You can spot them now. A message arrives, warm and well structured, three short paragraphs, a soft ask at the end. The first line compliments a talk you are fairly sure you never gave. Somebody fed a model your name and let it fill in the gaps, and the model did what models do when they run out of facts.
Most people don't reply. They aren't offended either, which is the part worth paying attention to. The message goes in the same mental bin as a parcel delivery text from a number nobody recognises.
We wanted to know how common that reflex had become, so we asked. In The Network Paradox, our survey of 500 US professionals, 66% said they're less likely to respond to a message that reads as AI generated. Only 10% said it makes them more likely to reply. Two thirds of the working population has developed an allergy to AI generated outreach in about two years, and most of the people sending it have no idea.
Here's the uncomfortable part. The messages people are ignoring aren't badly written. They're well written. That's the problem.
Good outreach used to be rare because it was expensive. Finding out that someone had just raised a round, or moved cities, or written a thing you actually read, took twenty minutes a person. The effort was the signal. When a stranger knew something real about you, it meant they'd spent something to find it out.
AI took the cost of sounding like you'd done the work down to nearly zero, and it did it for everyone at once. So the signal stopped working. Not because the messages got worse, but because the thing they used to prove is no longer proved by them.
That's why "but mine are personalised" misses it. Personalisation was never the point. Evidence of attention was the point, and personalisation used to be the cheapest available proof of it. It isn't any more.
The response rate finding was the one we expected. What we didn't expect was what's happening on the other side of the feed.
Nearly a third of the professionals we surveyed, 32%, said they've posted less about their own work because of how much AI written content is online now. Only 9% said they've posted more. People aren't just ignoring AI generated outreach in their inbox. They're pulling back in public, because they'd rather say nothing than be mistaken for a machine.
Think about what that costs over a few years. The people with the most useful things to say are the ones most alert to sounding generic, so they're the first to withdraw. The volume stays high and the signal drops. Then everyone's harder to find, harder to read, harder to approach honestly.
Alongside that, 59% agreed that LinkedIn rewards looking successful more than being useful. One respondent put it more bluntly than we would have: "Those people are there to inflate numbers, or to give off an appearance of success." That's a lot of people describing the main professional network as a performance they no longer believe.
When we asked who they'd actually turn to for something real, a job lead, an introduction, a favour, 67% named their close circle first. Just 9% said they'd post publicly and see who responded.
That's the whole shift in one pair of numbers. Trust didn't disappear. It contracted. It moved from the open channel, where anyone can now sound like anyone, into the small set of people whose messages arrive with a history attached.
Which means the question for anyone trying to reach someone has changed. It used to be how do I write something good enough to get a reply. Now it's closer to how do I get into the category of person this message is from, rather than the category it looks like.
Three things, and none of them is a template.
The first is specificity a model couldn't invent. Not "I loved your post on hiring" but "you said the thing about hiring your opposite and I've been arguing with it for a week". A model can produce a plausible compliment. It can't produce the detail you only have because you were in the room, on the call, or genuinely reading. If your opener could be true of fifty people, it will be read as true of none.
The second is an offer that costs you something. Most outreach is an ask dressed as an introduction. If you open by giving, an intro, a name, a piece of information someone actually needs, you've done the one thing automation still can't do cheaply, which is spend your own social capital on someone else's behalf.
The third is timing that proves you were paying attention. Reaching out the week something happened reads completely differently from reaching out whenever a sequence surfaced the name. This is where most people lose, and not for lack of care. They care, they mean to follow up, and then nine days pass and the moment is gone. We've written separately about how to follow up without feeling awkward, and almost all of the awkwardness comes from having left it too long to reference anything real.
If you're asking someone to make a connection on your behalf, the same rules get stricter, because now you're spending their credibility as well as your own. That's worth its own thought, and how to ask for a warm introduction covers the version that doesn't put the person in the middle in a difficult spot.
Here's what we keep coming back to. Almost nobody's outreach problem is a writing problem. Hardly anyone sits down unable to find the words. What they can't find is the material.
You met someone at a conference in March. You know you talked about something that mattered. You know there was a reason to stay in touch. Eight months later you have a name, a vague warm feeling, and nothing you could put in a first line. So you reach for a tool that will generate something plausible, and you send the kind of message two thirds of people now delete.
The fix isn't better generation. It's better recall. If you can remember what someone said, how you met, what you owe them and what they're working on, the message writes itself in about ninety seconds and reads like a person, because it is one. If you can't, no model is going to rescue it. It will just help you sound more fluently like everyone else.
That's the gap we built Goodword to close, and we'd think it was worth closing even if we weren't the ones doing it.
AI generated outreach isn't failing because the technology is bad. It's failing because it automated the proof rather than the thing being proved, and people worked that out faster than anyone expected.
Which leaves a strange and fairly hopeful situation. The bar for standing out has dropped, not risen. One honest sentence that could only have been written by someone who was actually there now outperforms a flawless paragraph, and it takes less time to write. You just have to have been paying attention, and be able to remember that you were.
