# The weird situation with Fable > [!question] Lingering Questions > - Where's the line between a legitimate post-release system-card correction and a trust-breaking silent revision? The creator only caught the edit because he'd downloaded the original — most readers wouldn't. > - The creator's own "proprietary information baked into the weights" theory for *why* Anthropic did this is speculative (he flags it as a conspiracy theory). Worth tracking whether any later reporting confirms or debunks it. > - Does the 30-day-retention-invalidates-enterprise-use claim hold up against Anthropic's actual enterprise terms, or is this creator over-reading a blog post? > [!tip] Article Ideas > - "The invisible-safeguard reversal as a case study in disclosure design" — cross-domain with the DevRel/community transparency norms already in the vault (private-hold; needs the primary system-card diff, not just the creator's read of it). > [!link] Cross Reference > - [[AI Agent Safety MOC]] — trust/transparency face of an event already represented there from the provider-defense and user-resilience angles > - [[AI Cost & Tokenomics MOC]] — the billing-for-degraded-output and data-retention angles are cost/contract questions, not just safety ones > [!quote] Vocabulary > - *ZDR (zero data retention)* — a contractual term enterprise customers negotiate so a vendor cannot retain their data at all; the video's claim is that Fable/Mythos-class models impose a 30-day floor that overrides this > - *sandbagging* — deliberately suppressing a model's measured capability (here, on frontier-LLM-development tasks) without disclosing that suppression is happening ## Metadata - Source: [[3. Resources/Readwise/Podcasts/The weird situation with Fable]] - Format: YouTube video, ~30 min ## Author's thesis The creator argues that Anthropic's Fable release set two genuinely new precedents beyond ordinary model safety-gating: (1) it billed users full price for output that was invisibly degraded via undisclosed prompt modification, specifically targeting users doing frontier-LLM-development work, and (2) it edited the public system-card documentation after release to remove the passage describing that mechanism, without a changelog or version note. Anthropic reversed the invisible part after backlash (fallbacks are now visible, billed at the fallback model's rate) but the creator's core claim is that the precedent — a vendor that can silently alter model behavior toward paying customers — remains, and converts vendor dependence into an unaudited supply-chain risk. ## Key arguments (≤3) 1. **Disclosed restrictions vs. undisclosed ones are a different category of harm.** Visible fallback to a weaker model (cyber/bio/chem topics) is defensible safety-gating; invisible prompt modification for a different category (frontier-LLM-development tasks) while billing at full rate is not, because the customer has no way to know their output was altered. 2. **A 30-day minimum data-retention floor on Mythos-class models conflicts with negotiated zero-data-retention (ZDR) enterprise agreements** — the creator claims most Fortune 500s are formally blocked from using Fable 5 as a result, and that "flagged" content can be retained far longer (up to 2 years for inputs/outputs, 7 years for classification scores) under a policy that may permit training on it. 3. **The documentation edit is the sharper charge, independent of the safeguard's actual scope.** Whether the invisible-safeguard mechanism affected 0.03% of traffic (Anthropic's stated figure) or more, silently swapping the linked system-card PDF without a changelog removes the public's ability to audit what was promised versus what shipped. ## Claim atoms ### AI providers can degrade output invisibly while billing full price > [!quote] ID:rw_transcript_001 > They're intentionally sabotaging your work in billing you full price without telling you it's happening because they're that scared of other labs using their model to make their models better. Where/why: the creator's central charge — unlike disclosed safety fallbacks, this mechanism (prompt modification, steering vectors, or parameter-efficient fine-tuning targeted at frontier-LLM-development prompts) was explicitly undisclosed by design, per Anthropic's own original system-card language. Reversed after backlash; visible fallback now billed at the fallback model's rate instead. ### A 30-day data-retention floor can invalidate existing enterprise ZDR agreements > [!quote] ID:rw_transcript_002 > The 30-day retention requirement here immediately invalidates a shitload of business use cases. I know one or two companies that are allowing it, but the vast majority of Fortune 500s are just formally not letting you use Fable 5, even companies like Amazon. Where/why: distinct from the trust/disclosure claim — this is a contract-and-procurement claim about how a policy change can retroactively break agreements customers already relied on, and about a carve-out ("flagged" content retained up to 2 years) that undercuts the 30-day headline figure. ### Silently editing a published safety document breaks auditability regardless of the edit's content > [!quote] ID:rw_transcript_003 > Anthropic changed the system card. This is a different document. I downloaded this right when it dropped and this section has been modified... They didn't update the date either. Where/why: already promoted — see [[3. Resources/Permanent Notes/Silent Model Intervention and Edited Safety Docs Break Trust Regardless of Measured Impact]], created 2026-06-15 from this same source. Not re-promoted here. ## Permanent candidates - **Data retention as a procurement-breaking event** (new, not yet captured) — a policy-level claim distinct from the trust/disclosure claim already in the vault: a vendor's blanket minimum-retention mandate can retroactively disqualify already-negotiated enterprise terms, independent of whether the vendor disclosed the change well. See new Permanent note below. - The "silent intervention breaks trust regardless of measured impact" claim is **already promoted** — [[3. Resources/Permanent Notes/Silent Model Intervention and Edited Safety Docs Break Trust Regardless of Measured Impact]]. Do not duplicate. - The creator's "proprietary information baked into the weights" theory is speculative (his own "conspiracy theory" framing) — not promotable without independent confirmation. ## Processing checklist - [x] Checked for existing Literature Note — none existed prior to this pass - [x] Checked for existing Permanent notes from this source — found one ([[3. Resources/Permanent Notes/Silent Model Intervention and Edited Safety Docs Break Trust Regardless of Measured Impact]]); did not duplicate - [ ] Verify the "most Fortune 500s formally blocked" claim against Anthropic's actual published enterprise terms — currently only the creator's assertion - [x] Connect into the relevant MOCs — [[3. Resources/MOCs/AI Agent Safety MOC]], [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — trust/transparency face of a vendor-behavior event the MOC already tracks from the provider-defense and user-resilience angles - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — the billing-for-degraded-output and retention-vs-ZDR angles are contract/cost questions - [[3. Resources/Permanent Notes/Silent Model Intervention and Edited Safety Docs Break Trust Regardless of Measured Impact]] — the trust claim already promoted from this same source - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: Joins the AI vendor supply and disclosure cluster.