Hidden signatures

Once you strip away Markdown, formatting, and HTML classes, you are left with raw data (either flat string text or raw JSON values). At this foundational level, qualifying copy as a "legitimate human-created asset" rather than "synthetic, machine-created copy" relies on two completely different methodologies: algorithmic/statistical markers and forensic provenance.
Because large language models (LLMs) have evolved past simple grammatical errors or stylistic "hallmarks," detectors and platforms look for hidden signatures that separate human behavior from machine probability.

1. Algorithmic Markers: Linguistic Chaos vs. Smooth Math

When an AI detector or an enterprise verification engine processes raw text, it ignores vocabulary and looks strictly at the mathematical distribution of the text. Humans are naturally chaotic and erratic writers; machines are highly optimized prediction engines. [1, 2]

A. Perplexity (The Predictability Metric)

  • Synthetic Copy: LLMs operate by calculating the statistical probability of the next word. Because they favor safe, mathematically sound language, synthetic text has low perplexity—it is highly predictable to another machine. [2, 3, 4]
  • Human Copy: Humans write with high perplexity. A human will use unexpected adjectives, localized idioms, irregular slang, or sudden stylistic left-turns that a predictive model would rank as a "low probability" choice. [1, 2]

B. Burstiness (The Pacing Metric)

  • Synthetic Copy: Machines have an incredibly stable, uniform pace. AI-generated paragraphs feature sentences of roughly the same length, structurally balanced dependent clauses, and regular placement of transitional words (e.g., Furthermore, Moreover, In conclusion). This results in low burstiness. [1, 2, 4, 5, 6]
  • Human Copy: Humans naturally write in "bursts". A human might follow a complex, meandering 40-word sentence with a fragment. Or a two-word punch. They write how they speak, resulting in highly erratic sentence rhythms (high burstiness). [1, 2, 4]

2. Algorithmic Watermarking (The "Green List" Secret Rulebook)

Major AI providers like Google and Anthropic use native, mathematical watermarking systems like Google DeepMind's SynthID-Text during text generation. [7, 8, 9] 
[AI Generation Stage] -> [Secret Key divides dictionary into Red/Green words] -> [AI subtly favors Green words] -> [Detector tallies Green words]
  • How it qualifies text: When generating an answer, the LLM's vocabulary database is secretly split into "Green words" and "Red words" via a pseudo-random key. The model is subtly nudged to pick Green words.
  • To a human reading raw text, it looks completely normal. But if a verification engine runs that text against the model's key and finds a statistically impossible cluster of "Green words," the entry is immediately flagged as synthetic. A human writer—having no access to this hidden mathematical rulebook—will hit Green and Red words at a baseline 50/50 distribution. [9, 10]

3. Forensic Provenance: How the Text Was Born

In an enterprise data pipeline, you do not just verify the text after it's written; you look at the metadata tracking how it was written. Legitimate human copy leaves a digital trail of physical creation.
Human Creation SignatureSynthetic Generation Signature
Keystroke Dynamics: The raw text entry is accompanied by telemetry showing varied typing speeds, backspacing, pauses for thought, and manual edits over time.Instantaneous Injection: A 2,000-word entry is pasted into a field or database in a single millisecond via an API call or Ctrl+V.
Iterative Revision: Draft versioning history shows structural evolution, deleted paragraphs, and gradual refinement.Single-Shot Output: The copy exists in its final, grammatically flawless state on version 1.0.
Hyper-Local Context: Inclusion of highly specific, unfalsifiable personal anecdotes, verifiable real-time physical experiences, or niche local dialect.Polished Generalizations: Highly articulate but broad prose that relies on consensus data ("Studies show...", "It is important to note...").

Summary

Ultimately, raw text qualifies as human-created if it passes the "Chaos Test" (high perplexity, variable sentence lengths, lack of an algorithmic watermark) or if it is supported by provenance data proving a human physically typed and revised the document over time. [2, 4, 10]
Are you asking this because you want to build a pipeline to filter out synthetic web-scraped data for training a model, or are you trying to verify the authenticity of content submitted by human writers?

Comments

Post a Comment

Empathy recommended

Popular posts from this blog

Supporting Artistes (SAs)

"Dave Watts, an experienced SA who has appeared in numerous superhero movies and major productions, has been scanned several times. He said there were wider implications for the industry. "'I already hear crew members saying: To be honest, we don’t even need to do this any more. We can just ask AI to create a crowd of 1,000 people based on information which has already been captured [by body scanning],' he said. "'If you don’t have your usual crowd of 100, 200 or 500 SAs on a big production, then you also don’t need the assistant directors that look after them, and you don’t need the hair and makeup people. You don’t need the costume people, the costume fittings, all the caterers, all the drivers and location marshals. There’s a whole range of jobs there that AI effectively puts at risk'."
Read more

Hamza Chaudhry

"Although a nuclear confrontation based on fake intelligence may seem unlikely, the stakes during crises are high and timelines are short, creating situations where fake data could well tilt the balance toward nuclear war. "The evolution of nuclear systems has led to further ambiguity in crises and shorter timeframes for verifying intelligence. An intercontinental ballistic missile from Russia could reach the U.S. within 25 minutes. A submarine-launched ballistic missile could arrive even sooner.  "Many modern missiles carry ambiguous payloads, making it unclear whether they are nuclear-tipped. AI tools for verifying the authenticity of content are not sufficiently reliable, making this ambiguity difficult to resolve in a short window. "The likeliest nuclear hotspots are also the arenas involving actors with low levels of trust on both sides — be that the U.S. and the Chinese Communist Party or Russia, or India and Pakistan. Even if communication is established in a...
Read more

Injection

Image
✨ AI Mode The mirrored text in the image can be read as: The Thai characters "มอเมร" (mo-maeh-ra) are mirrored. When unmirrored, they spell "re-mae-mo" in the reverse order. The Latin-alphabet text "enimnot mion" when unmirrored becomes "noim tonmine" which seems to be a non-word or phrase. The number "2169" when unmirrored is "9612". The text "9VIJEMot" when unmirrored becomes "to MEJIV9". The full unmirrored text is "re-mae-mo noim tonmine 9612 to MEJIV9 noim tonmine re-mae-mo". ✨AI can make mistakes, so double-check responses [Turned Google lens onto this image —that I created —to see whether it could not only read the text but also unmirror it…] [I'd never done anything like this before on a whim…] [But I knew that untried things sooner or later get tried by somebody, so why not me?] [I forget what I pointed at after highlighting the image. Interesting search questions appeared beneath...
Read more