- This topic is empty.
Viewing 1 post (of 1 total)
-
Posts
-
Prompting an artificial intelligence (AI) large language model (LLM) with a question can often produce an answer in minutes that may have taken hours to research. However, these answers can often contain false information that it has fabricated.
This made-up information is known as an AI hallucination and can occur for several reasons.
One potential cause for hallucinations is that the LLM is trained on insufficient data, given the information it is expected to generate for users.
When AI models are trained, they learn to make predictions using that data.
The accuracy of these predictions depends on the amount and completeness of the data used to train the model.
If biased data — limited data, which paints an inaccurate picture — is used, the model will produce a less accurate prediction than if trained on a more complete data set.
A helpful analogy is how people play chess. Professional chess players can compete at the level they do because they have memorised probable outcomes from studying and playing thousands of games.
Given this training, they would know the most effective move for a specific board state — often knowing how the next few moves will play out.
However, someone who has played fewer games would have a less accurate prediction of what to do and may find themselves in checkmate within a few moves.
Therefore, given the finite number of games that can be played, machine learning can quite easily master chess.
This is why computers surpassed humans’ skill levels at playing the board game in the 1990s when Deep Blue beat Garry Kasparov.
However, reasoning is far more complex than chess, as there is infinitely more data on which a model can be trained.
A lack of grounding, or the inability of an AI to understand real-world knowledge and factual information, may also cause a model to generate information that may seem plausible but is factually incorrect.
This occurred in the first demo of Google’s Bard LLM when asked, “What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?”
It answered that the telescope “took the first pictures of a telescope outside of our own solar system.”
Astronomers soon pointed out that the first photo of an exoplanet was taken in 2004, and the James Webb Telescope was only fully assembled in 2011.
Another factor is whether or not the data used to train the LLM is accurate.
When a model is trained on data from a website like Wikipedia or Reddit that may contain inaccurate information, it is likely to hallucinate.
However, inaccurate and limited data are not the only reason.
Large language models do not inherently have a way to fact-check the information they present. When generating text, they produce the next word or phrase they think makes the most sense.
Although AI system developers have released many improvements to increase the accuracy of the outputs produced by their LLMs, hallucinations still frequently occur.
AI hallucinations presented in South African court
A Pietermaritzburg-based law firm, Surendra Singh and Associates, recently made the mistake of not fact-checking answers obtained from ChatGPT.
Judge Elsje-Marie Bezuidenhout interrogated the documents and found that only two of the nine cases cited in the papers existed, adding that the citation for one was incorrect.
She concluded that the firm likely had used AI technology to source the fake legal citations, describing the action as “irresponsible and downright unprofessional”.
The firm was representing KwaZulu-Natal politician and suspended Umvoti Mayor Godfery Mvundla. He claimed the suspension was decided during an unlawful council meeting.
Mvundla secured an interim interdict against the municipality, but Judge Bezuidenhout discharged the interdict and rescinded the order. He then applied for leave to appeal her ruling.
In this application, his legal representation — Ms S Pillay, supported by article clerk Ms R Farouk — cited non-existent case authorities to support their submission.
Bezuidenhout asked Ms Pillay to provide copies of the cited cases, which she said had been provided with the references by an “article clerk” and that she had not seen the cases as she was under a lot of pressure.
It was later revealed that Ms Farouk had drafted the notice of appeal and was ordered to appear before Bezuidenhout.
Farouk said she had obtained the cases from law journals during her research, but when asked which journals she could not respond.
“I asked her (Farouk) if she by any chance used an artificial intelligence application such as ChatGPT to assist with her research, but she denied having done so,” Judge Bezuidenhout wrote in her ruling.
“It then came to light that the cases referenced had been sourced from an artificial intelligence chatbot, namely ChatGPT.”
“It seems to the court that they placed undue faith in the veracity of the legal research generated by artificial intelligence and lazily omitted to verify the research,” she added.
Viewing 1 post (of 1 total)
- You must be logged in to reply to this topic.