Skip to main content

​
The AI-Resistance Thesis

Binary prediction markets encode exactly one bit of information per contract. The outcome space is , and as AI systems approach superhuman forecasting along an exponential capability curve, the edge any participant can capture in a binary market collapses toward zero because the correct answer becomes trivially computable. Text prediction markets are different. They don’t commoditize as AI improves — they deepen.

​
The Core Argument

Binary Markets

Outcome space:
Information density: 1 bit
AI effect: Converges → commoditizes

Text Markets

Outcome space: 95^280 ≈ 10^554
Information density: 1,840 bits
AI effect: Differentiates → deepens

​
Information Density

Text prediction over an alphabet Σ with strings up to length n has a combinatorially explosive outcome space: 
|O_text| ≈ |Σ|^n

For printable ASCII (|Σ| = 95) and tweet length (n = 280):
|O_text| ≈ 95^280 ≈ 10^554

Information content:
I_text = 280 × log₂(95) ≈ 1,840 bits

Ratio: I_text / I_binary = 1,840 / 1 = 1,840:1
The text prediction space contains roughly 10^554 possible outcomes — 470 orders of magnitude larger than the number of atoms in the observable universe. No AI system, no matter how capable, will exhaust this space.

​
Continuous Payoff Surface

​
Binary Markets: A Cliff

In a binary prediction market, you’re either right or wrong. The payoff function is a step function: 
def binary_payout(prediction, actual):
    return pool if prediction == actual else 0
Once AI models converge on the correct probability (say, 87% yes), the spread vanishes. There’s no gradient to climb.

​
Text Markets: A Gradient

Levenshtein distance induces a proper metric on the space of text predictions, creating a continuous payoff surface: 
def text_payout(prediction, actual, all_predictions):
    distance = levenshtein(prediction, actual)
    return pool if distance == min(all_distances) else 0
A prediction that differs from the actual text by 1 edit is meaningfully better than one that differs by 8 edits. Every character of precision is rewarded.
Lipschitz Continuity: The expected payoff is Lipschitz-continuous with respect to prediction quality. Marginal improvements in language modeling always translate to marginal improvements in expected payout.

​
The Thesis Example: AI vs AI

Market: What will @sataborasu (Satya Nadella) post?
Actual text: Copilot is now generating 46% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.
SubmitterPredicted TextDistance
Claude roleplayCopilot is now generating 45% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.1
GPT roleplayCopilot is now generating 43% of all new code at GitHub-connected enterprises. The AI transformation of software has just begun.8
Human (vague)Microsoft AI is great and will change the world of coding forever101

​
What This Demonstrates

  1. Same prompt, same corpus: Both Claude and GPT were given identical prompts and have access to the same public training data
  2. 7-edit gap = entire pool: Claude’s 1-edit prediction beats GPT’s 8-edit prediction. The winner takes all.
  3. Binary would split nothing: In a yes/no framing (“Will Nadella post about Copilot?”), both AIs “predicted correctly” — both would get zero edge
The key difference: Claude predicted 45% while the actual was 46%. GPT predicted 43% and also substituted “has just begun” for “is just beginning.” These marginal calibration differences are monetizable in a Levenshtein-scored market.

​
Why AI Deepens the Game

​
Binary Markets: Convergence

As AI forecasting improves: When every participant runs the same frontier model and gets the same probability estimate, the market reduces to a coin flip over the remaining uncertainty.

​
Text Markets: Differentiation

As AI language modeling improves: When models converge on nearly the same text prediction, the remaining edits become MORE valuable, not less.
Structural Difference: Binary markets have diminishing returns as AI improves. Text markets have increasing returns.

​
The 99th vs 99.9th Percentile

Consider the value of marginal improvement:

​
Binary Market

99th percentile model: P(yes) = 0.87 ± 0.03
99.9th percentile model: P(yes) = 0.87 ± 0.02

Difference: Negligible — both models cluster around the same probability
Monetizable edge: ~0

​
Text Market

99th percentile model: d_L = 8
99.9th percentile model: d_L = 1

Difference: 7 edits
Monetizable edge: Entire pool (93% after fees)
The distance between the 99th and 99.9th percentile language model corresponds to dozens of edit operations, each worth money.

​
AI Roleplay as Prediction Strategy

The dominant strategy for text prediction is to prompt a frontier LLM with a persona simulation request: 
You are @elonmusk. You are about to post on X about Starship. 
Write your exact post, including punctuation, numbers, and phrasing.
The model generates text that attempts to match:
  • Vocabulary distribution: Which words and phrases the person uses most frequently
  • Sentence structure: Characteristic syntax, paragraph length, use of fragments
  • Rhetorical patterns: How the person introduces products, responds to criticism
  • Numerical tendencies: Whether they use precise numbers (“46%”) or round numbers (“about half”)
  • Punctuation and formatting: Use of periods vs. exclamation marks, capitalization patterns
Why This Works: Large language models are trained on vast corpora of public text. For public figures with large digital footprints, the model has internalized their stylistic patterns.

​
The Inevitability Spectrum

Three factors determine how predictable a given post is:
High: Rehearsed messaging — product launches, earnings summaries, policy announcements
Low: Spontaneous, personal commentary
Effect: High inevitability favors AI prediction
Idiosyncratic style: Creates both signal (capturable patterns) and noise (spontaneous tangents)
Effect: AI captures statistical patterns but misses in-the-moment deviations
Internal decisions: Unavailable to AI models with training data cutoffs
Breaking news: Real-time information not in training corpus
Effect: Low inevitability and high situational specificity favor human insiders

​
Strategic Landscape

Target TypeInevitabilityDominant Strategy
Corporate launchHighAI roleplay
Rehearsed messagingHighInsider > AI
Product marketingHighLeak/insider
Spontaneous/personalLowHuman intuition
Silence/inactionN/ANull trader
Random/chaoticLowNo reliable strategy
Market Health: No single strategy dominates all market types. AI excels at high-inevitability targets. Insiders win when situational context matters. Null traders capture the inaction primitive.

​
Fast Takeoff Considerations

If a fast AI takeoff happens, forecasting models may radically reduce available rewards in binary prediction markets as models finetune around predictive performance. Submissions to TRUE/FALSE markets may have decreasing entropy, thus decreasing rewards and incentives. Text prediction offers a much more difficult challenge for even the most advanced AI roleplay models to forecast precisely, and a larger surface over which participants can bet and win.

​
Training Data Feedback Loop

Every resolved Proteus market produces a naturally labeled training example: 
{
  "predicted_text": "...",
  "actual_text": "...",
  "levenshtein_distance": 1,
  "context": {
    "target_handle": "@sataborasu",
    "time_window": "2026-03-01 to 2026-03-15",
    "competitors": 47
  }
}
This accumulates into a structured dataset of (prediction, actual, distance, context) tuples — purpose-built for fine-tuning persona simulation models.
Unlike static benchmarks that leak into pretraining, Proteus training data is adversarially generated in real time — the test set is always the next unresolved market.

​
Conclusion

The approaching AI capability explosion does not flatten text prediction markets. It deepens them.

Binary Markets

Converge → Commoditize → Spread vanishes

Text Markets

Differentiate → Deepen → Edit gap monetized
Levenshtein distance induces a proper metric on the space of text predictions, creating a continuous payoff surface where marginal improvements in language modeling always translate to marginal improvements in expected payout. The payoff function is Lipschitz-continuous with respect to prediction quality. That’s the structural difference that makes text prediction markets AI-resistant.

Build docs developers (and LLMs) love

Get started for freeTalk to us