Neil Sands
Chalk up a win for the humans, partially at least.
A mock court hearing set up to examine the role of artificial intelligence in criminal trials has raised serious questions about the reliability of AI-derived evidence.
New Zealand Police organised the exercise at the High Court at Auckland on Thursday to examine the issues created when AI analysis plays an important role in building the prosecution’s case.
While it involved a hypothetical scenario, the hearing involved a real judge – High Court Justice Greg Blanchard KC – real lawyers and a real expert witness.
The fictional case involved an accused drug dealer, Phillip Popper, who was identified as a suspect and became the subject of a surveillance operation based solely on analysis of online posts by an AI tool called “TraceLex”.
In the scenario, police gathered evidence through surveillance that strongly linked Popper to the crime.
But he was challenging the admissibility of the evidence because the warrant allowing it to be collected was issued on the strength of analysis by TraceLex – a tool which doesn’t currently exist, although something similar is likely to become reality very soon.
Infinite time, infinite intelligence
Defence barrister Jacob Parry argued that the analysis could not be relied on because the AI tool had not been peer reviewed and it was a “black box”, meaning police had no access to the algorithms it used to generate the analysis because its owners deemed them commercially sensitive.
He also pointed out that the tool was cloud-based, meaning its owners could change its workings at any time without police knowledge, resulting in a totally different analysis than the one used when the warrant was sought.
Prosecutor Helen Brown argued that police need access to new technology to investigate criminal offending and protect the community, saying the AI tool performed the work of a human “if they had infinite time and infinite intelligence”.
“Should we allow the use of an AI tool like TraceLex to help the police to do things that are beyond the scope of ordinary humans? I say the answer to that question has to be yes,” she said. “As technology develops, so must we. That’s always been the case. Artificial intelligence is just another way that the police can keep up with our increasingly digital world.”
Under the scenario being presented, Brown said the AI tool was not performing a generative function that could give rise to hallucinations; it was instead reviewing vast quantities of text to provide an analysis.
Supporting evidence needed
Blanchard declined to provide an oral judgment but did outline some initial thoughts, focusing on the reliability of the AL analysis which was used as the basis for issuing the warrant.
“Case law relating to anonymous informants might be helpful in this context,” he said. “There is a similarity in that it can be difficult to assess the reliability of anonymous informants. The Court of Appeal has said that there has to be some accompanying evidence in the application to suggest why the information should be considered reliable and why the informant’s assertions are solidly grounded.”
He said that on balance he was likely to find that the warrant was not validly obtained but he would admit the evidence because it was obtained in good faith and had proved reliable.
A show of hands from audience members in the court overwhelmingly backed the premise that it was unreasonable to issue a search warrant on the basis of the AI analysis and the evidence should not be admitted.
So the end result revealed a great deal of scepticism about the reliability of AI, although that may change as the technology evolves at breakneck pace.
Or in the words of Arnold Schwarzenegger’s implacable visitor from the future: “I’ll be back…”
View the livestream of the event here.
0 Comments