Detecting AI-Generated Content From Web Structure Alone: A New Approach

Identifying AI Content From Structure, Not Text

As large language models become increasingly widespread, more and more web content is being mass-produced by AI. Traditional AI content detection methods mostly rely on linguistic features of the text itself — such as perplexity, vocabulary distribution, and sentence patterns. But these approaches share a common weakness: once the text has been manually rewritten or polished, detection accuracy drops significantly.

A post shared on the Hacker News community proposes a different angle — identifying whether content is AI-generated based solely on a webpage's structural information. In other words, instead of reading what the text says, the model observes external features like HTML structure, layout, and how tags are organized.

Perplexity is one of the most commonly used metrics for AI text detection. It measures how "surprised" a language model is by a given piece of text — AI-generated text tends to be highly predictable to language models, resulting in low perplexity, while human writing is typically more creative and unpredictable, yielding relatively higher perplexity. Detection tools like those from OpenAI and GPTZero rely on this principle to varying degrees. The fatal flaw of perplexity-based methods, however, is that a simple synonym substitution or sentence restructuring can significantly shift the perplexity distribution, rendering detection ineffective. This is precisely the core pain point that structure-based detection aims to address.

Why Structural Signals Might Work

The logic behind this idea is worth examining. A large volume of AI-generated web content — especially pages mass-produced for SEO or content farms — tends to share similar generation pipelines and templates. At the structural level, these pages leave identifiable "fingerprints":

• Highly uniform, formulaic paragraph breaks
• Mechanical repetition of heading hierarchies (H2/H3)
• Suspiciously consistent use of lists, bold text, and subheadings
• Page skeletons that look nearly identical, as if produced by the same generation script

By contrast, pages written and formatted by hand tend to be structurally more "organic" — with irregular paragraph lengths and personalized layout choices. These differences make structure itself a potential discriminating signal.

The Robustness Advantage of Structural Features

The most appealing aspect of structure-based detection is its robustness. Even if someone rewrites AI-generated text sentence by sentence and swaps out vocabulary, as long as the page still uses the same generation template, the structural features remain intact. This means such methods could potentially sidestep the "rewrite and evade" problem that plagues text-based detection, making them a powerful complement to linguistic approaches.

Potential Applications and Limitations

If models of this kind prove sufficiently reliable, they could be useful in several scenarios: search engines could use them to filter out content farms; platforms could flag pages suspected of being bulk-generated; and researchers could use them to quantify the proportion of AI-generated content on the web.

That said, this approach faces some clear challenges. First, structured layouts aren't exclusive to AI — many human-used CMS systems and web templates also produce highly uniform structures, making false positives a real concern. Second, once detection rules become public, content generators can deliberately scramble their structure to evade detection, triggering a new round of adversarial cat-and-mouse.

An Early-Stage Exploration Worth Watching

It's worth noting that this content has attracted limited discussion on Hacker News so far (5 points, no comments at the time of writing) — it's more of an early experimental idea than a mature solution. Its value lies in offering a fresh perspective: in the ongoing arms race of AI content detection, perhaps we should be looking at the "skeleton" of content, not just the text itself.

As AI-generated content occupies an ever-growing share of the web, multi-dimensional detection — combining text features, structural features, and behavioral signals — is likely the direction of the future. A single dimension is easy to circumvent; a combination of signals is far harder to fake.

Content farms are websites or operations that mass-produce low-quality web pages for the purpose of capturing search engine traffic. Before large language models became widespread, content farms relied on low-paid outsourced writers; today, AI has dramatically lowered the cost of bulk production, allowing a single operator to generate thousands of seemingly complete articles in hours. This type of content typically relies heavily on fixed templates — formulaic titles, cookie-cutter structures — which is precisely what gives structure-based detection something to work with. Search engines like Google have updated their algorithms multiple times (such as the 2022 "Helpful Content Update") specifically to demote such content, but the battle between detection and evasion never truly ends.