AI slop needs a safety layer, not a better prompt.
Language models made outreach cheap to generate. They also made a recognisable kind of email cheap: polite, fluent and empty. This is how to detect it before it ships.
What AI slop looks like in outreach.
In cold email, slop is rarely bad grammar. It is fluent text that treats the reader as a slot in a template: praise that could fit anyone, a familiarity that never existed, and an ask far larger than the relationship.
These patterns map onto the boundaries MailDignity checks, which is what makes AI slop detection actionable instead of a vague score.
- Generic flattery: “Loved your recent post, you're clearly a visionary.”
- Fake familiarity: a Re: line or “as we discussed” with no prior thread.
- Asymmetric asks: “Worth 15 minutes tomorrow?” on first contact.
- Guilt-trip follow-ups: silence framed as a debt.
- Unsolicited diagnosis: telling a stranger what is broken in their business.
Why the writer can't grade its own homework.
Asking the same model to review what it just generated tends to approve it. The prompt that produced the draft also shapes the judgment, and the verdict drifts with every prompt change.
A safety layer is independent: fixed boundaries, verbatim evidence for every issue, and a verdict computed from issue severity rather than a mood. See Beyond Prompts for how that works.
Where the outbound guardrail sits.
One API call per draft. PASS emails go out unchanged. REVIEW emails come back with a repaired body that keeps your references and tone. FAIL emails stop and wait for a person.
Lead data → GPT / Claude / agent → MailDignity → Instantly / Smartlead [PASS] send · [REVIEW] repaired_body · [FAIL] human
Block, repair or let through.
| Verdict | What it means | What your pipeline does |
|---|---|---|
PASS | No boundary crossed | Send the original |
REVIEW | Fixable issues | Send repaired_body or ask for approval |
FAIL | Pressure, hostility or deception | Route to a human |
Guardrails for agents, too.
Autonomous agents that write and send email need the same cold email safety layer. MailDignity exposes a single check_email tool over MCP, so ChatGPT, Claude and custom agents can audit a draft before sending — with a cap on re-checks so an agent can't loop.
Try it.
Paste a draft into the AI slop detector, or connect the API to gate every email your stack generates.