How do you distinguish AI content from real, human-made content?

I agree with the point about text and video.

But this is still an interaction, and interactions carry dynamics that shape how something feels: response timing and adaptability, tone, approach, and even the presence — or absence — of certain patterns.

So it’s not just about the two sides, but also the flow between them.

Ok first of all the question is on the point. Nowadays, everything from a simple email message to writing a description about your personal startup is full of AI. Sometimes it irritates me to see people using these AI tools without knowing like where to use these tools. The AI written text can be detected by the human eye itself, it is too polished, too perfect, sounds like written by direct oxford dictionary. and second if you talking about image it is visible too like a 3d type image, looking cartoon more than looking a natural brand photo. even the designers nowadays using image creation tools instead of creating them manually. Seriously, like where we are heading too.

Thanks,
Nexbloggy

I think the main differentiator between AI content and real human-made content is clearly the taste and the individual presence behind the post or content. You can tell when someone has genuinely put effort across in whatever they're doing, whether they used AI to generate it or not.

In highly technical domains the tell isn't stylistic — it's whether the content demonstrates live knowledge that can't be reconstructed from training data.

I write about project finance and M&A modeling. An AI answer in my domain gets the framework right — DSCR, debt sculpting, cash waterfall — because that's in the training data. What it can't do is tell you that a specific lender will require 18 months of DSRA for a merchant wind project but accept 6 months for a fully contracted solar plant, or why the P90 exceedance ratio that matters for debt sizing is the 10-year figure, not the 1-year. Those calibrations come from sitting in credit committee rooms, not from reading about them.

So my practical test for technical content: does the author commit to a specific number, a named counterparty, or a decision that would be wrong in the wrong context? Generic advice is always safe and always AI-compatible. The thing that can get you fired if you apply it to the wrong deal — that's the signal.

On images and video: agree with Saad that effort is the better frame than AI vs. not-AI. The tells are more about whether the visual is doing real work — showing something specific, dated, and contextually accurate — or just looking the part.

Landing pages and websites are the easiest to determine tbh - After having built enough prototypes it's incredibly easy to see which landing pages have no character but the standard template and animations.
Video is getting scary good but for people who know how to use the tools - what is the m-dash equivalent of the level of forced bokeh? for images, the tell tale signs of the past (6 fingers, wrong spellings) is also disappearing so one has to really look for imperfections for standard images. I have found that AI still struggles massively with illustrations and flowcharts for custom product flows, and I just rebuild the whole thing in Figma after I've seen a few suggestions from whatever AI tool I'm using that week.

For video and images it's getting really hard to tell honestly.

But I think the bigger lever will be platform policy. TikTok, Instagram, X — they'll decide how AI content gets labeled and how it gets distributed. Meta could easily say "content made with our tools gets a boost, everything else gets average reach." Human-made content might literally get a gold star one day.

That's actually the direction we're going with @Magic . Real footage with proper copyright, combined with CG or generative AI where it makes sense. Not pure generation, not pure human — somewhere in between where quality and authenticity both hold up.

The hybrid approach won't just be a creative choice. It might become a distribution strategy too.

interesting anti-grammarly instance - how crazy that the pendulum has swung so far haha - I'm starting to see a real pivot back - almost towards - embellished authenticity - more casual tonality - and not only in text but within video content / ads as well.

I think in text - I am trying to write in a way that is... more stream of consciousness.. which stylistically can be a nightmare or.. potentially come off as authentic.

That being said - I've messed around with a claude skill for this and even in its current state - the outputs are spookily close to home.

M dashes - its not X its Y - also another obvious one is the same word used within the pre header - header and sub header of a website. Ex.

"build stronger headlines

The tool that helps write with ease

better headlines - less friction."

You'd see like - build twice or headlines twice and oftentimes I think people don't even audit their AI and just ship it (guilty for some low stakes copy)

I think we're still in a place where AI video is relatively obvious - but for rich media - I care less for some reason as a consumer.

If a video is AI or a human made the graphics.. or if someone did a frame by frame video using claymation - I am not too bothered as long as the end product is of interest / gets the point across.

This is a topic I'm interested because almost everyone LinkedIn has turned into an expert at recognising AI content. 😂

I'll be reading the comments.

It's so sad that AI occupied the long dash.

Such a wonderful stylistic device.

The patterns Nika describes show up clearly in financial analysis too. AI-generated financial content tends to be structurally fluent but contextually flat — it can explain what a DSCR is, but it struggles with the judgment call: given this specific lender, this specific project, and this macro environment, what should the DSCR floor be, and why does it differ from the standard covenant? That requires having sat across a credit committee, not just having processed thousands of credit papers.

The tell I notice most: AI financial content defaults to the median. It gives you the textbook answer, not the deal-specific one. A practitioner knows that a 1.20x DSCR floor might be appropriate for a contracted solar project with an investment-grade offtaker, and completely wrong for a merchant wind project in a volatile power market — same metric, fundamentally different answer, and the difference requires experience the model doesn’t have.

For distinguishing it in documents: AI reads as authoritative on concepts but vague on specifics. Real deal analysis tends to be the reverse — telegraphic on the concepts (your reader knows what a DSRA is) but precise on the numbers, the counterparty, the stress case logic, and why those particular assumptions were chosen. That specificity is hard to fake at volume.