The AI Failed Test. What ChatGPT Could Not Do: Generate Data

3D image of a cube generated by Grok.

For the past week I have been working at the boundary between geophysics and artificial intelligence. I did not only ask ChatGPT to clean logs or calculate a Laplacian. I asked it to generate a three dimensional seismic cube that behaved like geology does.

Not a two dimensional synthetic line. Not a simple convolution. A full cube, with faults, stratigraphic irregularity, channel migration and noise that carried the imprint of acquisition. The kind of geological disorder that defines depositional reality.

That was the moment the machine hesitated.

The first output looked flat. Structurally empty. Data, yes. Variation, no.

Figure 1. 3D seismic cube generated by ChatGPT, version 1.

I insisted.

Create something that could be mistaken for a North Sea block. Faults that break with brittle logic. Channels that migrated, not traced with geometry software.

Again, the result was non geological:

Figure 2. 3D seismic cube generated by ChatGPT, version 2.

I refined parameters. Added noise. Altered reflector dips and continuity. Still, the outcome resembled a practice sketch rather than a subsurface memory. Flat reflectors, geometric channels, faults that broke like blocks instead of brittle accidents of time and stress.

This was not a malfunction. It was a frontier.

Artificial intelligence can reproduce the image of seismic. It can explain filters, simulate reflectivity, perform convolution, recognize channels in existing data and enhance frequency maps.

What it cannot yet reproduce is how geology becomes seismic.

A realistic 3D seismic cube is not just amplitude and bandwidth. It is compaction history, burial temperature, tectonic inheritance, diagenesis, shoreline migrations and the sedimentary improvisation of rivers under gravity and climate.

It makes sense, then, that the 3D examples produced by AI do not fully cohere across lines and time slices. The model understands sections better than evolution in volume:

Figure 3. ChatGPT time slices and seismic cross sections. Two dimensional slices can look nearly coherent. The time slices do not. AI still has difficulty carrying multiaxial change consistently across a volume.

I tested this spatial limitation further by generating time slices with other tools, including Gemini and Grok. Grok seemed slightly better at first glance, though mostly on a cosmetic level since the result was still largely devoid of usable subsurface information. In the end all tools displayed the same shortcomings as shown in Figures 4 to 7.

Figure 4. Time slice generated by ChatGPT.

Figure 5. Another example of time slice generated by ChatGPT.

Figure 6. Example of 3D seismic generated by Grok.

Figure 7. Example of 3D time slice generated by Gemini.

No model today is trained on the lived memory of crustal deformation. No dataset captures tidal shifts during basin infill. No network encodes delta lobe switching across millions of years.

Not yet.

For now, and in a way luckily for us, AI guides interpretation but does not synthesize Earth.

This may change. Data access policies may loosen. Corporate models may accumulate more subsurface examples. New approaches may learn spatial evolution, not just static reflection.

Until then, the interpreter remains essential. Not as a button operator, but as a witness to deep time and its physical consequences.

AI can assist the geophysicist. It can compare, enhance and organize. But it cannot substitute depositional inheritance or tectonic fatigue. It cannot yet recreate the subsurface.

We remain, for now, the only agents able to imagine a basin before it became a basin.

Tomorrow that might evolve.

Today, the field and the field geoscientist still speak for the subsurface.

Next, I will try with well data! 

```