Conversation analyzer: what it must read to say anything
A conversation analyzer works from written turns, senders and timestamps. What Portraitor replaces before any of it is sent, where the key to the real names stays, and what no amount of reading can reach.
A conversation analyzer reads written turns: who sent each message, when it arrived, and what it said. Everything it reports is worked out from those three fields. Portraitor adds one step before any of it is read, and that step is what decides how much of your conversation ever leaves the room.
Three fields, nothing more, and a masking step in between. That is the whole machine. Knowing which three fields it reads tells you what it can answer, and knowing where the key to the real names is kept tells you what it can never hand to anyone else.
A conversation is turns, timestamps and who sent them
An exported chat is a list of lines. Each line carries a sender, a time and some text. A conversation analyzer has nothing else to work from, so every number it produces, how often one person starts things, how long replies take, which words recur, is derived from those three fields and from the order they come in.
That is why the two questions a tool asks first are usually which person and which stretch of time. Portraitor’s own page asks exactly those, in that order. Its second step says to pick the one person you want a portrait of, and tells you that their messages are read while everyone else stays context. Its third step offers a preset span or a range you drag. The unit of analysis is therefore one participant’s turns inside one window, not a relationship and not a verdict on a thread.
The file has to be the kind the reader can parse, which is a separate question from how long it is. Which chat export files can be analyzed goes through the formats app by app, and the two controls that narrow a chat covers what picking a shorter window does and does not change.
What the masking step removes before anything is read
Portraitor’s privacy policy, version 1.3 of 30 September 2026, puts a step in front of the reading. Before any part of the conversation is transmitted, the browser removes names, email addresses, telephone numbers, postal addresses and credentials and replaces them with anonymous placeholders. The policy says this takes place entirely on your own device, and that if it does not complete successfully no data is transmitted and the payment is refunded automatically.
The same paragraph then states its own limit, which is the part worth reading twice. In the policy’s words, this removes who is speaking; it does not remove what was discussed. If the people in the conversation talked about their beliefs, their relationships or their bodies, those passages travel to the analysis provider inside the anonymised text, with no name attached to them. A masked conversation is not an empty one.
Attachments are handled differently again. The policy says chat exports often contain photographs, videos and voice recordings, that the service never opens them and never transmits them, and that only the written messages are read, on your device, with only their masked text sent. What the masking step does to names and what happens to photos in a chat export each take one of those halves further.
The placeholder map stays on your device, so the names cannot come back
Replacing a name with a placeholder is only useful if the list that joins the two is kept somewhere the text never goes. Portraitor’s policy says that for the other people in the conversation, who have not used the service themselves, their names, email addresses and telephone numbers are removed on your device by the same process, and that the list matching each placeholder back to a real detail never leaves your device. Its conclusion from that is flat: neither Portraitor nor the analysis provider learns who they are.
A later section says the same thing from the other direction, listing what the service keeps on your device and nowhere else. One item is a snapshot of the placeholder list, held so that an interrupted portrait can resume after the page is refreshed. The policy says none of those items is transmitted.
That is a claim you can hold a competitor to, because it is architectural rather than a promise about conduct. A service that receives your raw export holds both the text and the identities, and can only offer to delete them. A service whose re-identification key exists in one browser cannot restore the names later even if it wants to, because the key was never on its side of the wire. The honest remainder is the subject matter: the policy also says the provider does not use the conversation to train its models, but does retain submitted content for a limited period in order to monitor misuse of its own service, and that Portraitor is unable to disable that retention. Whether a chat is used to train AI sets out that chain in full.
Four things a conversation analyzer cannot reach
The first is anything that was never written down. Tone of voice, a pause on a call, a look across a room: none of it is in the file, so none of it is in the reading. The second is the media inside the export, for the reason above.
The third is whatever the messaging app itself leaves out when it builds the export. WhatsApp’s own help article on exporting chat history says that view once messages will not be exported, and in its WhatsApp Web version says that view once and disappearing messages will not be exported either. The export is a copy of what survived, not a copy of the conversation.
The fourth is the identity of the other participants, once masking has run. That is a limit by design rather than an accident, and it is the reason a portrait is written about one participant you chose rather than about the group. What a chat analysis cannot tell you works through the rest of that list.
What it will not tell you about the other person
Portraitor’s privacy policy lists what its analysis infers: personality traits, attachment and conflict style, emotional regulation, motivations, communication style, patterns of concern such as manipulation or coercive control, and darker personality tendencies, each with an estimate of confidence. It also says the analysis estimates the language of the conversation and whether messages appear to have been written with the help of an AI tool. Those are the policy’s own categories, and each arrives with a confidence estimate attached rather than as a finding.
The terms page, version 1.3 of the same date, draws the line around what any of that may be used for. Its acceptable use section says you agree not to use a portrait, or anything derived from one, to make or support a decision about a person’s employment, education, housing, credit, insurance, immigration status, or access to any service or benefit, and states that the service is built for personal reflection and is not designed, tested or suitable for any such purpose. The same section says not to use the service to surveil, stalk, harass or harm anyone, and that before payment you confirm you have permission to read the conversation. What you cannot use a chat analysis for quotes those clauses in full.
So the answer to what a conversation analyzer can say about someone is narrow and worth stating plainly: it can describe how a person came across in writing, to you, in the span you picked, with a confidence figure beside each description. It cannot tell you what is true about them. If that narrower thing is what you wanted, Portraitor’s own conversation analyzer runs the masking step in your browser first, and what comes back in the report shows the shape of the result before you spend anything.
Frequently asked questions
Frequently asked questions
What does a conversation analyzer actually read?
The written turns in an exported chat: the sender of each message, its timestamp and its text. Every statistic and every description it produces is derived from those three fields and their order.
Does a conversation analyzer see my name?
With Portraitor it does not. Its privacy policy, version 1.3, says names, email addresses, telephone numbers, postal addresses and credentials are replaced with anonymous placeholders in your browser before anything is transmitted, and that the list matching a placeholder back to a real detail never leaves your device. Other services differ, and the answer is in each one’s own policy.
Can a conversation analyzer read a phone call or a screenshot?
Not with Portraitor. Its policy says only the written messages are read, and that photographs, videos and voice recordings inside an export are never opened and never transmitted. A call leaves no written turns at all, so there is nothing in the file to read.
Can a conversation analyzer tell me what is wrong with someone?
No, and Portraitor’s terms say its portraits must not be used to make or support decisions about a person’s employment, education, housing, credit, insurance, immigration status or access to a service. What the policy describes is a set of inferences about communication, each carrying a confidence estimate.
Three fields, one masking step, and a key that stays in your own browser.
One person, one span, masked first
The names are replaced in your browser before anything is sent, and the list that joins a placeholder to a real detail stays on your device.
Read a conversation with Portraitor →- Portraitor. Privacy Policy, version 1.3, 30 September 2026, sections 2, 6 and 7.
- Portraitor. Terms of Service, version 1.3, 30 September 2026, sections 2, 3 and 5.
- Portraitor. Home page, the four step flow, the person step and the time period step.
- WhatsApp. How to export your chat history, iPhone version.
- WhatsApp. How to export your chat history, WhatsApp Web version.