otherwise have done. And solicitors don’t usually use the “throwing-spaghetti-at-the-wall” technique when applying their skills to a task at hand.

Gibberish

The Judge in the infamous ChatGPT lawyer case (Mata v Avianca) suggested what ChatGPT produced was “legal gibberish”.

As I said at the time:

Possibly a better and more realistic term than legal hallucinations!

Big Fat Lies

Alex Cranz, deputy editor and co-host of The Vergecast said:

AI might be cool, but it’s also a big fat liar, and we should probably be talking about that more.

ChatGPT is, after all, sometimes referred to as a Black Box Pinocchio.

But, maybe not an intentional liar like Pinocchio. Alex Cranz expanded:

The AI keeps screwing up because these computers are stupid. Extraordinary in their abilities and astonishing in their dimwittedness. I cannot get excited about the next turn in the AI revolution because that turn is into a place where computers cannot consistently maintain accuracy about even minor things.

If only the Legal Futurists took some cognisance of this fact.

Fabrications or Blagging

Jamie Wodetzki, Chief Product Officer at Catylex, co-founder of Exari (acquired by Coupa), said on LinkedIn:

I think we should stop calling them hallucinations and start calling them fabrications. The former implies that these mistakes jump out like obvious drug-induced nonsense. In reality they are deceptively believable and almost impossible to spot. Fabrications rings more true of this LLM tendency. Or blagging might be a better verb?

Bullshit

A more straightforward approach is to simply call it bullshitting.

Michael Townsen Hicks, James Humphries & Joe Slater take that view in their paper ‘ChatGPT is Bullshit‘:

Recently, there has been considerable interest in large language models: machine learning systems which produce human-like text and dialogue. Applications of these systems have been plagued by persistent inaccuracies in their output; these are often called “AI hallucinations”. We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs. We distinguish two ways in which the models can be said to be bullshitters, and argue that they clearly meet at least one of these definitions. We further argue that describing AI misrepresentations as bullshit is both a more useful and more accurate way of predicting and discussing the behaviour of these systems.

Any more words?

We started with hallucination v fungazi. I’m not convinced by “usefully wrong”. That is perhaps a fugazi. We also have gibberish, big fat lies, fabrications, blagging and bullshit. Anymore to add to that list?

Image credits: Dumbo © Disney; Wizard of Oz © MGM Studios; Black Box Pinocchio © The Time Blawg (prompted) and GenAI generated