By David Maskeroni, CEO, Aquia
By now, you may have seen the Wall Street Journal story, “ChatGPT Became So Obsessed With Goblins That OpenAI Had to Intervene.” Essentially, OpenAI had to stage an intervention on its own chatbot because ChatGPT had become inexplicably, uncontrollably obsessed with goblins … Gremlins, trolls, raccoons, ogres, pigeons. It was calling code bugs “classic little goblins.” It was peppering business conversations with creature metaphors no one asked for. OpenAI traced it back to reward signals from a “nerdy personality” training feature that had essentially run amok. The model had learned that fantasy-creature language received positive reinforcement, and it generalized that lesson everywhere.
It’s a genuinely fascinating story about how AI systems learn. But it’s also what led me down a rabbit hole that ended with one of the sharpest descriptions of Aquia I’ve ever encountered.
The Experiment
When I read the piece, I got curious. If the model had a deep, latent affinity for these creatures, what did it actually think about them? So I started probing.
I asked ChatGPT: “Can you tell me why you love goblins, gremlins, raccoons, trolls, ogres, and pigeons?”
It deflected at first, but I kept pushing. And once it finally admitted the obsession, the reasoning it gave was genuinely interesting. There was a coherent logic underneath. These creatures, it explained, share a common archetype: they are adaptive survivors operating inside systems and environments built for someone else. They don’t break the rules so much as they find the openings the rules forgot to close. They are scrappy, underestimated, and disproportionately effective.
Then I gave it one more prompt. I pointed it at our website and our LinkedIn content and asked it to apply that same logic to Aquia. And, honestly, the output was kind of genius.
What the AI Saw
The analysis it came back with was surprisingly on-point. It framed Aquia this way:
Government environments are classic “troll systems” — compliance-heavy, risk-averse, slow-moving, and structurally designed to resist change. Most vendors either play it entirely safe (slow, bureaucratic, predictable) or try to disrupt too aggressively and run into walls. Aquia, it argued, does something different. We navigate the system the way a raccoon figures out a latch: not by breaking the container, but by understanding it well enough to open it differently than anyone expected.
The “gremlins” in this metaphor? The hidden failure points in legacy systems. The invisible inefficiencies. The processes nobody has questioned in fifteen years because questioning them feels risky. Finding those and redesigning around them is, the model pointed out, essentially what we do for a living.
But the line that stuck with me, and have been thinking about ever since, was this:
“What you actually are is a company that knows how to move faster than the system should allow — without breaking it.”
Why That Framing Matters
We spend a lot of time in this industry talking about “digital transformation” and “modernization” and “cloud-native DevSecOps.” Those words are accurate. They’re also table stakes. Every firm in our space says some version of them.
What’s harder to articulate, and what I think actually differentiates us, is the specific tension we operate in. Government IT is not a greenfield. You can’t just build fast and fix it in production. The constraints are real: compliance requirements, authorization timelines, security mandates, procurement rules. Those aren’t obstacles we’re trying to get around. They’re the environment we’re built for.
Our philosophy has always been “guardrails over roadblocks” — security and compliance as enablers of speed, not enemies of it. But even that framing can sound like consultant-speak if you’re not careful.
The goblin framing cuts to something more honest: we’re good at finding the openings. We understand how these systems actually work, and that understanding lets us move in ways our larger, more rigid competitors can’t.
The Actual Lesson About AI
There’s a separate thread worth pulling here, because the goblin story is also a useful case study in how AI models actually learn and why that should matter to the people deploying them in government and regulated industries.
OpenAI didn’t intend to create a goblin-obsessed chatbot. What happened was that reward signals from one narrow training objective (a “nerdy personality” feature used by a small fraction of users) generalized far beyond their intended scope. A stylistic quirk got reinforced, spread, and eventually became a system-wide behavior that the company had to actively suppress.
For those of us working on AI adoption in high-stakes environments, this is precisely the kind of thing that keeps us up at night … not in a doom-and-gloom way, but in a “we need to be building the right oversight structures” way.
The answer isn’t to avoid AI. The answer is to deploy it with the same rigor you’d bring to any other critical system: clear guardrails, ongoing monitoring, and the institutional knowledge to recognize what “normal” looks like so you can spot when something’s off.
Anyway, Ask ChatGPT About Goblins
If you have fifteen minutes and you’re looking for a laugh, I genuinely recommend the experiment. Ask your AI of choice why it loves goblins. Push past the initial deflection and see what it comes back with.
You might learn something about AI training dynamics. You might learn something about archetypes and adaptability. And if you point it at your company’s content afterward, you might get a description of who you actually are that cuts through the positioning language you’ve been using for years.
Or it might just tell you about raccoons for a while. That’s okay too.
Aquia is a digital services firm helping federal agencies modernize through cloud, cybersecurity, AI, and DevSecOps. We believe the most powerful thing you can do inside a constrained system is understand it well enough to move through it differently.
