How a screenshot becomes a report
- Recognition.A vision model reads visible speakers, message order, timestamps, text, reactions, and describable stickers from the screenshots.
- User review.The user checks the recognized conversation before analysis. This matters because an OCR error can otherwise become false evidence.
- Redaction and parsing.Recognized phone numbers, email addresses, identity-card numbers, and web links are replaced with neutral placeholders before the conversation is analyzed.
- Deterministic metrics.The application calculates counts and rates from the parsed messages. These calculations do not ask a language model to guess the numbers.
- Constrained interpretation.A language model receives the redacted transcript, calculated metrics, fixed evidence excerpts, tone setting, and explicit safety limits.
- Evidence-linked output.The final report separates the overall interpretation, cited excerpts, observable flags, uncertainty, and actions the user controls.
What the numbers actually mean
| Measure | Definition | What it cannot prove |
|---|---|---|
| Message share | Each participant's recognized messages divided by all recognized messages in the selected sample. | Who cares more, who is “winning,” or why somebody writes more. |
| Average message length | Average non-whitespace character count for each participant. | Emotional depth or sincerity. |
| Question rate | Percentage of a participant's messages containing a visible question mark. | Curiosity expressed without a question mark or the quality of a question. |
| Short-reply rate | Percentage of a participant's messages containing four or fewer non-whitespace characters. | Dismissal. Context can make a short reply warm, practical, or entirely normal. |
| Consecutive bursts | Count of adjacent messages sent by the same participant. | Neediness, anxiety, or any personality trait. |
| Initiation estimate | With full timestamps, the first message and the first message after a gap of at least six hours count as session starts. | Who initiated offline, deleted messages, or conversations without usable dates. |
| Reply-time estimate | Average elapsed minutes when a message follows the other participant and both messages have parseable full timestamps. | The reason for any delay. Time-only screenshots are not treated as complete date evidence. |
| Balance score | A heuristic summary based on the gap in message share and question rate. | Relationship health, compatibility, safety, or future success. |
Progression and avoidance indicators are candidate signals, not relationship facts. The system can notice planning language or extremely brief deflections, but those counts remain internal context. The report must explain the actual behavior in plain language and must not present a keyword match as proof of intent.
Not every signal deserves the same confidence
Directly countable
Speaker, message count, visible question marks, recognized character length, and complete timestamps can be checked against the transcript.
Repeated behavior
Patterns such as one person repeatedly restarting or plans repeatedly staying vague require multiple messages and surrounding context.
Model interpretation
Tone, warmth, avoidance, or the meaning of a sticker can be ambiguous. These observations should be phrased cautiously and tied to visible evidence.
One message is rarely enough for a strong conclusion. Longer, continuous samples generally support better comparisons than selected highlights, but more messages do not eliminate missing offline context or recognition errors.
What the confidence label covers
The current confidence label is primarily a sample-size label: fewer than 35 recognizable messages is lower confidence, 35–79 is medium confidence, and 80 or more is high confidence. It does not certify that the screenshots are complete, that OCR is perfect, or that the interpretation is correct.
The report must not:
- diagnose attachment style, personality, or a mental-health condition;
- claim to know love, cheating, manipulation, or private motives as fact;
- use a sticker or reaction image as decisive evidence by itself;
- turn balanced evidence into a negative verdict merely to sound sharper;
- recommend surveillance, retaliation, coercion, or a reckless breakup;
- replace qualified help in situations involving violence, stalking, self-harm, or coercive control.
Related evidence—and the limits of borrowing it
Relationship research gives useful context for why responsiveness and conversational coordination matter, but it does not validate an individual AI report. Laurenceau, Barrett, and Pietromonaco's diary studies linked self-disclosure, partner disclosure, perceived responsiveness, and experienced intimacy. That supports treating responsiveness as meaningful context; it does not make message volume a measure of intimacy.
Ireland and colleagues found that language-style matching in speed-date transcripts and instant messages predicted later relationship outcomes in their samples. The authors also described the data as correlational. AI Chat Analyzer does not currently calculate their formal language-style-matching measure, and we do not apply their reported odds ratios to individual users.
The NIST Generative AI Profile emphasizes uncertainty, data provenance, human oversight, documented limitations, and risk management. Our practical response is to keep calculated metrics separate from model prose, preserve evidence excerpts, let users review recognized text, and state what the system cannot know.
Sources
- Laurenceau, Barrett & Pietromonaco (1998), “Intimacy as an interpersonal process,” Journal of Personality and Social Psychology.
- Ireland et al. (2011), “Language Style Matching Predicts Relationship Initiation and Stability,” Psychological Science.
- NIST AI 600-1 (2024), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.
How to challenge the method
Methodology should be inspectable. If a definition does not match the product's behavior, a source no longer resolves, or a report appears to make a prohibited claim, email support@aichatanalyzer.com with the relevant page or order reference. Please avoid sending private chat screenshots unless they are necessary and you choose to do so.
Read more about the publisher on the About page, data handling in the Privacy Policy, or start with the relationship text analysis guide.