01
Factual background & dispute
- The offender pleaded guilty to obtaining property by deception and tendered character references from relatives and other members of his community at sentencing.
- A reference from his brother used generic, repetitive language that did not reflect their familial relationship, leading the court to find that its language was consistent with AI generation.
- The author gave no evidence about how the reference was prepared, while counsel had only been instructed that computer translation might have been used.
02
Core issues & judicial focus
- What weight a sentencing court should give a character reference generated or rewritten with a large language model
- How automated translation may affect the linguistic detail and reliability of a character reference
- What enquiries counsel must make before tendering such material
03
Judicial finding & holding
- The court stated that character references used in sentencing should not be generated with the assistance of large language models and warned that automated translation may obscure linguistic nuances relevant to weight.
- Counsel should make appropriate enquiries and be able to tell the court whether a reference was written or rewritten with a large language model or automated translation program.
- The court placed little weight on the brother's reference and sentenced the offender to 21 months and 15 days' imprisonment, wholly suspended, plus a AUD 6,000 fine.
04
Practical risk implications
01The named author should verify any character reference, expert material, or factual statement prepared with generative AI and disclose how the tool was used.
02Lawyers should confirm the generation, rewriting, and translation process with the author and retain a record of that verification.
03Review should identify generic praise, mismatched relationships, and unusual repetition associated with generated text.