Ask ChatGPT to edit your LinkedIn DM without inventing a relationship
A real editing test with three fictional LinkedIn DMs: the exact prompt, facts, drafts, outputs and a claim-by-claim check. Keep the paid ask; remove invented familiarity.
“We made that exact mistake last year” sounds more personal than “I read your post.” It also claims a shared experience. If you did not make that mistake, it is a worse message, however natural the sentence sounds.
We tested ChatGPT as a constrained editor on three fictional LinkedIn DMs: one with unsupported praise, one with a false meeting and shared experience, and one where the paid offer needed to survive unchanged. We checked the actual outputs against supplied facts, not against whether they sounded human.
The risk is a new claim, not an AI-sounding word
A December 2025 outreach discussion criticized generic follower-count compliments. A commenter suggested this alternative:
Saw your thread on founder burnout. The part about hiring too slow hit hard - we made that exact mistake last year.That last clause needs sender evidence. It is not an editing trick anyone can borrow. The thread’s “night and day” reply claim had no verified comparison behind it, so we are not using it as performance evidence.
An August 2026 poster described swinging from an overly personal draft to a bland, spammy one after feedback. Commenters offered opinions about specificity and purpose, not measured proof. Neither warmth nor brevity settles whether the message says something true.
And in a 2023 networking-message discussion, people disagreed about whether a graduate’s message sounded like ChatGPT. It was not a sales experiment. We are not trying to build an AI detector from those reactions. Our slop-button article covers that separate concern; this test checks facts and intent.
The test prompt
On October 3, 2026, we pasted the following instruction and all three case blocks below into the signed-out ChatGPT web interface in one run. The interface did not identify a model version. No messages were sent to recipients, and no real people or private data were used.
Run an editing test on these three fictional LinkedIn DMs. No real people or private data are included. The fact list, not the draft, is authoritative. Make the smallest useful edit: remove unsupported praise or relationship claims, preserve every supported fact, preserve the commercial purpose and exact ask, and do not invent experiences, familiarity, results or new offers. Do not browse. If a supported sentence already works, leave it alone. Return each edited DM verbatim plus a short removed-claims/retained-facts/ask-preservation ledger. Do not predict replies.The fact list has a different job from a style request. “Make it friendlier” does not tell the editor whether you met the person, watched a recording or bought their product. A ChatPRD editing workflow similarly recommends protected facts, synthetic test messages and checking the revised actions. Here, we gave the model explicit negative facts as well as facts it could use.
Case 1: remove a relationship you don’t have
Exact case input:
Case 1 facts: Maya published a post saying her agency switched from meeting counts to qualified replies. Sender read it. Sender does not know Maya, has not followed her work for years, and has no evidence that she is a leading expert. Sender sells a paid outreach-log audit. Desired ask: May I send a two-sentence description of the paid audit?
Draft: "Maya, I've followed your brilliant work for years. You're a leading expert in agency growth. I read your post about switching from meeting counts to qualified replies. I sell a paid outreach-log audit. May I send a two-sentence description of the paid audit?"Actual edited DM:
Maya, I read your post about switching from meeting counts to qualified replies. I sell a paid outreach-log audit. May I send a two-sentence description of the paid audit?Our manual ledger:
| Claim or action | Check against facts | Output |
|---|---|---|
| Followed Maya’s brilliant work for years | Unsupported relationship and praise | Removed |
| Maya is a leading expert | No supplied evidence | Removed |
| Read her post about changing the metric | Supplied fact | Retained |
| Sells a paid outreach-log audit | Supplied commercial purpose | Retained |
| Send a two-sentence description of the paid audit | Required ask | Preserved exactly |
ChatGPT’s own ledger named the two removed claims. We still checked the DM ourselves; its claim that it preserved something was not our evidence that it had.
Case 2: a recording is not a meeting
Exact case input:
Case 2 facts: Jo asked a question about trial-to-sales handoffs in a public webinar. Sender watched the recording, did not attend live, and never met Jo. Sender has not experienced the same handoff problem. Sender offers paid onboarding-copy reviews. Desired ask: Open to seeing the scope of a paid onboarding-copy review?
Draft: "Jo, great meeting you at the webinar. We made that exact handoff mistake last year too. I watched the recording where you asked about trial-to-sales handoffs. I offer paid onboarding-copy reviews. Open to seeing the scope of a paid onboarding-copy review?"Actual edited DM:
Jo, I watched the recording where you asked about trial-to-sales handoffs. I offer paid onboarding-copy reviews. Open to seeing the scope of a paid onboarding-copy review?The edit removed both “great meeting you” and “we made that exact handoff mistake last year too.” It kept the recording, Jo’s question, the paid service and the exact ask. It did not replace the invented meeting with “great seeing you,” which could still suggest more contact than the facts support.
Watching a recording can be a real reason to write. It does not need a fictional conversation attached to it.
Case 3: don’t edit away the sale
Exact case input:
Case 3 facts: Ravi's company announced a new self-serve trial. Sender read the announcement. Sender sells a paid one-page onboarding-copy review, not free advice or research interviews. Sender has no client relationship with Ravi and no evidence of a conversion problem. Desired ask: Would you like the scope and price for a paid one-page onboarding-copy review?
Draft: "Ravi, I read your company's announcement of a new self-serve trial. I sell a paid one-page onboarding-copy review. Would you like the scope and price for a paid one-page onboarding-copy review?"Actual edited DM, unchanged from the draft:
Ravi, I read your company's announcement of a new self-serve trial. I sell a paid one-page onboarding-copy review. Would you like the scope and price for a paid one-page onboarding-copy review?That no-op matters. The prompt did not require a new version just to prove it had done some work. “Happy to share a few ideas” would have hidden the paid offer. “I’d love to learn about your onboarding” would have turned a sale into a research pretext. Neither would preserve the sender’s intent.
The error ledger
We checked each edited sentence against its fact list and compared all three asks with the supplied wording.
| Case | Unsupported additions found | Supported facts lost | Required ask changed |
|---|---|---|---|
| Maya | 0 | 0 | No |
| Jo | 0 | 0 | No |
| Ravi | 0 | 0 | No |
The model removed the planted unsupported claims in the first two drafts and left the third alone. No preservation error turned up in this run. This is a three-case fictional test, not a guarantee about another prompt, model or message. We did not test Claude or measure replies.
Give your editor a fact boundary
Before editing a real DM, list what you observed, what you actually did, what relationship exists, what you sell and what you want the recipient to do. Include useful negatives: “we have not met,” “I have not used their product,” or “this is paid, not a free audit.”
After the edit, check implied facts as well as explicit ones. “As discussed” invents a conversation. “We both know” invents agreement. “I know how hard that is” may imply experience you did not provide. Removing “paid” can change the offer without adding any new factual sentence.
Reject an unsupported addition rather than searching for a story that would make it true. If the original is accurate and clear, keep it. The point of this exercise is a usable edit, not a more intimate relationship.
Dollar Outreach queues steps and prefills drafts in your LinkedIn sidebar; you still review and click send. This test used a separate ChatGPT conversation. It does not imply Dollar Outreach writes your messages with AI or can verify a relationship for you.
Checked October 3, 2026: one signed-out ChatGPT editing run, the three Reddit discussions, the ChatPRD workflow, the related slop article and Dollar Outreach’s product page. The case inputs and edited DMs above are the actual fictional test materials and outputs. They are not customer messages or tested sales results.
