Waymo 高速公路召回與 AI 虛構引文裁決分析:聲稱更新已修復並非證據
"Fixed in the update" is a claim, not evidence: notes on Waymo's freeway recalls and an AI-citation ruling
VSO 核對 Waymo 的兩項高速公路召回及 Kansas Court of Appeals 對 AI 虛構引文的裁決,指出溯源記錄不能證明修正有效或 AI 生成內容真確。
Two things happened on 20 November 2026 that look unrelated. Waymo announced two recalls after its robotaxis drove into closed freeway lanes. And the Kansas Court of Appeals published an opinion about a self-represented litigant whose AI-drafted filings cited cases that do not exist.
A monitoring brief sent to VSO, which maintains the VAP provenance framework, bundled them together and presented provenance logging as the fix for both. We checked the claims against the primary sources and registered both cases (V1111WAYMOFREEWAY and V1120LINGLE). The brief was wrong in a way that is useful for anyone who builds systems that log decisions. Here is what the sources say, and what logging can and cannot do in each case.
Case 1: a remedy that has to be shown to work
The timeline matters, and the brief had it backwards.
- April–May 2026: 13 events in which Waymo vehicles entered freeway construction zones in Phoenix and the Bay Area. Recall 26E035 was filed in June (3,871 vehicles). Its NHTSA report attributes the defect to the system "inappropriately prioritizing the avoidance of other freeway hazards" or not recognising the zone. The software update began on 16 June, and freeway operations began to resume on 25 June.
- 5 November: three vehicles entered a construction closure on US 101.
- 6 November: according to AP, a vehicle tried to pass a highway patrol traffic break on I-280.
- 11 November: the California DMV restricted Waymo's driverless freeway operations. It announced this on 17 November, calling the operations an "imminent hazard to public safety" and noting that similar May incidents had recurred despite earlier corrective actions.
- 19 November: Waymo filed two new recalls (closed lanes, with the remedy still in development; manual traffic control, with the update done on 6 November). It announced them on the 20th.
The DMV's condition for lifting the restriction is that Waymo "demonstrates" its corrective actions are effective. That is the engineering problem worth noticing. "Fixed in version X" is a claim. Turning it into evidence takes, at minimum:
- Version-to-event binding. For every relevant event after the remedy date, which software build was running on that vehicle.
- Post-remedy event accounting. How many relevant events occurred after deployment, and whether the set you are showing is the complete set you recorded.
- Independent observations. Here, a CHP traffic break and a viral video. These are outside the operator's records, and no operator log replaces them.
A tamper-evident, externally anchored log helps with the second point in a specific, limited way. A third party can check that anchored records were not altered afterwards, and can detect events omitted after anchoring at batch granularity within a declared scope. In VAP v1.2 that is the Completeness Invariant, INT-008. It cannot tell you whether the fix was adequate. It cannot recover an event the vehicle never recorded. And it does not cover events that were never put into an anchored batch at all.
There is also a scope fact: VAP's automotive profile, DVP, is listed as "Planned" in the v1.2 registry, which means no specification document exists. Since a deployment must sit under exactly one domain profile (§4.6.3, rule 1), nobody can claim VAP conformance for a robotaxi fleet today.
Case 2: provenance is not correctness
In Lingle v. Subaru of Wichita (No. 129,581), the district court had warned the litigant several times to verify sources in motions it called "admittedly Artificial Intelligence-generated." It then sanctioned him in several separate awards. The brief said the appellate court imposed about $1,480 in sanctions. That was one of the district court's awards. The appellate court imposed no new sanction. It checked the appellate brief itself and found cases and statutory subsections that do not exist, plus 36 misstatements of 11 real holdings. It disregarded the fabricated authority, affirmed, and warned future filers. Its syllabus: a self-represented litigant has the same duty as an attorney to verify AI output.
If you build LLM features, the lesson is old but keeps getting relearned. A perfect record of how text was generated tells you nothing about whether the text is true. You could sign every prompt and every token, and a fabricated case citation would still be fabricated, just with a better audit trail. The check that caught these errors was a lookup against an authoritative source: the court reading each citation against the reporters and statutes. If your product emits citations, the verification step belongs in the pipeline, before the output reaches a user or a court.
VAP's own normative text says this. Conformance "does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records" (§1.6(c)).
What the brief claimed, and what the spec says
- An audit trail that "cannot be altered": VAP records are tamper-evident, which means alteration is detectable, not impossible.
- Action "before the shutdown": VAP is post-hoc only and has no role while a system runs.
- "Cryptographic hash chain": v1.2 replaced the hash-chain requirement with the more general Cryptographic Sequence Verifiability.
- "INT-008 shows which AI produced which citation from which input": INT-008 detects post-anchor omission and split-view. It establishes nothing about causation.
- "DVP-equivalent" and "CAP-equivalent" checks: DVP has no specification. CAP is the Content / Creative AI Profile.
- EU AI Act and GDPR obligations: both cases are US matters with no EU connection in any source.
Disclosures
VSO develops VAP. Zero external implementations of VAP and zero Evidence Packs accepted in any proceeding (disclosure reviewed 11 November 2026; count dated 25 October 2026). A contract figure published earlier was withdrawn on 24 October 2026; the withdrawal does not mean the count is zero, and no contract, customer or revenue figure is asserted. VAP v1.2.0 is a draft with no standing in any established standards body. This post was researched and drafted with Claude (Anthropic), which is not a party to either case.
Full analysis on Medium: https://medium.com/@veritaschain/two-questions-a-record-cannot-answer-4d3b4ed5ce99
Sources: Reuters via syndication · TechCrunch · AP via KPBS · NHTSA 26E035 · Lingle v. Subaru of Wichita · VSO incident registry · VAP v1.2 spec
來源:DEV Community · dev.to