S’pore not likely to use AI in sentencing in foreseeable future: Chief Justice

The use of AI in sentencing has attracted controversy and remains contentious, Chief Justice Sundaresh Menon said. PHOTO: ST FILE

SINGAPORE - The criminal courts in Singapore are not likely to start using artificial intelligence (AI) tools in sentencing for some time to come.

The use of AI in sentencing has attracted controversy and remains contentious, Chief Justice Sundaresh Menon said on Monday in his keynote address at a two-day sentencing conference held online.

The conference, organised by the State Courts and the Singapore Academy of Law, covers sentencing principles and frameworks, the use of AI in criminal justice and the sentencing of certain classes of offenders, such as youth and those with mental health issues.

In his speech, Chief Justice Menon addressed developments on the horizon, including the possible use of AI in sentencing.

He noted that several states in the US, including Kentucky, Pennsylvania and Wisconsin, have adopted AI risk assessment tools that use demographic data and offence-related data to assess an offender’s risk of recidivism.  The Wisconsin Supreme Court in 2016 dismissed a challenge by an offender against the use of this tool, but stressed that such a risk assessment could not be used to determine the severity of the sentence.

He noted that a study cited in the case found that the tool was far more likely to incorrectly assess black offenders as having a high risk of recidivism than it was to make the same error regarding white offenders.

In Malaysia, a pilot programme using AI to generate sentencing recommendations began in Sabah and Sarawak in 2020, covering the offences of drug possession and rape. Judges have followed the system’s recommendation in a third of cases, said Chief Justice Menon.

The Chief Justice said one criticism of both systems is that the underlying algorithms are opaque, and the offender does not have an opportunity to test the data and assumptions that underlie them, which could reflect biases.

“It remains to be seen if these and other concerns can be satisfactorily addressed. The use of AI in criminal cases requires careful consideration… I do not at this time see us moving in this direction in the criminal justice field within the foreseeable future,” he said.

Chief Justice Menon also addressed the setting up of the Sentencing Advisory Panel (SAP) earlier this year.

The function of the panel – which comprises members of the judiciary, the Law and Home Affairs ministries, the police, the Attorney-General’s Chambers and the criminal bar – is to issue sentencing guidelines, which would be persuasive but not binding on the courts.

The panel’s guidelines, when issued, can be assessed by the public on its website.

Chief Justice Menon recalled that in 2013, the judiciary formed a sentencing council with the goal of achieving greater consistency in sentencing. Since then, the Supreme Court has issued around 100 guideline judgments. But there are limitations to this approach, as an authoritative sentencing guideline can only be established when an appropriate case reaches the Supreme Court, he said.

“The work of the SAP has just begun. But I have no doubt that in the years to come, it will advance our overarching goals of promoting consistency and transparency in sentencing while enhancing public awareness of sentencing, and thus come to play a significant role in our criminal justice system.”

The Chief Justice’s speech also traced the evolution of the courts’ sentencing role and how sentencing discretion is exercised.

Sentencing frameworks serve to secure broad consistency in sentencing outcomes, by establishing patterns of reasoning that can guide the court to a sentence that is both appropriate on the facts of each case and broadly in line with the sentences imposed on similarly placed offenders, he said.

He noted that the courts in Singapore have laid down sentencing frameworks using at least five approaches: the “single starting point” approach, the “multiple starting points” approach, the “benchmark” approach, the “sentencing matrix” approach and the “sentencing bands” approach. 

The choice in approach turns on the features of the offence in question, he said.

For example, the benchmark approach may be viable when an offence typically happens in a certain way. One example is the benchmark of four weeks’ jail for a simple assault of a public transport worker, where a first-time offender pleads guilty.

Such an approach is unlikely to be helpful in cases where the offence tends to manifest in many ways. Thus, in crafting a sentencing framework for the offence of rape, the Court of Appeal eschewed the benchmark approach.

Join ST's WhatsApp Channel and get the latest news and must-reads.