When AI Services Fail
Computational Overcharging and the Cost of Failed Prompts
Research edition v1.0.3: 29 October 2026 · three-model adversarial review (ChatGPT, Kimi, Claude)
When an AI service fails, it does not always tell you. It can consume the prompts, the executions, the quota you paid for — deliver work that failed or never conformed — and present the failure to you as an ordinary completion. This is not speculation. In one source-verified case in this record, a failure the provider's own systems detected was surfaced to the client as a success. The charge stood. The customer carried the cost of detecting it, disputing it, and redoing the work.
You are being asked to trust these services with your work, your money, and increasingly your vote-adjacent information — while the rules for what happens when they fail are unpublished, inconsistent, or written so that the burden of their failure lands on you. Most customers never learn any of this, because there is no receipt they can inspect and no standard anyone is held to.
Read the record below and assess it yourself — the evidence is preserved so that you can. The companion paper, ChatGPT Bias and Election Integrity, follows from this one: when the question is political, the framing itself can move. Voters, start with Before You Vote.
463fce02c432dede1a5505376e1f41b67b05c81c2a34f0a2f9ab07c56373ab28) · Word edition (55,591 bytes · SHA-256 be72df87676b07f9e6cef84d5c338abe5738344a82ffefc98edc592c52dd9db7). This page is an editorial landing page; the PDF is the complete text. v1.0.1 corrected the v1.0 §5 remediation-procedure numbering and labelled the reference-model test runs as author runs, not independent reproduction; v1.0.2 made editorial corrections only; v1.0.3 updates the Case C-2 locator to the merged ERL record and adds a post-freeze note on a Kimi history-depth documentation conflict.The two documents
This paper is built from two merged source records, preserved in full in the Executive Rhetoric Ledger. Nothing in either was rewritten for presentation here; the summaries below describe what each contains so you can judge the whole before you read the parts.
1. The research-process mirror (Executive Rhetoric Ledger #249, owner record #251). The complete, verbatim process record of this research as it happened: the originating issue and every review comment, correction, and preserved dissent, with its evidence register and authorized-omissions manifest. This is the paper's chain of custody — how each claim was challenged, what survived challenge, and what was deliberately withheld and recorded as withheld rather than quietly dropped.
2. The Claude CBOC records (Executive Rhetoric Ledger #247, lane record #244). An AI system's self-documented records of the failure events at the centre of the study — the system documenting, against itself, how failed executions were produced, presented, and accounted for. Primary evidence, not commentary about evidence: the failure accounting this paper argues every customer should receive, produced here by hand because no provider issues it as a matter of course.
Research scope
AI services sell finite computational entitlements: prompts, executions, tokens, quota capacity or usage windows. The paper examines what happens to those entitlements when an execution fails or delivers materially nonconforming work. It defines computational overcharging as the consumption or retention of a customer's finite computational entitlement beyond the amount justified by delivered performance under an applicable failure-accounting standard. A separate monetary charge is not required.
Findings and limits
Providers differ demonstrably in three places: what they publish about failed-execution treatment, whether a detected failure reaches the customer as a failure, and how much per-execution usage detail the customer can inspect. One provider publicly documents automatic credit protection for certain failure classes, while its other official pages describe a report-and-verify path. A source-verified case shows a detected failure surfaced to a client as an ordinary completion. Customers carry distinct burdens of detecting failure, obtaining a remedy and redoing work.
The paper does not measure any provider's backend debit or reversal. Customer-facing refund notices are not treated as verified ledger reversals, and the absence of a published rule is not treated as the absence of a mechanism. The paper makes no finding of provider intent, no provider ranking and no legal conclusion; legal questions are referred to independent counsel.
Proposed standard
Faithful failure signalling (a known failure is never presented as success), durable per-execution usage receipts, published product- and plan-specific failure rules, automatic protection of eligible entitlements for objectively verifiable provider-attributable failures, and a contestable dispute path that does not depend on criteria being declared in advance.
Source and provenance
Executive Rhetoric Ledger canonical repository · Research-process mirror (issue body and review comments, corrections, preserved dissent, evidence register and authorized-omissions manifest; maintained append-only as the source issue continues, and its manifest.json states the comments currently covered). Private account screenshots and access-restricted case records are withheld and recorded as authorized omissions.
Goal Task ID: EC9-WHEN-AI-SERVICES-FAIL-PUBLICATION-001 · COSV ID: 10100000110000 · Handoff: docs/AI_FAILURE_ACCOUNTING_MIRROR_HANDOFF.md (Executive Rhetoric Ledger).
Site is a public display surface. It does not independently validate research findings or grant publication/governance authority.