The chatbot replies patiently, remembers details and never seems to be in a hurry. That conversational style can invite an intimacy a user would never offer a form. Yet the words remain inside infrastructure operated by a company. The experience of having a confidant and the conditions under which messages are retained can move in different directions.
“Crimen y algoritmo”, by Rodrigo Castillo Murillo and Francisco Figueroa Ogalde, appeared in El Mostrador on 23 August 2026. The column examines the distinction between expressing an idea and carrying out an act, alongside the privacy problem that arises when a conversation becomes a record. The link leads to the Spanish original. Source: El Mostrador
Context costs more than spotting words
Imagine three conversations containing the same violent description. One concerns a fictional screenplay; another reproduces a threat the user received while seeking help; the third prepares a threat to send to somebody. A detector could find similar vocabulary. Establishing what each exchange means requires understanding its function, its recipients and its relationship to external acts. The example is hypothetical, but it shows why an alert and a conclusion are not equivalent.
A policy that confuses those levels can intervene inappropriately while missing serious situations expressed in ordinary language. Its quality should not be judged merely by the number of messages it flags. What matters is what it distinguishes, the information required to do so and the consequences of mistakes. When an alert triggers action affecting a person, the procedure must be able to examine the leap from the signal detected to the interpretation assigned to it.
Trust is designed too
A provider may state in its terms that conversations are retained. That communicates part of the relationship. At the same time, affectionate replies and permanent availability can encourage a sense of confidential interaction. Examining both together makes it possible to ask which expectations the service creates and how it clarifies them when they actually matter. A notice buried beneath many pages of text may exist without guiding the user’s behaviour.
The alternative is not an absolute promise of secrecy for every use. A responsible policy should precisely explain the conditions for access and intervention, who decides to apply them and what information is disclosed where appropriate. Accumulating data in case it proves useful one day also has consequences: it increases the collection of conversations exposed to improper access, misinterpretation or uses their participants never anticipated.
An explanation must keep track of the decision
When an exchange raises concern, the process needs to distinguish what was observed from what is inferred. A review may require further evidence and may establish that the original signal was unfounded. If only the initial label survives, those receiving the file later could treat it as an established fact. Recording what was rejected matters as much as retaining what triggered the alert.
The issue reaches beyond criminal enforcement. Systems that classify conversations distribute attention, suspicion and opportunities to respond. Their decisions may protect someone or impose an unjustified burden. Making those decisions examinable requires assigning concrete responsibilities and allowing mistakes to be corrected under the rules governing each intervention. A machine’s capacity to listen to many things does not give it the authority to settle what they all mean. Nor can the company using its conclusions disappear behind that capacity.