What happened
A criminal case in São Paulo state centred on a video recording. The question was whether an offensive word could be heard in the audio.
According to Brazil’s Superior Court of Justice (STJ), the official forensic institute, the Instituto de Criminalística, examined the recording and found no signs of speech matching the word mentioned in the charge.
According to the STJ, a police investigator then produced a second report using generative artificial intelligence tools, which the court named as Gemini and Perplexity. That report reached the opposite conclusion and said the word had been spoken.
According to Conjur, the trial court accepted the charge. The São Paulo Court of Justice found no evidence that the report had been produced unlawfully and allowed it to stay in the case file.
On 7 April 2026 the STJ’s Fifth Panel held unanimously that “the report produced by a police investigator using generative artificial intelligence tools does not have the minimum epistemic reliability to be used as evidence in criminal proceedings”. It ordered the report removed from the file and directed the trial judge to decide again, without it, whether the charge should proceed. The court announced the decision on 8 April.
What an auditable version would have shown
The STJ’s account sets out what the forensic institute did: an analysis of the audio by the established methods for examining speech. It does not say what the investigator asked the tools, which files were supplied or what answers came back.
According to the court, the tools “do not process sound waves” and were not suitable for phonetic analysis. The legal news site Migalhas reported that the defence argued the report could not be checked or reproduced independently, and that the reporting judge treated the absence of those qualities as an obstacle to relying on it.
A record of the analysis would have shown which tool was used, what material was supplied to it, what it was asked and what it returned. It would also have shown whether anyone compared those answers with the forensic institute’s finding before the report was used to support the charge.
The full judgment was not retrieved for this entry. The sources reviewed do not show who decided that the tools should be used.
Where the gap was
In this entry’s reading, the gap was that nobody established whether the tools could do the task at all before their output was used.
According to the STJ, the reporting judge said that “one of the risks inherent in the use of generative artificial intelligence is hallucination”, which is the presentation of information that is inaccurate, unreal or fabricated. The court found that the report lacked “minimum epistemic reliability” to be admitted as evidence.
According to the STJ’s summary of the decision, the court also recorded a degree of confirmation bias in the state’s conduct, in that the official forensic finding was set aside in favour of the AI analysis.
A VerificationGate is designed to route a factual claim to a trusted source, rather than accepting the model’s answer, before the claim is used.
A ConductRecord keeps the tool, the input, the prompt and the output of each analysis, so that it can be reproduced and challenged.
VerificationGate and ConductRecord are Headlights designs.
What governance should have looked like
The STJ left the decision on the charge itself to the trial judge, and has not decided whether the word was said.
Before an AI tool is used on evidence, best practice would be to check whether it is capable of doing the job, and to have a qualified expert carry out the analysis by methods that other experts can check.
If an AI report contradicts the official forensic findings, the difference should be recorded and referred back to the forensic institute. The defence should be given the details of the AI analysis, including the material supplied to the tools, the questions asked and the answers produced.
Failure Pattern: a police investigator used generative AI tools to re-analyse audio after the official forensic institute found nothing, and the resulting report, which contradicted the forensic finding, was used to support a criminal charge and accepted by two lower courts.
Governance Principle: an investigating authority should be able to show, for any analysis it puts forward as evidence, what tool produced it, whether that tool is suited to the task, and how its result can be reproduced and checked.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Informativo de Jurisprudência no. 884, HC 1.059.475-SP (Superior Tribunal de Justiça, 14 April 2026)
- Relatório feito por IA generativa não pode ser usado como prova em ação penal (Superior Tribunal de Justiça, 8 April 2026)
- Entenda a Decisão traz tese sobre uso de relatório policial produzido por IA generativa em ação penal (Superior Tribunal de Justiça, 7 July 2026)
- STJ julga validade de denúncia com laudo feito por IA generativa (Conjur, 11 February 2026)
- STJ invalida relatório gerado por IA como prova em ação penal (Migalhas, April 2026)
- O STJ e os limites da IA generativa no processo penal (Conjur, 8 May 2026)