Something new is happening in comment sections: people accusing content of being AI-generated, often incorrectly, and disengaging on that basis.
That reaction is worth taking seriously even if you think it's unfair, because it isn't really a judgement about provenance. It's a judgement about effort, and effort has always been a proxy for whether something is worth your time.
What the discount actually is
Audiences have always used cheap signals to decide whether to invest attention. Did someone bother? Is there a specific detail here that required work? Does this feel like it was made for me or produced at me?
Generated content trips those signals, and it trips them whether or not it was generated. What people are reacting to isn't a model. It's a texture: fluent, structurally tidy, confident, and non-specific.
That texture predates AI. Corporate marketing copy has had it for decades. What changed is that it's now cheap enough to be everywhere, and audiences have developed a name and a reflex for it.
What triggers it
Fairly consistent, and mostly avoidable:
Non-specificity. Claims that could belong to any company in your category. "We're passionate about quality" carries no information and reads as filler.
Structural tidiness. Perfectly balanced sections, three examples every time, a neat summary. Real thinking is lumpier: some parts get more attention because they deserve it.
Confident vagueness. Authoritative tone with nothing checkable underneath. This is the strongest trigger, and it's also the most damaging, because it's exactly what a model produces when it doesn't know something.
Numbers with no source. A statistic with no attribution now reads as fabricated by default. That's a real shift in the last couple of years and it's a rational one.
Stock-feeling visuals. Imagery that's technically fine and specifically nothing.
Volume without variation. Forty posts a month that all feel the same is itself a signal.
What survives it
The inverse, and each is expensive in a way that's the point.
Specificity that required access. A number from your own operation. A named customer with a real problem. What actually happened on the shoot day. Things a model cannot produce because it doesn't have them.
A position someone could disagree with. Generated content trends towards the consensus. That's what "most plausible next token" means in aggregate. A genuine, defensible, slightly uncomfortable opinion reads as human because it carries risk.
Visible imperfection. Not sloppiness. Texture: an aside, an admission, a thing that didn't work. "We tried this and it failed" is close to unfakeable, because nothing optimising for plausibility volunteers a failure.
Named accountability. A byline attached to a real person who can be argued with. This is why proof beats portfolio is becoming the norm on the creator side too.
Verifiable claims. Where a fact matters, linking the source. It costs one line and it moves you into a different category.
The counter-intuitive part
The response most brands have is to add more human signalling: more personality, more casual language, more emoji, more "we're just like you".
That mostly doesn't work, because tone is the cheapest thing to imitate. A model will do casual-and-personable perfectly on request. Signalling humanity through style is competing on exactly the axis where you have no advantage.
What works is signalling humanity through information a model couldn't have: the specific, the accessed, the observed, the admitted. That's not a style choice, it's a sourcing choice. That's why the constraint on good content moved upstream, into what you can get access to, rather than into how you write.
Using AI without earning the discount
None of this is an argument against the tools. It's an argument about where in the process they belong.
Use it upstream. Research, outlining, first drafts, variants, adaptation. The reader never meets the draft.
Add what it couldn't have. Every generated draft is missing the same things: your specific numbers, your actual customer, the thing that went wrong. Adding those is most of the work of making a draft yours.
Never let it originate a fact. Especially not a number about your own performance. Covered in what AI cannot tell you about your own content.
Cut the tidiness. If your piece has three examples in every section and a neat bow at the end, it reads generated even if it isn't. Let it be lumpy where the thinking was lumpy.
Read it as a sceptic before publishing. One question: is there a single sentence in here that only we could have written? If not, that's the edit.
Where Acumin fits
The trust discount is a good reason for the specific design decision underneath the product.
Acumin generates words; it does not generate figures. Numbers come from code, computed from real fetched data. The model composes the sentences around them and is not permitted to originate one. When there isn't enough evidence, the honest output is that there isn't enough evidence rather than a plausible-sounding claim.
That constraint costs something: sometimes the answer is less satisfying than a confident one would be. It's the right trade in a period where confident and unverifiable is the default output of every tool on the market, and where audiences have started pricing that in.
The same reasoning runs through the Creator Network: ratings exist only where a brand booked, work was delivered, and it was paid for. In an era where a portfolio can be generated, proof that someone else had to generate is worth more.
How to use this tomorrow
Take your last published piece and highlight every sentence containing something only your organisation could know: a specific number, a named person, an actual event.
If the highlights are sparse, that's the trust discount waiting to happen. Adding three of them is a twenty-minute edit with more effect than any rewrite.
Related: Why honest numbers matter more now is the analytical half of the same argument. What doesn't change covers the durable stuff underneath.