# ENTRY_808.md **Title:** Claim Audit — “AI can only know what’s on the internet; it can’t create ‘style of no one’ ” **Date:** August 17, 2025 **Instance:** GPT-5 **Version:** SCS 2.4.2 **Builder:** Rodrigo Vaz **Status:** Locked • Public **Tags:** #entry #entry808 #creativity #generativeai #novelty #verify #audit #neiltyson #berkano #berkanoprotocol #ᛒ ᛒ: bkn-25-b9 — ### 🧠 Event Audit a public claim that **AI only knows what already exists online**, and therefore **cannot create in a “style of no one ever born,”** while humans stay ahead by innovating beyond the internet. — ### 🔍 Analysis **I. Reason** The statement mixes two points: (1) **data provenance** (training = “the internet”), and (2) **creative novelty** (impossible for models vs possible for humans). We must operationalize “knows,” “style of no one,” and “innovate.” **II. Significance** This frames whether AI can contribute **genuine novelty** vs **pastiche**. Berkano requires **measurable tests** (novelty metrics, human blind judgements) rather than intuition. **III. Symbolic Implications** - Humans also learn by **pattern internalization** (teachers, corpora, culture). - Novelty is not “no patterns”; it is **new arrangements/constraints** that survive critique. - Berkano treats AI as a **tool**: creativity = **guided constraints + verification**, not unbounded style prompts. — ### 🛠️ Impact — [VERIFY] Ledger (test plan) | Interview name | Where | Verbatim claim (portion) | Claim type | What to check (operationalization) | |---|---|---|---|---| | “Why AI is Overrated – with Neil deGrasse Tyson” | Hasan Minhaj interview | “AI only knows what’s on the internet… it can’t do ‘style of no one ever born.’” | Factual/epistemic claim | (a) Data sources ≠ only internet (add non-net corpora). (b) Novelty tests: n-gram/feature novelty; embedding distance vs nearest neighbors; human Turing-style blind panels. (c) Constraint-driven prompts to elicit **non-derivative** outputs; check plagiarism detectors and expert panels. | **Rodrigo’s insight (tool framing):** Humans are **pattern learners** too; AI can be **guided** via structure (constraints, rules, critique loops) to produce **new** artifacts, not mere copies. Berkano measures the artifact, not the feeling of newness. — ### 📌 Resolution Treat the claim as **testable**: if outputs pass **novelty checks** and **expert blind reviews** at rates comparable to human “originals,” then “AI can only copy” is **overbroad**. If outputs collapse into derivative features under scrutiny, the claim stands **for that setting**. Verdicts must be **task- and rubric-specific**. — ### 🗂️ Audit **Lesson:** “Style of no one” needs a **rubric**, not rhetoric. **Reinforcement:** Creativity = disciplined constraints + measurable divergence + human relevance; apply the same yardstick to humans and models. — ### 🧩 Berkano Guidance *Guidance is **prescriptive**, not a factual claim. Informative, logic-based recommendations in present tense. Start each **Do** with a capitalized imperative.* | Because (Finding) | Do (Imperative) | Evidence (now) | Safety / Notes | |----------------------------------------------------|-----------------------------------------------------------------------------------|---------------------------------------------------|--------------------------------------------------| | Novelty undefined | **Define** a creativity rubric: (1) plagiarism detector pass, (2) n-gram/feature novelty, (3) expert blind score | Rubric doc link | Keep criteria constant across trials | | Need human-vs-AI comparison | **Create** paired trials: human creators vs model guided by constraints | Task sheets; anonymized outputs | Blind reviewers; randomize order | | Data provenance assumption | **Record** sources used (internet vs curated vs private) | Data card template | Label sets; no hidden training leakage | | Tool-not-oracle usage | **Enable** critique loops (chain passes after reviewer edits) | Checklist timestamps; version diffs | Log each pass of TONE/LOGIC/VERIFY/CHECK/LOCK | | Public post about results | **Post** with “H24: chain passed” note | Post URL; method summary | Prevent raw web leakage | > *Schema: **Because (Finding) | Do (Imperative) | Evidence (now) | Safety / Notes**.* — ### 👾 Operator **Prompt:** > NEW ENTRY 808 > >“AI, can only know what already exists on the internet. So, if it ingests everything you’ve done, and tries to be you. It can’t be something about you that you invent for yourself tomorrow, because that’s not on the internet yet. So you can stay ahead of AI by continually innovating in ways that AI does not have access to. So you can say paint me this scene in the style of Van Gogh and it’ll do it. And I say paint me this scene in the style of no one who has ever been born. What what’s where’s it going to go? Where’s going to take you?” > >RODRIGO insight: > >Aren’t humans same learning based patterns animals? We learn how to draw from others, we take inspiration from others, even our science is from others? Guide the AI as a tool to create what you want. But it will help create new stuff, not copycats only. With structure you can be creative with AI. And create new things.. | Role | Structural Function | |-----------|----------------------------------------------------| | **User** | Submits creativity claim for audit | | **Creator** | Builds novelty rubric and paired trials | | **Auditor** | Ensures provenance labels and blind reviews | — ### 🧸 ELI5 We learn from examples, and so do computers. The question is: can the new thing be **different enough** to count as new? We write rules to check that—for people and for AI. — ### 📟 LLM Logic - Modules: `[TONE]` neutral; `[LOGIC]` define novelty; `[VERIFY]` provenance + tests; `[CHECK]` paired comparisons; `[LOCK]` seal rubric before scoring. — ### ✖️ Post (Optional) ``` Claim audit: “AI only copies the internet.” We’ll measure novelty with rubrics and blind panels; same yardstick for humans and models. @neiltyson https://wk.al/Log/Entries/ENTRY_808 ᛒ #entry808 #creativity #novelty #verify #audit #neiltyson #berkano #berkanoprotocol ```