A 36-year-old man was arrested in Espírito Santo on suspicion of devising a plan to murder his own 8-year-old son. The objective would be to avoid paying alimony to his ex-partner. The investigations began after OpenAI, creator of ChatGPT, detected user messages detailing the criminal scheme and reported the data to American authorities.
The FBI, in turn, forwarded the information to the Brazilian Ministry of Justice, which forwarded it to the Civil Police of Espírito Santo. The arrest warrant was served on June 19, in the rural area of São Gabriel da Palha, located in the northwest of the state. The Civil Police reported that, in interactions with artificial intelligence, the suspect mentioned his intention to hire a gunman, and made reference to weapons, ropes and poisonous substances. He also reportedly expressed a desire to carry out attacks against institutions such as schools, churches and public bodies.
The individual was subject to preventive arrest and search and seizure, with the police investigation still ongoing.
The episode brings to light several complex legal issues, such as the confidential nature of conversations with artificial intelligence, the possibility of using them as legal evidence and whether the mere planning of a crime can lead to criminal liability.
To shed light on these and other aspects, the portal heard from the specialist in Digital Law, Luiz Augusto Filizzola D’Urso, and the criminal lawyer Maíra Beauchamp Salomi.
Legal assessment on planning a crime and criminal liability
Criminalist Maíra Salomi clarifies that, as a general rule, the simple planning of a criminal act does not generate criminal liability.
It details that the process of a crime, known as *iter criminis*, is divided into four stages: consideration, preparatory acts, execution and consummation of the crime.
The lawyer explains that cogitation, which consists of the mental idealization of the crime, is not of interest to Criminal Law. Next come preparatory actions, such as acquiring tools, analyzing the location or formulating the best way to commit the act. These conducts, as a rule, are also not subject to punishment, except when there is a specific legal provision.
Criminal imputation only occurs when the agent goes beyond this phase and begins to carry out the criminal conduct described in the law.
In the specific case of a homicide, for example, accountability begins at the moment the verb “kill” is executed, going beyond the simple planning or preparation of the victim’s death.
For this reason, the criminalist states that the mere planning of a crime will rarely have relevance in the criminal sphere. She also points out that the Penal Code establishes that the adjustment, determination, instigation and assistance of a crime are only punishable if the crime reaches, at least, the attempt phase.
However, D’Urso highlights that, in this specific case, this will be one of the crucial issues to be examined by the Judiciary.
According to the expert, it will be up to the judge to determine whether the interactions with artificial intelligence remained in the field of preparatory acts or whether they already represented the beginning of the execution of the crime, a scenario in which criminal liability could arise.
The validity of conversations with artificial intelligence as evidence in processes
Maíra Salomi states that conversations with artificial intelligence tools can indeed be used as evidence, as long as the admissibility, authenticity and reliability criteria inherent to digital evidence are met.
The lawyer emphasizes that, like any type of electronic evidence, conversations are susceptible to manipulation or editing, making it essential to preserve the chain of custody.
It will also be necessary to verify the origin of the evidence, the authenticity of the messages, who was actually using the tool at the time of the conversations and whether all the material was obtained lawfully.
If produced and validated under the scrutiny of adversarial proceedings, these interactions have the potential to even support a criminal conviction.
The criminalist, however, emphasizes that a significant challenge remains: proving the user’s true intention when interacting with artificial intelligence in those dialogues.
D’Urso adds that conversations with artificial intelligence can have repercussions that go beyond the criminal sphere.
In the expert’s analysis, even if the content is not sufficient to constitute a crime, it can support other judicial decisions, such as child protection measures, including custody issues, whenever the Judiciary identifies a concrete risk demonstrated by the conversations.
ChatGPT rights to share interactions with authorities
According to D’Urso, OpenAI can indeed share users’ conversations with authorities, as long as such actions comply with the conditions established in the platform’s terms of use.
The lawyer explains that ChatGPT ensures the privacy of conversations before third parties, but also foresees specific circumstances in which OpenAI may have access to this content. These conditions include the improvement of artificial intelligence models and situations that pose a risk to the life or physical integrity of any person.
In these scenarios, he states, the company has the prerogative to forward the information to the competent bodies to safeguard a legal asset considered superior: life.
OpenAI’s obligation to report the case to the authorities
In D’Urso’s assessment, there was no legal obligation for OpenAI to report the case to the authorities.
He explains that, in current Brazilian legislation, there is no rule that imposes on companies developing artificial intelligence the obligation to alert authorities when identifying risks to the life or health of their users or third parties.
This decision, therefore, is up to each company’s internal policy, according to its terms of use.
With regard to OpenAI, the lawyer clarifies that the company chose to include this possibility contractually, a choice that he considers appropriate given the seriousness of the situations that may arise.
Analysis of the impact of the STF thesis on the responsibility of digital platforms in AIs
Another point of discussion generated by the case involves the recent thesis established by the Federal Supreme Court, which expanded the circumstances in which internet application providers can be held responsible for illicit content published by third parties, making the protection previously offered by article 19 of the Marco Civil da Internet more flexible.
However, D’Urso emphasizes that this understanding does not currently apply directly to conversational artificial intelligence, such as ChatGPT.
This is due to the fact that the STF’s decision was drawn up based on platforms that host and disseminate content created by users to other users, a dynamic that differs from generative AI tools.
The expert, however, assesses that this reality could change in the future. As artificial intelligence expands its functionalities and begins to assume roles more similar to those of traditional digital platforms, debates about civil liability on the internet may, in fact, also encompass this type of service.