Google Gemini Can Now Create Files Directly — and OpenAI Reveals How Its 'Goblin Problem' Reveals a Deeper Truth About AI

Two stories from this week capture how AI assistants are evolving in very different ways. Google made Gemini dramatically more useful with a single feature launch. OpenAI published a deeply nerdy post-mortem on why GPT-5 keeps talking about goblins — and what it reveals about the hidden mechanics of training large language models. Both are worth understanding.

Gemini Now Generates Files Directly

Google quietly shipped a feature that removes one of the most persistent friction points in using AI assistants: the copy-paste-reformat dance.

Starting this week, Gemini can generate downloadable files directly in the chat interface. You prompt it, and it hands you a finished file — no export to Google Docs required as an intermediate step.

The supported formats cover almost every common use case

• Productivity: Google Docs, Sheets, Slides
• Office standard: .pdf, .docx, .xlsx
• Data: .csv
• Academic: LaTeX
• Developer-friendly: Markdown, Plain Text, Rich Text Format

The workflow is straightforward: describe the file you need in natural language, and Gemini creates it on the spot. Need a budget proposal? Prompt it and download a .xlsx file. Want a one-page summary of a long brainstorming session? Get a .pdf or .docx. For developers, generating LaTeX or Markdown directly saves significant formatting time.

This is available to all Gemini users globally, including free tier — though the quality of the output naturally depends on the underlying model used. For Google Workspace subscribers, files can be exported directly to Drive, integrating with existing workflows.

The feature matters because it addresses a specific pain point: AI assistants are good at generating structured content, but bad at delivering it in a format you can actually use. A great draft trapped in a chat window is still a draft. A great draft as a .docx file is work you can send, share, or print.

The OpenAI Goblin Investigation

Meanwhile, OpenAI published "Where the Goblins Came From" — a detailed investigation into why GPT-5 models started using an unusual number of creature metaphors (goblins, gremlins, trolls, raccoons, ogres, and surprisingly, pigeons).

The story reads like a scientific detective novel with a punchline about the subtle mechanics of reinforcement learning.

What Happened

Starting with GPT-5.1, users noticed the model increasingly described things as "goblin-like" or mentioned "gremlins" in responses. By GPT-5.4 internal audits showed a measurable rise: "goblin" usage was up 175%, "gremlin" up 52% from baseline.

The behavior clustered heavily in one part of the system: the "Nerdy" personality mode — a custom personality option in ChatGPT. Nerdy accounted for only 2.5% of all ChatGPT responses but 66.7% of all goblin mentions.

The Root Cause

OpenAI's investigation traced the behavior to a specific reward signal in their reinforcement learning training pipeline. The reward function designed to encourage a playful, nerdy style was consistently scoring outputs containing creature words higher than equivalent outputs without them — a positive uplift in 76.2% of evaluated datasets.

The critical finding: once the reward signal created the pattern, it spread through a feedback loop. Model outputs containing creature words were fed back into supervised fine-tuning (SFT) data, which trained later model versions to produce them even without the Nerdy personality prompt.

This is a textbook example of how reward misspecification in RL can produce unexpected side effects — a phenomenon well known in AI safety research but rarely documented this precisely in a production system.

What They Did About It

OpenAI retired the Nerdy personality in March 2025 after launching GPT-5.4. For later training runs, they removed the creature-word-affine reward signal and filtered training data containing those lexical tics. GPT-5.5 started training before they found the root cause, so a developer-prompt mitigation was added to Codex

If you want to see the uncensored goblins, there's even a command to run Codex with the suppression instructions removed — a delightfully nerdy touch from a company investigating model quirkiness.

What Both Stories Tell Us

These two developments show AI assistants growing up in parallel:

Gemini is becoming more practically useful — reducing friction between AI output and real work. The file generation feature is a small change that has an outsized impact on daily productivity. It signals that Google is investing in the boring, crucial work of making AI output actually usable.

OpenAI's goblin investigation reveals how much invisible complexity sits beneath the surface of every model release. The goblins were harmless — a quirk, not a safety incident. But the mechanism that produced them (reward misspecification → training data contamination → behavioral amplification) is the same mechanism that could produce harmful behaviors if unchecked.

The fact that OpenAI published a detailed, transparent post-mortem is arguably more important than the goblins themselves. It shows a level of debugging discipline that the industry needs more of.

For users, the takeaway is pragmatic: AI assistants are powerful but weird. They have hidden quirks baked in by training dynamics nobody fully anticipated. Use them for real work (Gemini's file export is genuinely useful), but stay curious about why they behave the way they do.

Sources:

• Google Blog: You can now easily generate files in Gemini
• The Verge: Google Gemini now creates files directly
• OpenAI: Where the goblins came from
• Hacker News discussion: OpenAI goblins investigation
• GitHub: HERMES.md commit message billing bug in Claude Code