I asked AI to create a well log. This is what happened.

Well Log generated by ChatGPT

After watching generative AI produce eerily realistic faces, paintings, and even seismic sections, I decided to push the experiment into more hostile territory.

Not art.
Not text.
Not seismic images.

Well logs.

If there is a graphical object geoscientists believe they understand instinctively, it is the one dimensional well log. Gamma ray, resistivity, depth. A narrow strip of ink that quietly encodes lithology, fluids, and depositional history. You cannot bluff your way through it.

So I asked three large language models ChatGPT, Gemini, and Grok to create a synthetic well log.

Not calculate one. Just create one.

The outcome was not subtle.

At first glance, all three models produced something that looks right. Depth axes. Wiggles. Labels. Shaded zones. Enough familiar structure to satisfy a casual glance. Enough visual confidence to pass in a presentation.

And that is already dangerous.

Because the moment you stop looking and start reading, the illusion collapses.

Let us begin with the simplest case: gamma ray.

Well Log? generated by Genimi.

Well Log? generated by Grok.

A real gamma ray log is noisy, but not arbitrary. It has memory. Shale packages persist. Clean intervals are rare and therefore meaningful. Vertical coherence reflects depositional process, not artistic preference. The curve is ugly because geology is ugly.

The synthetic gamma ray curves produced by these models are different. They spike. They wander. They decorate the track. But they do not commit. There is no stratigraphic intent. No long range continuity. The curve moves because movement looks realistic, not because the rock requires it.

This is not a minor flaw. A gamma ray log without vertical correlation is not a gamma ray log. It is a waveform.

Now add resistivity, and the pretense fails outright.

Resistivity is where geology stops tolerating excuses. Shales conduct. Clean hydrocarbon bearing sands resist. Evaporites dominate. Invasion leaves fingerprints. These are not conventions. They are consequences of physics operating across orders of magnitude.

In the AI generated logs, resistivity behaves like a decorative accessory. It wiggles independently of gamma ray. It contradicts lithology. It refuses to tell a fluid story. The supposed reservoir intervals do not stand out. There is no depth where an interpreter would stake a decision.

This failure is consistent across ChatGPT, Gemini, and Grok.

Their styles differ. Their layouts vary. Their aesthetics improve with each iteration. But they all make the same conceptual mistake.

They treat each track as an independent drawing.

A real well log is not a collage. It is a constraint system. Gamma ray limits what resistivity is allowed to do. Lithology constrains both. Fluids impose further restrictions. When one curve lies, the others expose it.

The models do not understand this coupling.

That is why well logs are such a brutal stress test for generative AI. These systems excel at surface realism. They recognize visual grammar. They reproduce fonts, grids, layouts, and familiar noise. They are excellent illustrators.

They are not yet geologists.

They draw curves, not rocks.

Could these images fool an expert. Briefly. In a slide deck. In a report skimmed too quickly. In the hands of someone who trusts appearance over interpretation.

And that is the uncomfortable part.

This is not about whether AI can replace geoscientists. It cannot. Not even close. This is about whether it can produce artifacts that look authoritative without being constrained by reality.

Well logs expose that gap mercilessly.

Until an AI can generate a log that survives cross track scrutiny, not just visual inspection, it remains a stylist, not a scientist. When that day arrives, it will not be subtle. It will be obvious. And it will be unsettling.

Until then, the well log remains a quiet act of resistance.

A narrow vertical line where physics still wins.

I also tried to go further.

I attempted to build an application that would generate electrical well logs. Not images, but curves. Something interactive. Something that would let you assemble lithologies, thicknesses, and responses, and produce something that could pass as a real log.

I failed.

Not because the code did not run. Not because the math was impossible. I failed because every time I looked at the result, my geological intuition rejected it immediately. The curves were plausible in isolation, but wrong in relationship. They looked right, but they did not behave right.

That failure was instructive.

Electrical logs are not drawings. They are not textures. They are the emergent result of physics, chemistry, and geometry acting together in a very constrained vertical space. You cannot fake them by assembling rules the way you might fake an image.

I now suspect that the only way to generate something truly realistic is not through prompt engineering or handcrafted logic, but through machine learning trained on real log data. Not to imitate the look, but to internalize the coupling between measurements. I will try later when I have the time. programming has become a real passion and even a meditative process. 

In other words, you do not teach an AI what a gamma ray log looks like. You teach it what rocks do.

At this stage, I failed, and that is ok. And that is an important result. It tells us where the boundary currently lies. Generative models can draw the surface of geoscience. They cannot yet reproduce its constraints (for now).

When that changes, it will not be because the curves look better. It will be because the relationships stop breaking.

Until then, failure is the correct outcome.