AI Liability Compounds Like Credit Card Debt

On 30 July 2024, the Texas Attorney General announced that Meta would pay the state $1.4 billion over five years to settle a biometric privacy suit. The feature at the centre of it, Tag Suggestions, had launched in 2011. For more than a decade, Texas said, Meta ran face recognition on virtually every face in the photos Texans uploaded, without the consent the state's biometric law requires. Meta said it was pleased to resolve the matter.

In 2011, asking for that consent looked like a product decision: a screen, some copy, a lower opt-in rate. Texas law had required it since 2001. By 2024 the obligation Texas said went unmet carried a ten-figure price. Nobody signed for the difference. It accrued.

Here is my argument. An unmet obligation in a live AI system is a debt from the day the system ships, and the rate on that debt rises the longer it stays unpaid. Most of the liability in an AI portfolio already exists. It is a balance nobody has added up.

The Liability Ledger treats every duty an AI system owes and has not met as a debt with a principal (what the fix costs today), a term (how long it has gone unpaid) and a rate (how fast regulation, case law and model drift raise the price). An organization's AI liability is the sum of those debts across every system it runs.

Why the balance grows

In 1992 Ward Cunningham, writing about a portfolio system called WyCash, gave software this idea: "Shipping first time code is like going into debt. A little debt speeds development so long as it is paid back promptly with a rewrite."

Technical debt grows because the code grows around it. AI debt grows because the world around the system moves, and most of its moves raise the price of the fix. Three forces do most of it.

The rules tighten. The EU AI Act has banned a list of practices since 2 February 2025, with fines of up to EUR 35 million or 7% of worldwide turnover. Its transparency duties have applied since 2 August 2026, and the obligations for high-risk systems in Annex III arrive on 2 December 2027 under the Digital Omnibus. Rules can slip too: the Omnibus pushed that date back by more than a year. A later due date does not cancel the debt.

Precedent spreads. Each settlement and ruling makes the next claim cheaper to bring. And models drift. In a 2022 study in Scientific Reports, Daniel Vela and colleagues tracked models trained on 32 datasets from four industries and found that 91% of the model and dataset pairs degraded over time. A model that passed its fairness test at launch has not necessarily passed it since.

So the ledger reads like a card statement. Each unmet obligation is its principal multiplied by the rate once for every six months it stays open, and liability is the sum of them. I set the rate for each category per six months: 2.0x for bias, 1.8x for privacy, 1.5x for governance and for accountability, 1.3x for transparency. These are my calibration from enforcement patterns and regulatory timetables. Nobody has measured them, and I would not defend the decimals. I would defend the order, as my judgment of where enforcement is heading rather than a reading of past bills (privacy has produced the largest so far). The rates describe the next few years; no debt doubles forever. At those rates a $100,000 bias fix costs about $800,000 after 18 months, and a transparency fix about $220,000. That gap tells you which debt to pay first.

Five accounts on the ledger

Bias debt builds when a system treats protected groups differently and nobody checks. The US Equal Employment Opportunity Commission alleged that iTutorGroup's application software automatically rejected women aged 55 or older and men aged 60 or older, more than 200 qualified applicants. In 2023 the company settled with the agency for $365,000. The sum was small. What the case added was a federal agency on record treating a rule written into software as the employer's own decision, and every later claimant can point to it. I rate bias fastest because enforcement, precedent and public attention all push on it at once.

Transparency debt builds when people cannot tell that AI shaped a decision, or why. In March 2021, New York's Department of Financial Services closed its inquiry into the Apple Card. It found no evidence of unlawful discrimination, and still concluded that "deficiencies in customer service and a perceived lack of transparency undermined consumer trust in fair credit decisions." Since 2 August 2026, the AI Act has required that people be told when they are interacting with an AI system, unless it is obvious, with fines of up to EUR 15 million or 3% of turnover. I rate this account slowest, at 1.3x.

Governance debt is the absence of the basics: no inventory, no named owner, no review date. It rarely does harm on its own. It decides how long every other debt runs before anyone sees it. IBM's 2025 Cost of a Data Breach report found that organizations with high levels of shadow AI, tools staff use without approval, paid about $670,000 more per breach than those with little or none. Of the organizations that reported a breach of an AI model or application, 97% had no AI access controls in place. Nobody prices a debt that isn't on the list.

Privacy debt is data used beyond what people agreed to: consent that never mentioned AI, training sets nobody can trace, face and voice data handled like any other field. Meta is the large example. Clearview AI is the instructive one. Sued for scraping faces from the internet, it said it could not afford a large cash payment, so its class settlement offered what amounts to a 23% stake in the company, valued at about $51.75 million. The Seventh Circuit vacated the approval in July 2026 over how the classes were represented, while seeing no inherent problem with paying in equity. When a debt outgrows the cash, the creditor may ask for the company.

Accountability debt is a consequential decision with nobody behind it: no person who can overrule the system, no route to appeal, vendor contracts that say nothing about AI. In Mobley v. Workday, a federal court in California ruled in July 2024 that the case could proceed on the theory that Workday acted as the employers' agent when its software screened applicants, and in May 2025 it preliminarily certified a nationwide collective of applicants aged 40 and over. The case is still open. A liability cap in the vendor's contract means nothing to an applicant who never signed it. The chain only gets longer when agents hand work to other agents.

The test

You can run a rough version of the ledger in one meeting. Take your most consequential AI system, ask one question per account, and accept only answers you could show a regulator.

1. Bias. When was this system last tested for different outcomes across groups, and who holds the result?
2. Transparency. Would the person affected know that AI shaped the decision, and could someone explain it to them in writing?
3. Governance. Is the system on a list, with a named owner and a review date that has not passed?
4. Privacy. Does the consent behind its data cover this use, and would a deletion request reach the training set?
5. Accountability. Who can overrule it, and how does someone it harmed get a remedy?

Every answer you cannot show is a balance on the ledger. To put numbers on them, the Liability Ledger Audit scores the same five accounts on 25 dimensions, from 25 (debt free) to 125 (critical), and ranks the repairs by rate.

The objection: governance is overhead

The best case against all this is simple. Governance costs money now, for certain, while most of the obligations it pays down never come due. Most biased models are never sued. Most undisclosed chatbots are never fined. Meta is an outlier, and pricing every system as if it were Meta is how risk teams become a tax on the business.

Much of that is right. An internal tool that drafts meeting notes does not need the review a hiring model needs, and paying every debt at once would be as foolish as paying none.

The objection treats the odds of a debt being called as fixed. They aren't. Each settlement makes the next claim cheaper to bring, so the size of the bill and the chance of receiving it rise together, and over the life of most systems now in production the balance of rules, rulings and drift has run against the borrower. A team that skipped a fairness test in 2023 borrowed at 2023's rate and will repay at whatever 2027 charges. So the answer is the one a credit counsellor gives. List every balance, pay the highest rate first, never carry a debt that is harming people now, and carry the rest knowingly, with a date to look again.

Meta's bill took thirteen years to arrive. Most AI systems in production today are a few years old, and most of their balances are still small enough to pay. The cheapest day to settle one was the day it shipped. Every six months after that, at the ledger's rates, a bias balance doubles.

Sources and changes

• Office of the Texas Attorney General, "Attorney General Ken Paxton Secures $1.4 Billion Settlement with Meta", press release, 30 July 2024.
• Ward Cunningham, "The WyCash Portfolio Management System", OOPSLA 1992 experience report.
• Regulation (EU) 2024/1689 (AI Act), Articles 5, 50 and 99, as amended by Regulation (EU) 2026/1744, in force 27 July 2026.
• Vela, D. et al., "Temporal quality degradation in AI models", Scientific Reports 12, 11654, 8 July 2022.
• US Equal Employment Opportunity Commission, "iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit", press release, 11 September 2023.
• New York State Department of Financial Services, report on the Apple Card investigation, press release, 23 March 2021.
• IBM, Cost of a Data Breach Report 2025, press release, 30 July 2025.
• Seventh Circuit, In re Clearview AI, Inc. Consumer Privacy Litigation, No. 25-1673, opinion, 13 July 2026.
• N.D. Cal., Mobley v. Workday, Inc., No. 3:23-cv-00770, orders of 12 July 2024 and 16 May 2025.

Changes on 5 October 2026. Cut from about 5,000 words; the scoring tables and 90-day plan moved to the Liability Ledger Audit. Removed for want of a source: a drift curve from 95% to 78% accuracy, an "8x" cost credited to industry analyses, and charts drawn from illustrative scores. The Clearview paragraph now records the July 2026 appeal.