By -

In a first for Australia, a Supreme Court judge has addressed the use of artificial intelligence (AI) in character references presented during a sentencing proceeding.

Last week during a sentencing hearing, ACT Supreme Court Justice David Mossop raised concerns about a character reference suspected of being generated by ChatGPT, highlighting potential issues with its reliability and the importance of genuine references. 

The case involved Majad Khan who, along with two co-offenders, arranged to meet with a Facebook Marketplace seller to purchase a large quantity of e-cigarettes for $63,600. 

During the exchange, the offenders loaded the e-cigarettes into a SUV which Khan drove away. His co-offender then handed the seller envelopes filled with pieces of paper made to appear like money. But when the seller noticed the counterfeit cash, he contacted ACT police. 

Khan was captured by CCTV driving the SUV into the carpark of the building where a co-offender lived. Police officers executed a search warrant two days later at the property and found envelopes containing fake paper money. 

Khan pleaded guilty to obtaining property by deception, an offence which carries maximum penalties of up to a $160,000 fine and/or 10 years imprisonment. 

In his sentencing hearing, Khan submitted several character references however, Justice Mossop flagged one supposedly written by his brother as peculiar. 

“I have known Majad both personally and professionally for an extended period, and I am wellacquainted with his unwavering commitment to his faith and community,” the reference read. 

“Majad’s commitment to cleanliness and order is another facet of his character that stands out. He maintains a meticulous approach to his surroundings, expressing a strong aversion to disorder.  

The reference continued to praise “his proactive attitude towards cleaning, both inside the house and in the community, reflects a sense of responsibility and respect for the environment. His efforts extend to keeping the streets and driveways clean, a testament to his commitment to a wellmaintained and orderly community.” 

Mossop flagged in his sentencing judgment that it would be expected that a reference written by his brother would mention this relationship rather than by having known him “personally and professionally for an extended period”. 

“It is certainly possible that something has been lost in translation.” 

“He may well be committed to cleanliness. However, the non-specific repetitive praise within the paragraph which places such an emphasis on his proactive attitude towards cleaning and strong aversion to disorder is strongly suggestive of the involvement of a large language model.” 

Khan’s defence counsel argued that the reference was produced with the assistance of computer translation rather than a large language model. 

However, Mossop stated it was inappropriate for references to be generated with the assistance of large language models as it “becomes difficult for the court to work out what, if any, weight can be placed upon the facts and opinions set out in them.”  

He also addressed the use of computer translation as undesirable as the subtleties of language “will not necessarily be accurately reflected in the automated translation.”  

Mossop concluded that “the use of language within the document is consistent with an artificial intelligence generated document” and handed Khan a suspended sentence of 21 months and 15 days imprisonment, and a fine of $6000.  

Image credit: ACT Courts