Español

Ban AI From Simulating Specificity

AI cold email's deadliest move is faking specificity.

A line like "I loved your recent post" reads as exactly what it is: a mail-merge token dressed up as research. Recipients have seen ten thousand of them. The fake-personal opener doesn't just fail to help — it actively signals that no human looked at them before hitting send, and it kills your reply rate. When you let a language model write outreach, its instinct is to simulate the appearance of research: warm, specific-sounding, and completely hollow.

Why simulated specificity backfires

A model is a fluent pattern-completer. Ask it to "write a personalized cold email" and it will produce something that has the shape of personalization — an admiring reference, a plausible detail — without any grounding in a real fact about the recipient. It invents the anchor. And invented anchors are worse than none, because they land in the uncanny valley: specific enough to claim familiarity, vague enough to expose the lie. "Your work in the space is impressive" is a tell. So is a confidently wrong detail, which is not just weak but a credibility landmine.

The reader's read is instant: this is automated, and the sender either didn't check or is fine misrepresenting. Either way, delete.

The fix: one real anchor, or it isn't ready

Ban the model from inventing anchors. Make the rule explicit and make it a gate:

Rule: Every email must cite ONE real, verifiable observation about the
recipient — a launch, a hire, a funding round, a public commit pattern,
a talk, a specific article claim.

If no such anchor is provided, DO NOT invent one. Output:
  "NO ANCHOR — draft not ready. Gather one of: [suggested sources]"
and stop. Do not write the email.

Now "no real anchor" produces a flag, not a fabricated compliment. The workflow inverts: instead of the model faking familiarity, it tells you to go gather the anchor first. You (or a research step) supply a verifiable fact, and only then does the draft get written around it.

Power moves

• Feed anchors, don't request them. Put the verified observation in the prompt as data. The model's job is to write around a fact you already trust, never to source it.
• Make the anchor falsifiable. A good anchor names a specific thing (a date, a repo, a claim) that the recipient can confirm is true. Vague praise fails this test on purpose.
• Gate on it in the pipeline. If you're automating outreach, treat "no anchor" as a hard stop that routes to a research task, not a soft warning the model can talk past.
• Prefer one email with a real anchor over a hundred with faked ones. Volume with fake specificity trains recipients to ignore you. One grounded message earns the reply.

Resources

• CAN-SPAM Act compliance guide — U.S. FTC
• OpenAI — Prompt engineering guide
• Anthropic — Reducing hallucinations
• GDPR — official text (EU)