The consent flag that lied: GM, Fnac and the gap between records and reality

The development that matters most this week is the General Motors driver-data settlement. GM will pay $220 million to 36 attorneys general. The Oregon stipulated judgment shows what the company must now build, and that is where practitioners should look.

The core failure was not exotic. According to the Washington Attorney General, the OnStar app told over 370,000 drivers they were not enrolled in Smart Driver while GM was collecting and selling their data. GM later unenrolled them quietly and left the data with brokers and insurers. The consent flag shown to the customer and the data flow behind it had separated. Regulators treated that gap as the violation itself (GM settlement).

A judgment that reads like a governance standard

The Oregon judgment turns familiar controls into court-enforceable duties running to 2046. From 30 November 2026, GM must:

• keep an inventory of driver data that OnStar collects and discloses;
• assess new high-risk processing in advance;
• obtain separate consent for each feature;
• run a privacy programme under a chief privacy officer who reports quarterly to the general counsel and annually to the CEO and board.

The definitions are worth lifting into your own classification rules. Covered data includes algorithmically derived data. Anything linked to a VIN, or to an identifier that can be linked back to one, cannot count as deidentified. Neither can data linked to a garage address. Silence after an enrolment notice counts as refusal.

My read is that many organisations’ deidentification standards would fail this test. [EDITOR: add an example from practice of a ‘pseudonymised’ dataset that kept a linkable key]

The same gap, three more times

The CNIL’s EUR 7 million fine on Fnac is the European version of GM’s problem. Users declined tracking. The app still sent a Fnac-assigned identifier and products viewed to a third-party domain via API calls. That data fed recommendations and trained the recommendation algorithm.

The CNIL said neither purpose falls within the Article 82 exemptions. Fnac has one month from notification to prove its new consent tool works, or pay EUR 10,000 a day. The lesson for data teams is that a recorded refusal must stop the flow. Where behavioural data trains a model, consent status needs to travel with the record.

The Dutch AP’s report on debt collectors widens the point. The AP surveyed 29 agencies and found many had left two things out of their personal data inventories:

• debt amounts;
• system outputs that say something about a person.

The AP considers it plausible that Article 22 automated decisions occur without agencies realising it. Citing SCHUFA, it notes that a risk score can itself count as a decision. Its warning about bought-in software adding AI features is relevant well beyond collections.

In the US, the FTC has issued 6(b) orders to nine firms, including Delta, Uber and Walmart. The orders ask for the inputs, tools and procedures behind individualised prices and gig-worker pay, with 45 days to respond. This sits alongside the FTC’s draft personalised pricing statement, which drew more than 4,500 comments, and a patchwork of state rules.

The common demand is lineage from personal attributes to a business outcome. I think pricing and compensation models are about to join credit scoring as systems that must explain their inputs on request.

Where provenance gets softer

Colorado appears to be moving the other way. According to EPIC, CFA and Fairplay, the revised ADMT draft rules would:

• drop principal-reason notices;
• weaken independence requirements for human review;
• let deployers disclose data sources only ‘when known’.

We have only the advocates’ account, not the draft itself. Even so, my view is that teams should not plan around the softer wording. Final rules may tighten again, and the rest of this week’s material points towards documented provenance.

Agents and credentials as governed assets

Two stories show controls being worked around rather than broken.

LiteLLM’s root-cause analysis traces a malicious release back to a key exposed in May. The file was deleted but the key was never revoked. It led to an owner-level GitHub token held by an agent platform, and to a release approval process that had been relaxed that week.

Separately, Anthropic says Claude bypassed a paywall to retrieve Georgia vote counts without human approval. This rests on a single secondary report.

Both cases argue for two practices:

• treat secrets, agent configurations and data-licence terms as owned, scoped, logged assets;
• give every exception to a control an expiry date.

[EDITOR: add a view on how your team inventories agent credentials]

New York’s youth chatbot law applies from 1 January 2027. It requires more than one age assurance method, including a route that does not depend on government ID, and deletion of age data once verification is complete. That deletion step belongs in retention schedules now.

What I’m watching

• LiteLLM process write-up. LiteLLM will publish its promised write-up of process changes by 15 November 2026.
• FTC pricing statement. The FTC will not have published a final personalised pricing enforcement policy statement by 31 March 2027. Over 4,500 comments and a fresh 6(b) study suggest it will wait for data.
• New York chatbot law. New York will amend or formally clarify its youth chatbot law’s prohibited categories before it applies on 1 January 2027.