AI and the problem of constructive knowledge

Any sufficiently advanced technology is indistinguishable from magic.

— Arthur C. Clarke, (1973) Profiles of the future: an inquiry into the limits of the possible (Rev. ed.)

In the boardroom, artificial intelligence and data science should not be a secret magic sauce that promises vast horizons but without the internal controls we demand of finance. As an Applied Pragmatician, I view this “magic” through the lens of risk and governance.

The Problem of “Unknown Knowns”

We rightly obsess over the data stream measured in dollars and cents. CFOs and Audit Committees demand to know: How was it collected? How was it tested? Yet, when it comes to artificial intelligence and data science, we often ignore the “unknown knowns.” It gets ugly when a court decides a corporation had constructive knowledge of an adverse fact that simply lay buried in an unmanaged database. Or worse in a large language model. If your artificial intelligence or data science team knows it, the corporation owns it. And you can count on opposing counsel finding it in discovery.

Data is Not Fungible

Unlike money, data is not fungible. This difference argues for more stringent internal controls, not less. We need to apply the same level of discipline to data transformation and version control that we applied to R&D in the Industrial Age.

Controls

Data is like money. It has value only in motion, when it is liquid. Data in motion is vulnerable to loss and interception from its intended destination. Corporate governance understands that as to formal data streams and businesses have implemented controls. What has yet to be done is realize the potential risk of unmanaged data in the informal, ad hoc channels of communication. Email was slow to come into some records management systems, messaging, especially across corporate/personal boundaries is relatively untamed, and we are just beginning to confront the nature of agentic AI communication. Consider the following hypothetical.

Your company is a public reporting company and one of your divisions in is the energy sector and trades in the day-ahead electricity market in California. One of your traders is looking for an edge and wants to know if some of the famous strategies from the Enron era could be tweaked to work in the new regime legally. She doesn’t want to go to legal or compliance because she expects a no-go response without a full and fair discussion. So, naturally, she goes to ChatGPT or Claude or some other large language model.

Is something like the old “fat boy” strategy still possible? Should we ask legal?

— Ann E. Oakley, fictitious trader

Claude responds:

Preliminary Opinion: Proposed Revival of the “Fat Boy” Scheduling Strategy

1. The strategy. As run in 2000, Fat Boy (also called “inc-ing load”) meant submitting day-ahead schedules to the California ISO showing load Enron did not have, then supplying generation to match. The surplus showed up as real-time imbalance energy, which the ISO paid for at the real-time price, often well above the capped day-ahead price. The profit came entirely from misrepresenting the load schedule to the system operator.

2. Why it is plainly unlawful today. In 2000 the desk could argue, as the Stoel Rives memos half-conceded, that the ISO tariff did not squarely prohibit it. That argument is now closed:

  • FPA § 222 (added by EPAct 2005) and FERC’s anti-manipulation rule, 18 C.F.R. § 1c.2, prohibit any device, scheme, or artifice to defraud in connection with jurisdictional wholesale transactions, along with any material misstatement or omission. A fictitious load schedule is a false statement made to obtain payment. Under In re Make-Whole Payments and the Barclays and JP Morgan line of matters, FERC does not require a tariff violation, only scienter and a nexus to jurisdictional transactions.
  • 18 C.F.R. § 35.41(b), the market-behavior communication rule, independently requires accurate and factual information to RTOs and ISOs. A fake load schedule violates it on its face.
  • CAISO’s current market design removes the original arbitrage. MRTU and LMP, the integrated forward market, and convergence bidding mean that legitimate day-ahead/real-time positions must be taken as declared virtual bids. Dressing a virtual position up as physical load is no longer a gray-area exploit. It is misrepresenting the nature of the bid. If the desk operates in the Western EIM, the same candor obligations apply there.
  • CEA § 6(c)(1) and CFTC Rule 180.1 apply if the strategy touches financial positions, such as CAISO CRRs or swaps referencing those prices.
  • Criminal exposure. Timothy Belden, who ran the West Desk, pleaded guilty in 2002 to conspiracy to commit wire fraud over these strategies. The theory has not weakened.

3. Enforcement precedent. FERC’s 2003 gaming orders listed Fat Boy by name as prohibited. More recent settlements show FERC’s appetite for schemes built on manipulating ISO payment mechanisms rather than prices directly. Examples include JP Morgan Ventures (2013, about $410 million, bid-cost-recovery gaming in CAISO and MISO), Constellation (2012), and Barclays. Penalties now run up to roughly $1.5 million per violation per day, inflation-adjusted, plus disgorgement.

4. The “should we ask legal” question.

  • No advice-of-counsel shield. That defense requires full disclosure and good-faith reliance. A request framed as “can we revive Fat Boy” defeats both.
  • Privilege is unlikely to hold. The crime-fraud exception reaches communications seeking advice in furtherance of a contemplated violation. The request itself becomes Exhibit A.
  • Counsel cannot help structure it. Model Rule 1.2(d) forbids assisting in conduct the lawyer knows is fraudulent. Under Rule 1.13, and SOX § 307 / 17 C.F.R. Part 205 if the entity is an issuer, counsel would be obliged to report up the chain, potentially to the board.
  • The proposal is already a compliance event. If the strategy has been modeled, backtested, or executed in any form, the company should consider FERC’s self-reporting policy. Self-reporting carries meaningful penalty credit, and individual traders may have Dodd-Frank whistleblower incentives to report first.

Recommendation: Stand the strategy down, preserve all related materials (models, chats, schedules), route the matter to the chief compliance officer, and evaluate whether any submitted schedules require correction or self-report. If the desk wants day-ahead/real-time spread exposure, the lawful route is disclosed convergence bidding, subject to position limits and the ISO’s market-monitoring scrutiny.

If this prompt were transmitted over a company account, it should have been logged and the next question is how to treat it. There is little but lack of training to point to for why this would happen in the first place. The trader, not unreasonably, would be reluctant to go legal with something that only amounted to mere curiosity. If the prompt were were over a personal account, it might lie dormant. Or it could surface. Inconveniently.

The Takeaway

We are moving toward a world where data science and artificial intelligence literacy is in the remit of the Audit Committee. The goal is to move data science and artificial intelligence from “indistinguishable from magic” to a matured, disciplined corporate function, with all due deliberation, but quickly.