The geoscience industry has been digitizing, standardizing, and debating data management for the best part of fifty years. So when AI arrived promising to transform how we find, develop and manage the Earth’s resources, the key issue wasn’t the technology, it was whether the data could support it.
That was the territory covered in our recent webinar – From Field Tapes to Interpretable AI: The Hard Lessons Behind AI-Ready Geoscience Data – a fireside chat with Joe Reilly, President of the Society of Exploration Geophysicists (SEG) and former Chief Research Geoscientist at ExxonMobil, interviewed by Jasmine Tran, Head of Strategic Partnerships at Ovation Data. What emerged was less a vision of AI’s future and more a frank account of what stands between the industry and AI-ready data; and why that gap is harder to close than might be imagined.
The legacy problem is older than you think
Joe Reilly began where his career started, at the very beginning of digital geoscience. “I started work in the age of nine track tapes, paper seismic sections and well logs. Technical reports were handwritten and only final reports typed by a secretary.” Each of the formats that followed – 490s, optical platters, exabyte tapes, Digital Linear Tape (DLT), Linear Tape-Open (LTO) – solved a problem of their time and created one for the future. All of those media types still exist somewhere in archives, each requiring different hardware equipment to read it, much of it no longer in commercial use.
Whether data gets digitized often comes down to whether someone sees enough cost-benefit to be ready to pay for it. “Data is only transcribed from legacy media if there’s a perceived need of some value in what is recorded,” Reilly observed. “So this does not happen automatically, all at once, and it’s a significant financial investment.” The tapes sitting in cupboards and storage facilities represent a genuine intelligence asset, indeed, in some frontier environments, 1970s surveys simply cannot be reacquired, but unlocking that asset requires capital, judgment, and expertise that is in short supply.
‘AI-ready’ is not the same as ‘digitized’
Getting data off legacy media is a necessary first step, but not enough on its own for our AI-enabled future. Reilly was precise in drawing a distinction between what public large language models (LLMs) can do and what AI for geoscience actually requires. “The current generation of public LLMs are absolutely great for general knowledge, but they’re simply translating and interpreting text…whereas fundamental data analysis requires problem-focused machine learning and AI [agents] that understand the base data as well as the context around it.”
That context (processing history, acquisition parameters, interpretation lineage) is exactly what is missing or inconsistent across much of the industry’s archive, creating the conditions for confident errors at scale. “Poor data, well labelled, properly reformatted,” Reilly said, “is easily misused.”
AI, in Reilly’s view, doesn’t change that: “I don’t think AI is an excuse for sloppy data management.” The discipline required to make data AI-ready is the same discipline that good data management has always demanded; the stakes are simply higher now.
The judgement that doesn’t go away
One of the conversation’s most practical threads was the “good enough” question; when is data sufficiently fit for purpose, and who decides? “Perfection can get in the way of good enough,” said Reilly, describing the guidance he often gives when teaching university classes. Legacy, 1970s data may be entirely appropriate for identifying frontier leads; it is not appropriate for reservoir management or mineral resource estimation. What matters is whether the data is good enough for the specific decision at hand.
The same data that is good enough for one application can be actively dangerous in another, particularly when AI-generated outputs are presented without clear audit trails or qualification. Jasmine Tran explained the importance of keeping human expertise in the decision-making process: “The expert value is knowing when the data doesn’t fit.”
The human geoscientist, in other words, remains the critical control. Not as a bottleneck to AI adoption, but as the judgment layer that makes AI outputs trustworthy. As Reilly put it in his closing remarks, to work effectively in this environment, you need to “know your data better than any AI agent at that time can assess.”
What this means in practice
Progress is being made. The SEG-Y rev 2.1 standard (an updated version of the SEG-Y seismic data format) extends headers to anchor processing reports and interpretation files directly to the seismic data, keeping project context intact and cloud-accessible. The Open Subsurface Data Universe (OSDU) is developing reformatters that allow data to be accessed and converted without depending on the software vendor that owns the original format. The SEG’s own SEG Advanced Modelling (SEAM) AI project is building publicly licensed, well-labelled datasets to test and develop stratigraphic AI agents against real geological complexity.
Reilly and Tran agreed that these are practical steps and the right framing for the industry at the moment. Its analytical foundations were built over five decades of accumulated data, formats, and institutional knowledge, and it will not become AI-ready through any single platform or initiative. It will get there through rigorous, unglamorous work; labeling, versioning, provenance tracking, and the human expertise to know what the data is actually telling you.
As Reilly put it – speaking directly to the next generation of geoscientists: “You’re not late. You’re right on time.”
This blog draws on Ovation Data’s February 25, 2026 webinar From Field Tapes to Interpretable AI: The Hard Lessons Behind AI-Ready Geoscience Data, featuring Joe Reilly, President of the Society of Exploration Geophysicists, and Jasmine Tran, Head of Strategic Partnerships and Global Key Accounts at Ovation Data.