Oil Reserves and AI: A New Risk for Investors

Oil companies use numbers called “proved reserves” to show how much oil they own. Those numbers help set the company’s value and decide how many wells to drill in the next five years. A new study shows that many firms are now using computer models, especially artificial intelligence (AI), to create those numbers. The old rules that let investors trust the reserves were written long before AI existed.

The company Permian Resources said in its latest report that using AI is a risk. It also repeated the old warning that reserve estimates are not exact. The report did not say Permian books reserves with AI; the 2025 reserves were made by an outside firm, Netherland Sewell & Associates. The bigger issue is that the whole oil industry is moving to AI models, but the rules for valuing reserves have not changed.

Traditionally, engineers drew a curve from past oil production and used it to predict future output. That method is easy to check because anyone can see the math. It works well for simple reservoirs but fails when oil moves in complex ways, like in shale formations. AI models can look at more data and find hidden patterns, which makes their predictions more accurate.

However, AI models are hard to explain. A decline curve shows how the estimate was made; an AI model does not reveal which inputs mattered most. Engineers know this problem and are working on “explainable AI” that ties the math back to physics. Still, a model that changes its output when retrained is hard to defend as a reliable reserve estimate.

The SEC’s rule that defines a proved reserve says it must be “reasonably certain to be economically producible” using deterministic or probabilistic methods. The rule does not mention AI, but it allows “reliable technology” that has been field‑tested. The 2008 update to the rule was meant to let companies use better tools, but it did not foresee AI’s lack of repeatability.

When a company reports new reserves, it must give a short summary of the technology used. This can say “proprietary machine‑learning model” without giving details, which protects trade secrets. Investors must decide whether that description gives enough confidence.

Outside auditors like Ryder Scott or DeGolyer & MacNaughton check the reserve numbers. They compare the model’s forecast to actual well data and judge whether the estimate is reasonable. Their certification does not depend on which software was used; it only looks at how well the forecast matches reality. This works for traditional curves but may not cover AI’s hidden logic.

The risk is real. Shareholders rely on a number that comes from a tool they cannot fully see. Banks lend money based on these reserves, and auditors face legal pressure when a reserve is later found to be wrong. The SEC has already asked companies to explain the technology behind their reserves, and future letters may push for clearer AI disclosures.

Investors should read the reserve number as an estimate that comes from a method. Look for the company’s explanation of its technology and how outside auditors have verified it. The industry must update its rules to match the new tools, or investors will have to do more digging on their own.