Gen AI’s Killer Feature Isn’t Generating. It’s Reading.
Two modes, one machine: Why reading mode delivers value and writing mode delivers fiction, and why hallucinations and creativity are the 2 faces of the same coin.
You feed ChatGPT 100-page report. It unearths patterns you missed, connects dots you never saw, and synthesizes insights that were buried in the noise all along. It feels like magic.
You ask the same AI to write a marketing strategy from scratch for company X. It hallucinates company initiatives that don’t exist and invents statistics with decimal-point precision. It feels like a liability.
It’s the same machine. The difference is the mode.
One function reads patterns from a reality you provide. The other writes fictions based on probability.
The core problem of the AI era is that most users can’t tell which mode they’re in, leaving them stranded between breakthrough and bullshit.
The AI’s Split Personality
Every Large Language Model has a split personality. We’ve been treating it as a single entity, which is why it seems brilliant one moment and delusional the next. The solution is to recognize the two machines hiding inside the one.
Reading Mode: The AI acts as a superhuman analyst. It processes, synthesizes, and finds patterns within a specific context you provide. It operates on what exists.
Writing Mode: The AI acts as a probabilistic storyteller. It generates new content based on the statistical patterns of its vast, chaotic training data. It creates what could plausibly exist.
Once you see this distinction, the entire industry’s behavior snaps into focus.
This explains why the gold standard for useful AI in the enterprise is RAG (Retrieval-Augmented Generation), a technical term for a simple command: “Read these documents first, then write.” Microsoft’s Copilot reads your files. Perplexity reads the live web.
Every successful AI tool is a tacit admission that the machine must be forced into reading mode to be reliable. As Shopify CEO Tobi Lütke says, the new frontier isn’t prompt engineering; it’s “context engineering”, the art of providing the right reality for the AI to read.
The Hallucination Engine
Hallucination isn’t a bug in Writing Mode. It’s the very engine of creativity we demand from it.
The mechanism is a parameter called “temperature.” which can be set between 0 and 1.
At low temperature (T≈0.2), the model produces quite predictable text—perfect for Reading Mode summaries. But at high temperature (T≈0.8), the setting physicist Stephen Wolfram identifies as optimal for creative writing, the model is encouraged to make surprising lexical leaps.
In Writing Mode, this high-temperature state is a fountain of both genius and gibberish. It generates novel metaphors and original ideas. It also generates plausible-sounding “facts” that are pure invention. They are products of the exact same process. You cannot have one without the other.
Research from the University of Washington confirms this is not fixable. When they “aligned” base models for safety—tuning them to reduce hallucinations—the process systematically lobotomized their creativity. The safe AI couldn’t write good poetry or even generate random numbers effectively. This is why GPT-4 can score in the top 1% on creativity tests while simultaneously inventing legal precedents. It’s why Stanford researchers found LLMs can generate ideas judged “significantly more novel than those from human experts,” but are incapable of recognizing their own fabrications.
So what do we do ? Midjourney co-founder David Holz offers the best metaphor. After generating 40,000 images, he felt he was “drowning” in a “torrent of water.” His insight: “We don’t try to get rid of water because it’s dangerous. We build dams. We build pipes. We build irrigation.” Trying to eliminate hallucinations from Writing Mode is like trying to make water not wet.
Mastering the Mode Spectrum
The path to mastery isn’t just recognizing Reading versus Writing Mode, it’s understanding the full spectrum of operations within each.
In any five-minute conversation with AI, you might need:
Extraction “What exactly does the contract say about termination?”
Synthesis: “Find the common themes across these customer interviews”
Critical Analysis: “What are the weaknesses in this business proposal?”
Creative Extension: “Generate ten campaign ideas based on these brand values”
The result of switching between these demands is often mode collapse, like a radio stuck between stations. You ask for facts, then ideas, then analysis, but the machine is still operating at the temperature and context of your first question.
The consequences of mode confusion are everywhere. When consultant Claudia Ng asked an AI to analyze a market, it “fabricated an entire analysis” with fake customer comments. She thought she was in Extraction Mode, but the AI had drifted into Creative Extension. Meanwhile, artist Caroline Zeller deliberately pushes into pure Creation Mode: “I want to go beyond what I want,” using AI to “reveal images that are in the unconscious.”
One user stumbled into the wrong mode and got fiction. The other chose her mode and found art.
The Mode-Conscious Approach:
Declare your mode explicitly: Don’t just ask questions—frame the operation.
“Extract only the facts from these documents, add nothing”
“Critically evaluate this proposal as a skeptical investor would”
“Creatively explore what these themes might suggest for our brand”
Reset between mode shifts: When switching from creative brainstorming to factual analysis, start fresh. The residual context from Creation Mode will contaminate your Extraction Mode.
If you have a tool which allows you to set temperature, Match it to task:
Low (0.2): Legal documents, data extraction, summaries
Medium (0.5): Strategic analysis, critical thinking, pattern finding
High (0.8): Brainstorming, creative writing, possibility exploration
Feed appropriate context for each mode:
Extraction needs comprehensive source documents
Analysis needs comparative examples and frameworks
Creation needs inspirational seeds, not restrictive rules
Conclusion: The Context Architect’s Toolkit
The individuals and companies mastering AI aren’t building better prompts. They’re designing information environments that keep AI in the appropriate mode for each task. They explicitly declare which operation they need and reset contexts between shifts. They understand that you don’t eliminate the flood; you build different channels for different purposes.
Gen AI isn’t one tool but a spectrum of them, packaged in a single interface. Master the mode spectrum, and you transform an unreliable oracle into a suite of powerful, specialized instruments.
The machine has always had multiple personalities. The breakthrough is learning to call forth the right one.
I like the emphasis on context engineering rather than prompt engineering. Feeding AI the right reality before asking it to generate output seems obvious in hindsight, but it’s a game-changer for reliability.
Super good advice, thank you!