Does AI-assisted content actually perform worse?

Three credible studies asked whether AI content underperforms and got three different answers. The disagreement is the finding — and it tells you exactly where the line is.

8 minute read

The three studies, and what each actually measured. They disagree because they asked different questions.

StudyDesignWhat it measuredResult
Buffer (2024)1.2M posts, ~15,000 posters, within-person mediansAI-assisted, unlabelled5.87% vs 4.82% — AI-assisted wins
Fleseriu & Fron (2025)24 posts, one 770-follower account, alternating orderAI-generated, unlabelledNo meaningful difference
Electronic Markets (2026)Two experiments, n = 325 and n = 371AI-labelled vs human-labelledLabels reduce engagement, worst on emotional content
Sprout Social (2026)2,300+ consumers, 1,200+ marketersWhat consumers say they wantUnlabelled AI is the #1 thing to stop

Three studies, three answers

If you have gone looking for whether AI-assisted posts perform worse, you have probably found a confident answer in both directions. That is because the credible research genuinely disagrees. Buffer analysed 1.2 million posts and found AI-assisted content earned a higher median engagement rate than human-only content. A peer-reviewed field experiment on Instagram published twelve human posts and twelve AI posts on the same account and found no meaningful difference. A pair of controlled experiments published in Electronic Markets found that labelling content as AI-generated reduced engagement compared to identical human-created content. None of these is wrong. They are measuring three different things, and once you see which, the practical answer falls out.

The biggest dataset says AI-assisted wins

Buffer's analysis is the largest of the three by a wide margin: 1.2 million posts across Facebook, LinkedIn, Pinterest, Threads, TikTok, and YouTube. Critically, the final analysis was restricted to about 15,000 people who published both AI-assisted and non-AI-assisted posts, and compared each person's median engagement rate against their own. That within-person design matters — it removes the obvious objection that AI users might simply have bigger or better audiences. Across all platforms, AI-assisted posts reached a 5.87% median engagement rate against 4.82% for non-AI posts. Threads showed the widest gap, roughly double. LinkedIn showed the narrowest, 6.85% against 6.22%.

Threads11.11%LinkedIn6.85%TikTok6.14%Facebook6.13%Pinterest4.35%YouTube3.90%X (Twitter)3.70%
Median engagement rate of AI-assisted posts by platform. Non-AI medians for the same users were 5.56%, 6.22%, 4.17%, 4.89%, 3.86%, 3.7% and 2.8% respectively. Buffer, 1.2M posts from ~15,000 posters.

The controlled experiment says it is a tie

Fleseriu and Fron ran a within-subjects field experiment on a real Instagram account with 770 followers, publishing 24 single-image culinary posts — twelve human-created, twelve AI-generated — in alternating order, and testing the difference with Mann-Whitney U. Their conclusion was that AI content has the potential to generate similar outcomes across likes, comments, shares, saves, reach, impressions and engagement rate. The differences that did show up are interesting: human posts drew more comments and more hashtag impressions, while AI posts drew more profile visits and more reach among non-followers. It is one account in one niche, so do not over-read it. But it is a properly controlled test, and it did not find the collapse that AI sceptics predict.

The third study finds the actual penalty

The Electronic Markets paper is the one that resolves the other two. Across two online experiments (n = 325 and n = 371), participants viewed Instagram profiles whose content was labelled as human-created, AI-enhanced, or AI-generated. Labelling content as AI-enhanced or AI-generated reduced both affective and behavioural engagement compared with human-created content — and the penalty was largest for emotional content. There is a second finding buried in there that is genuinely useful: emotional content beats rational content when it is labelled human-created, that advantage shrinks as AI involvement rises, and for fully AI-generated content it reverses, with rational content doing slightly better. In other words, the emotional register is exactly where AI attribution costs you most.

Assisted, generated, disclosed: three different questions

Line the three up and the contradiction disappears. Buffer measured AI-assisted posts — drafted with an assistant, then published by a human who could edit them, with no label attached. The Instagram experiment measured AI-generated posts, unlabelled, in a low-emotion niche. The Electronic Markets experiments measured AI-labelled content, holding everything else constant. The penalty does not attach to using AI. It attaches to the audience knowing, and it lands hardest on the content that was supposed to feel personal. The comparison table above is deliberately a table rather than a chart: the three studies measured different metrics on different populations, so there is no shared axis to plot them on.

What your audience says it wants

Sprout Social's 2026 content strategy research, drawn from 2,300+ consumers and 1,200+ marketers, found that the single most common thing consumers want brands to stop doing is posting AI-generated content without clearly labelling it. Its Q1 2026 pulse survey of 2,250 social users across the US, UK and Australia found 56% see low-quality AI content often or very often, and 66% say they are more selective about what they engage with than a year ago. Read those next to the Electronic Markets result and you have a genuinely awkward position: audiences say they want disclosure, and disclosure measurably costs engagement. Anyone selling you a clean answer here has not read both papers.

What to actually do with this

The defensible reading is that AI is fine in the parts of the job nobody was going to credit you for anyway, and expensive in the parts where being you is the product. Use it for the mechanical layer — cutting clips, drafting variants, writing alt text, pulling the quotable line out of a transcript, reformatting one recording for four platforms. Keep your hands on the parts that carry your judgement: the take, the story, the reply in the comments. That is also the honest reading of Buffer's own result, since every post in its AI-assisted group passed through a human before it published. Nobody measured what happens when you publish the first draft, because that is not what those 15,000 people were doing.

FAQ

Do social platforms suppress AI-generated content?

There is no documented ranking penalty for AI use on the major platforms. The measured penalty in the research comes from audiences, not algorithms: when content is labelled as AI-generated, engagement falls. Platforms do act on synthetic media that is deceptive, and several apply their own AI labels to detected content.

Should I label my AI-assisted posts?

It depends on how much AI is in them. Platform rules generally require disclosure for realistic synthetic media — a photoreal image or a voice that did not happen — not for using an assistant to draft a caption. Beyond the rules, the Electronic Markets research shows labels cost engagement, and the Sprout data shows audiences resent undisclosed AI, so the safe ground is to keep AI out of the parts that claim to be a real moment.

Why did Buffer find AI-assisted posts perform better?

Because every post in that dataset was edited and published by a human, and none carried an AI label. It measures assistance, not automation. It is also a within-person comparison — each poster is compared against their own non-AI posts — so it is not explained away by AI users having better audiences.

Does AI-written content rank on Google?

Google's guidance targets quality and scaled content abuse rather than production method. That is a separate question from social engagement, and we cover it in detail in our post on whether AI content ranks.

Sources

Related pages