💰 Deloitte refunds $290k, admits AI use

Hi, and happy Tuesday.

Deloitte issued a refund of $440k AUS (~ $290k USD).

It was for a report delivered in July to Australia’s Department of Employment and Workplace Relations. 

Chris Rudge, a Sydney University researcher, discovered that the 272-page report included citations to nonexistent academic papers, and misquoted a federal court judgment.

We’ve found these errors - known as "hallucinations" - to be the biggest concern clients have about Generative AI. (Hallucinations are when AI produces confident but incorrect information).

While the AI labs have worked hard to reduce the occurrence of hallucinations, they still happen.

It happened to me as recently as this past Friday.

I was working on a legal issue with GPT-5. 

The AI provided three citations. One of them said the opposite of what the AI claimed it to say.

So how can we catch these hallucinations?

Here is what I did to catch the hallucination I came across:

(1) For each of the three citations the AI provided, I found the original documents it referenced. 

(2) For each document, I created a new chat thread. In that chat, 

1. I uploaded the cited document 

2. I pasted in the conclusion from the original chat thread.

3. I gave the AI the prompt:

With reference to the conclusion pasted above, how does the attached document help or detract from what it says?

(3) In one of these document-specific threads the AI responded:

The attached document decision hurts your position.

(4) From there, I could rework the original chat thread with this information.

Why did this approach to catching the hallucination work?

Every question you send the AI transmits everything in the thread so far.

The longer your chat thread and the more documents and data it contains, the more the AI will struggle to process all the context and provide accurate answers.

The smaller the context provided to the AI, the more accurate it will be.

Why? 

Just as with humans, the smaller the amount of data the “digital brain” (neural network) has to process, the more of its “brainpower” (layers in the network) it can use to process just that portion.

It’s for this reason that I started a fresh new thread for each document with only the information the AI needed to evaluate each one.

What did I learn from this experience?

Generative AI has a last mile problem!

While it can help us quickly gain new knowledge and skills, without the prerequisite training in those fields, it’s a case of the “blind leading the blind”.

And it’s risky to make high-stakes decisions without the “last mile” confidence a seasoned professional provides.

I later spoke to a lawyer about the cases cited by the AI.

He said there had been no rulings on the topic by a higher appellate court, so these cases were not as strong as they seemed.

The AI did not think to educate me on that.

What did Deloitte learn from their experience?

Deloitte has doubled-down on Generative AI, rolling out Anthropic’s Claude AI to its nearly 500,000 employees worldwide. 

Why did they do this?

In the same way rounding errors arise at the limits of calculators, hallucinations arise at the limits of Generative AI.

And even the calculator created press and protest in its early days.

In the 1980s, calculators were the future, just as Generative AI is now. Deloitte recognizes this.

Ready to up-skill on Generative AI?

If you want to go deeper on how Generative AI works, prompting, hallucinations and more - join my complimentary AI Masterclass.

Here’s the invite form: https://docs.google.com/forms/d/e/1FAIpQLSf8ZbGGY1ZdaiUxxCOTTS3CIFmkO9f-oSzU0cfYLjO_cQn7dw/viewform

Talk soon,

Dino