Do AI-written LinkedIn messages perform worse than human ones?
On acceptance they did, in Expandi's data; on replies they tied. Expandi, a LinkedIn automation company that offers AI message generation itself, compared AI-generated campaigns with human-written ones in its State of LinkedIn Outreach H2 2026 report. The page shows no publication date; its metadata shows a last update on 17 August 2026. Expandi says the analysis covers 118 campaigns from 73 accounts, run between October 2025 and June 2026.
It ran two comparisons, and they point in opposite directions:
| Comparison | Metric | AI-generated | Human-written | Difference |
|---|---|---|---|---|
| AI campaigns against all non-AI campaigns | Acceptance | 20.4% | 17.3% | +3.1 points for AI |
| AI campaigns against all non-AI campaigns | Reply rate | 6.5% | 4.6% | +1.9 points for AI |
| AI campaigns against the same accounts' own copy | Acceptance | 20.4% | 23.2% | -2.8 points (Expandi: about 12% lower) |
| AI campaigns against the same accounts' own copy | Reply rate | 6.5% | 6.6% | Flat |
Expandi calls the first comparison misleading and draws its conclusion from the second: AI-personalised messages don't yet outperform human-written templates, and its advice is to use AI as a drafting tool and rewrite the draft yourself.
Why did AI look better in the first comparison?
Because of who used it. Expandi says the accounts that adopted AI tend to be stronger operators across the board, and their non-AI campaigns already beat the platform average. Setting their AI campaigns against everyone else's campaigns produced what Expandi calls an apparent 18% acceptance lift, and Expandi describes that comparison as misleading.
The second comparison sets each account's AI-generated campaigns against that same account's human-written campaigns. The sender and profile stay the same, which controls for differences between senders, though Expandi notes the audiences chosen for AI may still differ. On that basis, human copy won on acceptance and the two tied on replies.
The lesson goes beyond Expandi. When any vendor says AI-written messages get more replies, ask what they were compared with. A comparison against average users mixes the tool's effect with the skill of the people who chose it.
- Weak comparison: AI users against all other users. If stronger operators are the ones adopting the tool, as Expandi found, AI looks good.
- Better comparison: the same accounts, AI copy against their own copy, over the same period.
- Best comparison: one list split at random between AI drafts and your own messages, sent from one account in the same weeks. Expandi's report does not describe a test like this.
What the comparison does not show
It shows a pattern on one platform; it doesn't prove that AI writing lowers acceptance. Expandi says all of its findings are observational and show association, not causation. Before you act on the 12%, note these limits:
- A small sample next to the rest of the report. 118 campaigns from 73 accounts, against 13.2 million connection requests in the report's main dataset. Expandi doesn't say whether the 118 are only the AI campaigns or how many human-written campaigns the same accounts ran, and the non-AI side of its first comparison is a broader population of unstated size.
- Expandi's own three caveats. AI copy may feel formulaic to recipients who have seen the pattern before. Users may be putting AI on their hardest, coldest audiences. And the AI campaigns are newer, so late acceptances may not have come in yet.
- One AI feature on one platform. Expandi's FAQ says the feature launched in late 2025. In its H1 2026 report, published 13 January 2026, Expandi said it did not yet have enough data to benchmark its AI capabilities.
- No meetings and no reply quality. The comparison reports acceptance and reply rate only, and gives no separate definition of reply rate for this analysis.
- Different scale from the headline numbers. These acceptance rates (17.3% to 23.2%) sit below Expandi's 28.5% platform average. Its methodology says rates from its research dataset and its main dataset are not directly comparable in absolute terms.
- Unclear scope. The report doesn't say whether AI wrote the connection note, the follow-ups or both. Its section heading calls them unedited AI-generated messages but doesn't explain how editing was tracked.
Why might AI-written messages get accepted less?
Expandi's first explanation is that the copy can read as personalised to the sender while feeling formulaic to a recipient who has seen the pattern before. Our reading is that profile-based tools tend to reach for the same openers: a job title, a recent post, a company milestone. If several messages a week open the same way, the detail stops reading as personal.
A separate line of research bears on whether to say AI helped. In a paper in Organizational Behavior and Human Decision Processes (volume 188, 2025), Oliver Schilke and Martin Reimann of the University of Arizona's Eller College of Management report that across thirteen experiments, people who disclosed their AI use were trusted less than people who did not. The effect held whether disclosure was voluntary or mandatory. It was weaker than when a third party exposed the AI use, and smaller, but not gone, among evaluators who felt positive about technology. In an article by the authors on the University of Arizona's news site (29 May 2025), they put the participant count at more than 5,000, and they write that being found out later cost the most trust, so being upfront may be the better policy.
Their abstract describes tasks from communications to analytics and artistry and names no sales task, so treat this as a pointer. Taken with Expandi's data, it suggests a test for every AI draft: would you be comfortable explaining on a call where each detail came from and why you wrote to this person?
How to use AI for LinkedIn messages without the penalty
Use AI for research and a first draft, then make the message yours before it goes. These steps are our advice; the last one copies Expandi's within-account method so you can check the result on your own audience.
Give the AI the reason, not just the profile
Feed it the event that made you write: a post about the problem, a new role, a hiring push. A profile alone tends to produce compliments; a reason produces a point.
Check every detail
Open the post or page the draft refers to. If a fact is old, wrong or vague, cut it. One true, specific detail beats three generic ones.
Rewrite the first line and the question
These are the parts recipients judge. Delete stock openers such as "I came across your profile" or praise for someone's "impressive journey", and ask a question only you would ask.
Keep it to about two sentences
In Expandi's separate length analysis, single messages of 151 to 200 characters had the best reply rate among well-sampled lengths, at 8.5%.
Test against your own copy
Split one list in two at random. Send AI-assisted messages to half and your usual message to the other half, from the same account in the same weeks. Compare acceptance, replies and meetings booked.
Here is what the rewrite step looks like in practice. These are examples we wrote, not messages from Expandi's data:
| What the AI was given | Typical first draft (example) | After a human rewrite (example) |
|---|---|---|
| Profile only | "I came across your profile and was impressed by your experience in operations." | No rewrite: there is no reason to write yet. Wait for a signal. |
| Their post saying month-end close takes eight days | "Loved your recent post about finance! Would love to connect." | "Your post said close takes your team eight days. Is the slow part approvals or bank reconciliations?" |
| A new role as head of sales | "Congrats on the new role! Exciting times ahead." | "Congrats on the move. Are you reviewing the sales tools in your first quarter, or keeping what is there?" |
For first messages you can adapt by hand, see our LinkedIn follow-up message templates.
How Siela's AI SDR writes LinkedIn and X messages
Siela's AI SDR writes each LinkedIn and X message for that person from their profile and the signal that surfaced them, such as a post about the problem you solve or a role change. A second AI reviewer then checks the message before it is sent. Once you press Save and start, it sends on its own within the daily goals you set.
We have no acceptance or reply rates of our own to publish, so we make no claim against Expandi's numbers. What Expandi's comparison does give every buyer of AI writing, ours included, is a fair test: judge the tool's messages against your own, from your own account, over the same weeks, and count meetings as well as replies. Our post on whether AI SDRs work covers the other questions to ask before you hand over outreach.
Frequently asked questions
Are AI-written LinkedIn messages less effective than human-written ones?
On acceptance, in Expandi's own data, yes. When the same accounts used both, human-written campaigns were accepted at 23.2% and AI-generated ones at 20.4%, about 12% lower. Reply rates were flat at 6.6% and 6.5%. Expandi cites 118 campaigns from 73 accounts between October 2025 and June 2026, without the AI and human split, and the analysis is observational, so it shows a pattern, not a cause.
Why did AI messages look better in Expandi's first comparison?
Because of who used them. Against all non-AI campaigns, AI campaigns showed 20.4% acceptance against 17.3%. Expandi says the accounts that use AI tend to be stronger operators whose non-AI campaigns already beat the platform average. Compared with those same accounts' human-written campaigns, the AI campaigns were accepted less.
Should you tell prospects that AI helped write your LinkedIn message?
Research suggests disclosure costs trust. Across thirteen experiments, Oliver Schilke and Martin Reimann found people who disclosed AI use were trusted less, but being exposed by a third party cost more trust, and the authors suggest being upfront may be the better policy. Their abstract names no sales task. Whatever you decide, send only details you checked and would stand behind on a call.
How do you test AI-written LinkedIn messages fairly?
Compare AI drafts with your own messages from the same account over the same weeks, as Expandi did in its within-account analysis. Better still, split one list at random between the two versions. Track acceptance, replies and meetings booked, because a message that gets more replies can still book fewer calls.
Does Expandi's result apply to every AI writing tool?
Not necessarily. It measured AI-personalised campaigns on Expandi's own platform, a feature its FAQ says launched in late 2025. Expandi lists other possible reasons for the gap, such as AI being used on colder audiences and newer campaigns still collecting acceptances. Other tools write differently, so test any tool against your own copy before relying on it.
Sources
- Expandi: The State of LinkedIn Outreach H2 2026 (no date on page; metadata shows update 17 Aug 2026; checked 20 Oct 2026)
- Expandi: The State of LinkedIn Outreach H1 2026 (published 13 Jan 2026)
- Schilke and Reimann: The transparency dilemma: How AI disclosure erodes trust (abstract, OBHDP vol. 188) (2025)
- University of Arizona: Being honest about using AI at work makes people trust you less (published 29 May 2025)
- Siela: AI SDR (checked 20 Oct 2026)