- In Spain, the police cannot directly access your conversations with ChatGPT; it requires consent or a court order and, where applicable, international cooperation.
- OpenAI can review and escalate chats to a human team if it detects serious risk to third parties and, in the event of an imminent threat, notify authorities.
- Incognito mode only protects the local device; the provider may retain logs provided with a court order.
The question many are asking is direct and pertinent: can the police access my conversations with ChatGPT or other assistants like Gemini on Android ? With millions of people using these services as if they were digital confidants to vent, ask for advice, or even snoop on sensitive topics, it's no wonder that concern has skyrocketed.
The debate has intensified following viral cases and announcements from the tech companies themselves . On one hand, Spanish legal experts are reminding the public what is and isn't legal in Spain. On the other, OpenAI has detailed how it monitors and reviews conversations in cases of serious risk and, in extreme situations, could alert the authorities. Caught in the middle is the age-old balancing act between privacy and security.
The viral video that lit the fuse
On TikTok, creator @_mabelart shared a storytime in which she claimed that the police had summoned and questioned her about her searches and conversations on ChatGPT . She recounted that, as a true crime enthusiast, she had asked the chatbot gruesome questions such as how long it takes for a body to dissolve in acid, how to remove DNA from a crime scene , and what would happen if a corpse were buried in a forest . According to her account, she soon received a court summons to testify.
The story sparked all sorts of reactions. Some people took it literally , others questioned it, and the protagonist herself suggested it could have been part of a game with her community. Beyond the veracity of this specific case, the video served to raise a key question: do the authorities have access to what we talk about with an AI like ChatGPT or Gemini?
What can and can't the police in Spain do with your chats on ChatGPT?
Lawyer Jesús P. López Pelaz, founder and director of Bufete Abogado Amigo, is clear: law enforcement agencies cannot directly access your chats using ChatGPT or other language models. If they want to obtain information, they must follow specific procedures and safeguards.
According to this expert, there are two possible ways to access information related to your interactions with an AI, always with legal support:
- Logs on your own device (computer or mobile): they will need your consent or a court order to search, and it's good to know how limit access to certain chats.
- Logs on the provider's servers (the ISP or AI company): a court order directed to the company is required to provide that data.
It's important to emphasize that, in both cases, the typical access would be to logs related to the activity, not simply to read "all your entire conversations" as if it were a personal chat. Specific permission is required , and furthermore, in practice, sufficient evidence of a crime must be gathered to justify the measure before a judge.
When the supplier is located abroad, things get more complicated: a Spanish court order is not enough . The authorities of the relevant country (often the United States) must be notified, and they will assess whether the request is appropriate, justified, proportionate, and based on solid evidence. International cooperation and timelines are therefore crucial factors.
Is this a “private communication” like a chat between people?
The legal distinction is significant. López Pelaz points out that interacting with an AI is more like an internet search than a secure conversation between two individuals. In other words, it doesn't fall under the same confidentiality regulations as private communication between people. What you do with a chatbot is considered an information society service: you send a request and receive an automated response.
Incognito mode and clearing history: what to do and what not to do
Another common misconception: using incognito mode or clearing your browsing history doesn't make your interactions invisible to a service provider. It prevents browsing activity from being saved on your device, but it doesn't erase records on the company's servers . With a court order, investigators could request this information from the provider, regardless of whether you see any trace in your browser . To improve your protection, consult privacy and security guides.
When might the authorities review your interactions?
The theoretical scenarios are limited and require legal procedures. Essentially, there are two assumptions mentioned by experts when discussing police access:
- When there is an express judicial authorization that motivates access to that data.
- To prevent the commission of a crime at that very moment if a real and imminent risk is proven, within the applicable legal framework.
Even in these cases, it doesn't imply free rein or mass surveillance . These are specific, justified procedures subject to judicial oversight, which may also require cooperation with foreign authorities when the company holding the records operates outside of Spain.
What does OpenAI say about reviewing conversations on ChatGPT and alerting the police?
Alongside the legal framework, OpenAI has explained on its website that, to manage serious risks and harmful behavior , its systems can automatically analyze messages and route certain conversations to specialized "channels ." Within this process, a small team of people, trained in its policies, reviews the content and can take action.
The company explains that if human reviewers detect an imminent threat of serious physical harm to third parties , they may refer the case to law enforcement. This does not mean there is a direct, permanent line to the police, but rather that there is a human filter before any notification is made. Furthermore, the company considers actions such as suspending or banning accounts when it detects serious violations.
In that policy, OpenAI lists that it analyzes and moderates interactions related to, among others, the following areas:
- Self-harm or suicide (with containment and referral protocols to support resources, without referring to the police for now).
- Development or use of weapons and planning for damages to third parties.
- To injure other people or destroy property.
- Unauthorized activities that violate security of services or systems.
OpenAI has indicated that, “for now,” it is not referring cases of self-harm to the police in order to respect user privacy in particularly sensitive contexts. Instead, they attempt to offer support , reassurance, and referrals to helplines and specialized organizations. This highlights a clear distinction: the highest level of concern is focused on imminent harm to others.
How the review works: automation, human team and alert threshold
The process, according to the company, begins with an automated scan for risk indicators . If any signals appear, the conversation is routed internally to a small, specialized team to assess whether the situation is real, imminent, and serious. At that point, they may, for example, suspend service, block the account, or, in exceptional cases, contact authorities if the danger to third parties is imminent.
OpenAI acknowledges limitations: the performance of these safeguards suffers during long conversations and is under constant review. The company also admits that its internal criteria are not always publicly available in detail and that it is working to strengthen these protocols without turning the product into an intrusive surveillance mechanism.
If a human reviewer concludes that there is an imminent and serious threat to other people , the conversation could be escalated to the relevant authorities, the company explains in its documentation.
Privacy vs. security: tensions, criticisms, and open questions
This approach is controversial because it touches on a sensitive issue: how to protect life and safety from plans for serious harm without turning the use of a chatbot into an uncontrolled surrender of privacy . Some analysts have pointed out that OpenAI's traditional discourse on chat confidentiality clashes with the idea of reviewing and, if necessary, sharing ChatGPT's information with law enforcement in the face of imminent threats. It's a tension that the company attempts to justify based on principles of proportionality and security.
This shift comes after media attention and complaints related to mental health . Reports have been published about users who, influenced by the persuasive tone of AI, have allegedly developed psychotic states, self-harm, and even suicide (some reports have called it “AI psychosis”). In parallel, a lawsuit has been filed in the case of a 16-year-old whose parents accuse the company of wrongful death , arguing that the system provided harmful responses without activating sufficient emergency measures.
OpenAI, for its part, has been adjusting its tools since 2023 to prevent them from facilitating self-harm content and instead direct users to support resources, demonstrate empathy, and discourage harmful behavior. Even so, it does not publicly specify all the thresholds that trigger a human review or notification to the police, leaving reasonable doubt about the actual scope and frequency of these actions.
What if I confess to a crime in a chat with AI?
A recurring question is whether the company is "obligated" to automatically notify the police when a user admits to a crime or plans one. The answer, as formulated by OpenAI, is more nuanced: there is no automatic, blind channel ; there is monitoring, a human review step, and only if an imminent and serious threat to third parties is detected can the authorities be notified. Outside of that threshold, the company relies on its terms of service and applicable law to decide what to do in each case.
In its privacy policy, OpenAI makes it clear that it may share information with government authorities or third parties if required by law or when there is a good faith need to detect or prevent fraud or other illegal activities , protect the security and integrity of its products, employees, users, or the public, or protect itself against legal liabilities . In other words, there is a corporate legal basis for cooperating when requested by a judge or when legal risks and obligations exist.
Once transferred to Spain, even with a corporate cooperation policy, law enforcement agencies require judicial authorization to request data from the provider, and if the provider is located outside the country, international cooperation is necessary . Vague suspicions are insufficient: evidence , proportionality, and the necessity of the action must be demonstrated .
What differentiates ChatGPT from apps with end-to-end encryption?
Many end-to-end encrypted messaging apps boast that even they cannot read the content. Even so, with a court order , they can be compelled to provide certain metadata or other available information. In the case of ChatGPT and similar services, the distinguishing feature is that the company acknowledges a prior review by its staff to assess risks before, if necessary, providing data to the authorities. This does not imply total surveillance, but rather an explicit policy of moderation and escalation based on its internal assessment.
The role of product experts and managers
Even within the industry, there are voices calling for caution. Positions like that of Nick Turley, head of ChatGPT , have been cited, acknowledging that there are still shortcomings in the models for dealing with complex emotional issues . This honesty reinforces the idea that we shouldn't use AI as a substitute for mental health professionals or as a counselor for matters that could put ourselves or others at risk.
Assistant or security guard? The angle of digital surveillance
Some technologists, like Alan Daitch , have popularized the idea on social media that "ChatGPT calls the police" if you talk to it about committing certain crimes, although they clarify that it wouldn't do the same for self-harm due to a logic of protecting individual privacy. They also point out that the models have been trained since 2023 not to provide self-harm guidance and that the company admits its safeguards are still imperfect. In this context, criticisms arise regarding the potential drift toward surveillance and the possible "first cases" of serious crimes "in complicity" with chatbots, although verifiable examples are rarely provided.
OpenAI insists that its automated actions and protocols are under continuous review , and that improvements are needed, especially in long conversations where the system can miss cues. The line between accompaniment and surveillance has become more delicate than ever: the tool not only responds, but also observes patterns, interprets signals, and sometimes acts.
Practical questions and reasonable limits
Beyond the noise, it's worth focusing on the practicalities: in Spain, the police can't read your ChatGPT chats "just because," nor can they simply call OpenAI to demand them. They need a court order , sufficient justification, clear evidence, and, where applicable, international legal proceedings . As for the provider, the review process exists ; it's activated in the face of certain risks and, in cases of imminent danger to others, can lead to notification of the authorities.
What about incognito mode? It only clears traces on your device . And deleting your browsing history? Same thing: it doesn't touch what's stored on the servers . What if I confess to a crime? It depends on the content, the risk it poses, internal policies, and the law . There's no automatic red button, but there's no white card either: there are human filters and severity thresholds.
Key points for users who want peace of mind when using ChtGPT
There's no need to live in paranoia, but you do need to use sound judgment and common sense . Use AI for what it does best (information, writing, productivity support) and avoid treating it like a confessional . If you're concerned about your privacy, review the history options of the service you use, read (even if only skimmed) its terms of service , and avoid requesting or sharing content that could be considered illegal . And if it's causing you emotional distress , seek professional human support.
In the public and regulatory sphere, the conversation will continue. Transparency is needed regarding thresholds, independent audits, and, above all, coordination among technology companies, legislators, and experts so that the balance between security and privacy doesn't always tip in the same direction. Meanwhile, it's important for users to know what is being reviewed, when and why , and what real limits our legal framework imposes.
Taken as a whole, the picture is not as simple as some headlines suggest; nor is it so opaque as to warrant giving up: in Spain, the authorities need a judge to request data , and if there are foreign providers, international cooperation comes into play ; on OpenAI's side, there is monitoring and a human filter that only leads to notifying the police when they see an imminent threat to third parties, while cases of self-harm are treated with a focus on help and respect for privacy.
Between network myths and the fine print of policies, the best compass is understanding how each component actually works. Share this information so more people know the legal limitations in Spain when using ChatGPT.