Practical AI guide
July 22, 2026
Reclaim reality: how to check messages with AI
Gaslighting can erode trust not only in one conversation, but in your right to remember at all. AI can help turn a digital trail into a verifiable timeline — if it is never asked to play judge and privacy comes first.
This is not a playbook for proving to a manipulative person that they are wrong, and it is not a way to obtain a diagnosis of someone else from a model.
It is a method for yourself: export a conversation, locate exact language, rebuild a timeline, and verify every result against the original. Dates, quotes, and context can give your perception something solid to stand on.
The core limit is simple: AI does not establish truth. It helps you reach verifiable fragments of a digital trail faster.
Keep the guide
The complete version is in the PDF
13 pages covering the safety protocol, data export, chunking, a master prompt, four ready-to-use requests, and an output verification table.
Section 01
AI does not decide who is right. It returns links to facts
A useful output is not a confident psychological verdict. It is a list you can open and verify: file, author, date, exact quote, and adjacent messages.
What gaslighting changes
Gaslighting is a pattern of manipulation that makes a person question their memory, perception, or sanity. The argument shifts away from the event and toward your reaction.
A message archive cannot tell the whole story of a relationship, but it preserves a digital trail that can be indexed, searched, and arranged by date.
What to search for
- Direct denials such as “I never said that.”
- Language that casts doubt on perception.
- A promise, its change, and its later denial.
- Repeated insults, apologies, and topic shifts.
What not to delegate
- A diagnosis of the other person.
- A legal classification of the archive.
- A final judgment about intent.
- A decision about your safety.
Do not ask AI who is right. Ask for a date, an exact quote, and context.
The limit of the method
A model can miss an episode, confuse speakers, or interpret language too broadly. Its report is a map for manual review, not evidence on its own.
Section 02
Safety first. Model second
A product name — ChatGPT, Claude, Codex, or Claude Code — does not guarantee that the data stays on your device. Identify the real processing boundary first.
Three processing modes
The most private option is a model that actually runs locally with cloud features disabled. A business or API cloud setup can be reasonable only after checking training, retention, deletion, region, and integrations.
A consumer chat should not be the default destination for a complete, identifiable archive. Policies vary by product, plan, region, and settings.
Before upload
- Check whether someone else can access the device, cloud, or backups.
- Work on a copy and store the original separately.
- Replace names, addresses, phone numbers, and identifiers for cloud use.
- Disable unnecessary plugins, web search, and external tools.
Check the service
- Whether inputs and outputs are used for training.
- How long content is retained.
- Who may access it.
- Which features send content to third parties.
A local interface and local processing are not the same thing.
If the device is monitored
If discovering the files could increase danger, do not begin with a technical setup. Start with a personalized safety plan and a trusted person or specialist service.
Section 03
Preserve the original. Analyze a working copy
Reliability begins before the prompt: keep the original export unchanged, anonymize a working copy, and give every fragment a clear place in the sequence.
Export and chunking
In Telegram Desktop, open the chat, choose Export chat history, disable media, and save JSON for machine analysis or HTML for manual reading. In WhatsApp, export the chat without media.
Split a very long TXT into ordered parts. JSON needs a script that preserves valid structure; a blind split can break it.
Minimum folder structure
- original/ — unchanged export.
- working/ — anonymized copy.
- chunks/ — sequential parts.
- index.txt — files and date ranges.
TXT command
split -d -a 3 -l 1000 chat_working.txt chunk_
- -d: numeric suffixes.
- -a 3: three digits.
- -l 1000: one thousand lines per part.
Every report row must point backward: file → message → date → context.
If legal action is possible
Rules for digital evidence and recordings vary by jurisdiction. Preserve the original and ask a qualified local lawyer about procedure.
Section 04
Require a source, not confidence
A narrow question and a strict output format leave less room for flattering, generalizing, or inventing connections.
Master instruction
Tell the agent to work only from the folder, never invent motives or events, and support every statement with an exact quote, date, time, file name, and context.
When evidence is missing, it must say “insufficient data.” Facts and cautious interpretation should appear in separate sections.
Check a phrase pattern
“Find messages from [name] that use ‘you are making this up,’ ‘you imagined it,’ ‘I never said that,’ and close semantic variants. Explain why variants qualify and show the message before and after.”
Check a promise
“Review [date–date]. List direct promises, indirect plan confirmations, and later cancellations or changes separately. If there is no direct promise, say so.”
Apology balance
“Count apologies by each person using the supplied dictionary. Exclude quotes, sarcasm, and non-apology uses. Show counts, percentages, matches, and disputed cases.”
Output schema
- Date and time.
- Exact quote.
- Author and adjacent context.
- File, message ID, confidence level.
A confident tone does not make evidence stronger.
A useful control
Repeat the key search with a simple script or a different keyword set. Agreement between independent methods is more useful than one elegant explanation.
Section 05
Verify the report — and do not turn it into another argument
Even a precise timeline cannot force accountability. The method exists to return footing to you, not to win a presentation in front of a manipulative person.
Verification loop
Open every source, compare the quote word for word, and check author, date, timezone, and neighboring messages. Mark verified fact, disputed interpretation, and model error separately.
Keep the final table apart from raw data. Do not send the full log in the middle of a conflict, and do not make a major decision from an AI output alone.
Do not
- Use the report to force a confession.
- Leave the archive on a shared device.
- Treat message counts as a moral verdict.
- Ignore ambiguous or inconvenient fragments.
Use the result to
- Keep a verified timeline for yourself.
- Talk to a trusted person or therapist.
- Show original material to a lawyer if needed.
- Support a personalized safety plan.
You do not owe proof of your reality to someone who systematically denies it.
When to seek help
If there is a risk of violence, stalking, or device monitoring, contact a local specialist service and make a personalized safety plan. In an emergency, call your local emergency service.
Fact → source → context → verification → human decision
AI is useful here not because it is “more objective than a person,” but because it can index a large archive quickly — if the task is constrained, the data is protected, and every claim points back to the source.
Reclaiming reality does not begin with a confident verdict. It begins when you can say: I can see what happened, I can verify it, and I make the next decision.