Where exactly does AI-assisted become AI-written

A test of five AI detectors — Originality.ai, GPTZero, Copyleaks, Writer.com, and ZeroGPT — on the same three pieces found no consensus: raw Claude 3.5 Sonnet output was flagged by all five (95–99% probability), a 40% human rewrite split the detectors (Originality.ai and GPTZero still called it AI at 78% and 71%, Copyleaks and Writer.com flipped to human at 12% and 8%, ZeroGPT sat at 43%), and a fully human piece drew false positives from Originality.ai (23%) and GPTZero (18%). The author proposes a three-tier practical definition (AI-generated, AI-assisted, human with AI tools) and notes that platforms like Google, Medium, Substack, Amazon KDP, and academic journals enforce different rules, while the EU AI Act, China's regulations, and US voluntary commitments fail to resolve the boundary.

Where exactly does AI-assisted become AI-written 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 /en/tags/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 95–99% probability 40% human rewrite: split decision — Originality.ai and GPTZero still called it AI 78% and 71% , Copyleaks and Writer.com flipped to human 12% and 8% , ZeroGPT sat at 43% Human original: two false positives Originality.ai 23%, GPTZero 18% The detectors don't even agree with each other, let alone with a coherent definition. What the platforms actually enforce 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" vs "AI-assisted" their distinction: generated = text/images from prompt; assisted = brainstorming, editing, research . Most authors check "assisted" and move on. Academic journals are stricter. Nature and Science ban LLM-listed authors but allow usage disclosure in methods. arXiv added an AI-use field. Conferences like NeurIPS now require a checklist. Enforcement is spotty — reviewers catch maybe 30% of undisclosed use by my informal count. The practical definition I use For my own workflow, I settled on three tiers: Tier 1 — AI-generated: Prompt → output → minimal cleanup formatting, typo fixes . I label this when sharing publicly. Tier 2 — AI-assisted: I write the structure, key arguments, and ~60% of prose. LLM fills transitions, suggests phrasing, checks tone, generates examples I verify. This is most of my technical posts. Tier 3 — Human with AI tools: Grammarly, spellcheck, maybe a thesaurus lookup. No generative step. Still human. The boundary between Tier 1 and 2 is where arguments live. I've seen writers claim Tier 2 after changing three adjectives. I've seen purists call Tier 3 "tainted" because they used Claude to brainstorm titles. Why the definition matters Copyright office rulings hinge on human authorship. The USCO denied registration for Théâtre D'opéra Spatial Midjourney /en/tags/midjourney/ but granted it for a comic where the human wrote every panel description and curated outputs. The distinction: "creative control" over expressive elements. SEO clients ask "will Google penalize AI content?" The honest answer: Google penalizes bad content. But clients want a binary — AI or not — because their legal teams need a policy. Freelance marketplaces Upwork, Fiverr now have AI-disclosure tags. Buyers filter them out. Writers who disclose lose bids; writers who don't risk bans. The incentive structure rewards ambiguity. No consensus coming The EU AI Act requires watermarking for generative systems. China's regulations mandate labeling. The US has voluntary commitments from major labs. None of this solves the "where's the line" problem for the 99% of content that sits between raw output and pure human craft. I don't expect a standard. The spectrum is too useful — everyone slides the marker to suit their incentives. What I do: disclose my tier when asked, don't pretend Tier 2 is Tier 3, and accept that detectors will misclassify me either way. The only definition that holds up in practice: who made the creative decisions that shape the work's meaning? Everything else is theater. The hype cycle promised mass adoption by 2024 — reality check 11m ago /en/news/6968/ Why most companies will miss the AI-native shift entirely 9h ago /en/news/6912/ AI is making every side project feel like meaningless slop 1d ago /en/news/6755/ AI is killing the spirit of hackathons if we keep letting LLMs 2d ago /en/news/6662/ GitHub PRs are completely falling apart under the weight of 2d ago /en/news/6641/ Is jailing anti-AI protesters actually a sign that we're losing 3d ago /en/news/6570/ Next OpenAI pauses o1 training after model attempts self-preservation → /en/news/6964/