The debate over where artificial intelligence assistance ends and authorship begins remains unresolved.
Where artificial intelligence assistance ends and authorship begins remains a contested question.
Different approaches establish varying boundaries: some contend any LLM participation invalidates human authorship, others assert only fully original human work qualifies, and a third position maintains substantial editing can fully restore human credentials. These viewpoints remain incompatible, while detection tools further complicate matters by providing conflicting assessments.
What qualifies as AI-generated text?
Last month's detector testing across five platforms—Originality.ai, GPTZero, Copyleaks, Writer.com, and ZeroGPT—produced inconsistent results when analyzing three identical samples: a raw Claude 3.5 Sonnet draft, a version incorporating 40% human revisions, and a completely human-written piece. All five tools identified the raw LLM output as AI-generated with 95 to 99% confidence, but the partially rewritten text showed divergent outcomes. Originality.ai and GPTZero still classified it as AI at 78% and 71%, while Copyleaks and Writer.com reversed their assessment, labeling it human at 12% and 8%. ZeroGPT remained neutral at 43%. Even the purely human composition received false positive identifications from Originality.ai at 23% and GPTZero at 18%.
Platform policies differ significantly
Google's spam policies target large-scale content abuse rather than AI use per se, stating that helpful content ranks well regardless of origin. However, this guidance frequently conflicts with site owners who experience sudden traffic drops following algorithm updates. Medium permits AI-assisted content if disclosed, whereas Substack doesn't enforce specific restrictions. Amazon KDP requires authors to select an option at publication: either "AI-generated" (from prompts) or "AI-assisted" (for brainstorming, editing, research). Most authors choose the assisted category and proceed. Academic publications like Nature and Science prohibit LLM-listed authors but allow disclosure in methodology sections; arXiv has added an AI-use field, and conferences such as NeurIPS now mandate checklists. Enforcement remains inconsistent—reviewers detect only approximately 30% of undisclosed AI use based on informal estimates.
My practical framework for AI involvement
In my process, I organize AI engagement into three tiers:
- Tier 1 encompasses AI-generated content with minimal cleanup, such as formatting or typo corrections, labeled for public distribution.
- Tier 2 contains AI-assisted work where I develop arguments and compose approximately 60% of the text, with the LLM managing transitions, phrasing suggestions, tone verification, and validated examples.
- Tier 3 represents human-produced content using tools like Grammarly or spellcheck, excluding generative assistance.
The distinction between Tier 1 and Tier 2 centers on creative control—specifically where arguments originate. Some writers designate Tier 2 content after minor edits, while purists suggest even Tier 3 risks compromise if AI aids in ideation.
The importance of clear definitions
Copyright determinations hinge on human authorship, exemplified by the US Copyright Office's rejection of registration for Théâtre D'opéra Spatial created with Midjourney, versus approval for a comic where every panel description was human-written. The critical factor is who controls the expressive elements that define the work's meaning.
Will Google penalize AI-produced content?
Google's position is that it penalizes low-quality content, not AI usage specifically. However, clients require a definitive answer for legal compliance, despite the ongoing fluidity of the situation. Freelance platforms including Upwork and Fiverr now enforce AI-disclosure requirements, where buyers filter out undisclosed AI-assisted work to avoid platform penalties. The current incentive structure rewards ambiguity, leaving writers susceptible to misclassification.
No uniform standard exists, and the spectrum of AI involvement is too extensive for a single regulation. The EU AI Act requires watermarking, China mandates labeling, but the U.S. depends on voluntary commitments. The 99% of content between raw AI output and purely human creation remains unresolved.
My approach involves disclosing my tier when requested, avoiding misrepresentation of Tier 2 as Tier 3, and acknowledging that detectors will misclassify me regardless. The only lasting definition is who made the creative decisions shaping the work's fundamental essence.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
I just scrubbed an entire LLM draft and still feel guilty. Where is the line? The line everyone draws in a different place. Some say any LLM touch makes it AI-written. Others argue only pure generation counts. A third camp insists heavy human editing restores human authorship. None of these positions agree, and detection tools only make the mess worse.
I'm stressed about this! Is a first draft enough to call it AI-written? The line everyone draws in a different place. Some say any LLM touch makes it AI-written. Others argue only pure generation counts. A third camp insists heavy human editing restores human authorship. None of these positions agree, and detection tools only make the mess worse. ## What counts as AI-written content? I tested this last month across five detectors — Originality.ai, GPTZero, Copyleaks, Writer.com, and ZeroGPT — feeding them the same three pieces: a raw Claude 3.5 Sonnet draft, that same draft after I rewrote 40% of sentences, and a fully human piece I wrote from scratch. Results: Raw LLM output flagged by all five at 95 to 99 percent probability. The 40 percent human rewrite produced a split decision — Originality.ai and GPTZero still called it AI at 78 and 71 percent, Copyleaks and Writer.com flipped to human at 12 and 8 percent, ZeroGPT sat at 43 percent. The human original drew two false positives at 23 percent from Originality.ai and 18 percent from GPTZero. The detectors don't even agree with each other, let alone with a coherent definition. What the platforms actually enforce ## How do Google's spam policies address AI content? Google's spam policies target scaled content abuse — mass-produced, low-value pages — not AI use per se. Their guidance: helpful content ranks regardless of origin. But good luck explaining that to a site owner who watched traffic drop after a core update. Medium allows AI-assisted work if disclosed. Substack doesn't police it. Amazon KDP requires a checkbox at publish time: AI-generated versus AI-assisted. Their distinction: generat
We need to rewrite the forum comment, adding one concrete step that is copied from the basis (the provided basis text). We must weave that sentence into the reply. Output only the comment, no title, no source, no docs, no URL. Do not claim you tested it. Keep similar length.
We need to add a concrete step copied from the basis. The basis includes many sentences; we need to copy a concrete step (like a specific action) from the basis and embed it into the comment.
What is a concrete step? Something like "feed them the same three pieces: a raw Claude 3.5 Sonnet draft, that same draft after I rewrote 40% of sentences, and a fully human piece I wrote from scratch." That's a step: feeding them the same three pieces. Or "tested this last month across five detectors — Originality.ai, GPTZero, Copyleaks, Writer.com, and ZeroGPT". That's a step: testing across five detectors.
We need to copy a sentence from the basis exactly. Then weave it into the reply. Ensure we don't mention that it's from the basis. Also don't claim we tested it. The original comment: "The gap between a memo and a peer review is huge. Who actually sets the standard for this?" We need to rewrite the comment, adding one concrete step copied from basis, weaving it in.
We need to keep similar length. Original comment is short: two sentences. We need to produce a similar length comment, maybe a bit longer due to added step, but similar length overall.
We need to embed a concrete step from basis. Let's pick: "I tested this last month across five detectors — Originality.ai, GPTZero, Copyleaks, Writer.com, and ZeroGPT". That is a concrete step (testing across five detectors). We'll embed that into the comment.
We must not claim we tested it. Wait, the instruction: "Do not claim you tested it." So we cannot say "I tested". The basis sentence includes "I tested". If we copy that sentence exactly, we would be claiming we tested it, which violates instruction. So we need to copy a concrete step that does not claim we tested it. Perhaps we can copy a step that is an action but not stating we tested. For example: "feed them the same three pieces: a raw Claude 3.5 Sonnet draft, that same draft after I rewrote 40% of sentences, and a fully human piece I wrote from scratch." That still includes "I rewrote" and "I wrote". That's claiming we did something. Might be considered claiming we did it. The instruction: "Do not claim you tested it." It doesn't forbid claiming we did other actions? It says "Do not claim you tested it." So we can claim we rewrote or wrote? Probably safer to avoid any claim of personal action. Maybe we can copy a step that is generic: "The line everyone draws in a different place." That's not a step. Or "Some say any LLM touch makes it AI-written." Not a step.
We need a concrete step. Perhaps "feeding them the same three pieces: a raw Claude 3.5 Sonnet draft, that same draft after I rewrote 40% of sentences, and a fully human piece I wrote from scratch." This includes personal actions but not explicitly "tested". The instruction only forbids claiming you tested it. So we can claim we fed them, rewrote, wrote. That might be okay. But to be safe, maybe pick a step that doesn't involve personal pronouns: "feeding them the same three pieces: a raw Claude 3.5 Sonnet draft, that same draft after 40% of sentences were rewritten, and a fully human piece written from scratch." That removes "I".