The Maze: NP Digital reviewed social content from 100 companies and found a production machine running ahead of its audience. AI generated 52.1% of new posts in the reported categories, while human work represented 14.5%. Yet human posts averaged 15.7 likes — 2.15 times AI's 7.3 — and three times as many saves. A human edit did not close the gap. The useful lesson is not “ban AI.” It is that inexpensive output does not create an expensive idea. AI can multiply drafts. Original observation, judgment and taste still decide whether anyone cares.
The volume engine is working. The affection engine is not. AI-generated posts supplied 52.1% of new content, and AI plus a human edit supplied another 27.4%. Human-generated posts were only 14.5%. The three reported categories total 94%, so six percentage points remain unexplained. Even with that gap, the operating pattern is clear: companies have industrialized production much faster than they have industrialized relevance.
Human work wins the higher-intent signals. Human posts averaged 15.7 likes, versus 7.3 for AI and 6.8 for AI edited by a person. Saves ran 0.3, 0.1 and 0.2 respectively. The lightly edited group underperforming pure AI on likes matters. Moving words after generation is not the same as contributing a premise, lived observation, concrete example or opinion worth remembering.
Shares stop this becoming an anti-AI sermon. AI and hybrid posts both averaged 0.2 shares, compared with 0.1 for human posts. All three figures are tiny and rounded, but the direction still blocks a lazy conclusion that human content wins every action. Different signals reward different jobs. A useful explainer may travel. A distinctive point of view may earn approval and saves. One engagement metric cannot stand in for attention, trust and distribution at once.
Disclosure can change the audience before quality does. A 2026 controlled study showed identical social content to users with human-created, AI-enhanced or AI-generated labels. AI labels reduced emotional and behavioral engagement, especially for emotional posts. That supports an authenticity mechanism, not NP Digital's suggestion that algorithms suppressed its sample. No reach, label or ranking data was published for the 100 companies.
AI performance is a design choice, not a fixed law. A separate 892-person experiment found an advantage for AI-created social content in its tested Facebook, Instagram and X formats. The contradiction is useful. Prompt quality, topic, platform fit, disclosure and the originating idea all matter. Teams should stop debating whether AI is “good” and start measuring where it adds leverage without removing the reason to pay attention.
Why it matters: Content economics have split in two. The cost of producing a competent post is collapsing, while the value of a distinctive thought is rising. The winning workflow uses AI for research, alternatives, structure and speed, then spends human time where imitation fails: choosing the tension, adding evidence, making a judgment and deleting the generic ending. Measure likes, saves, shares and qualified response separately. If the team optimizes only for output, it will manufacture more of the cheapest resource on the internet — plausible sentences — and wonder why attention remains scarce.
