Risk Ascent | Exploring the Edges of Risk

For 20 years, companies treated data like oil – drilling for it, extracting it, and refining it. But we’ve hit a crisis: the oil wells are drying up.

High-quality public data which fueled every major AI model is disappearing. To build the next generation of AI, extraction isn’t enough. We have to manufacture it.

Enter Synthetic Data – the renewable energy of the AI world. ⚡

1️⃣ What Is Synthetic Data?
🛫 Think of it like a hyper-realistic flight simulator. It isn’t “fake.” It is manufactured reality – computer-generated data that behaves exactly like the real world but contains zero real people.

📊 Real Data: Actual patient records (Finite • Sensitive • Regulated)
📈 Synthetic Data: Invented patients with real-world patterns (Infinite • Programmable • Safe)

2️⃣ Why Do We Need It?
🚧 We are hitting the Data Scarcity Wall. Research from Epoch AI estimates that between 2026-2032, we may exhaust the high-quality public data needed to train frontier AI.

We face three “breaking points”:
🧱 The Scarcity Point: There isn’t enough human-written content left to make AI smarter.
🔒 The Privacy Point: Real data carries regulatory friction to move. Synthetic data travels safely across teams and clouds.
⚠️ The Rare-Event Point: You can’t wait for a real car crash to occur. AI must learn from manufactured extreme scenarios.

3️⃣ Who Is Already Using It?
🚗 Waymo (Self-Driving Cars): Uses billions of miles of simulation to manufacture dangerous “edge cases” (like a child running into traffic) – allowing the AI to master safety-critical scenarios that are unethical to test in the real world.

🏥 Roche (Healthcare): Uses Synthetic Control Arms to create “virtual” placebo/control groups from historical data – allowing every real patient to receive the active drug and accelerating the path to a cure.

💳 JPMorgan (Banking): Uses synthetic financial crime patterns to train AI agents on complex money laundering schemes – enabling the detection of rare fraud without ever exposing sensitive customer data.

4️⃣ The Governance Gap (Where Firms Will Fail)
Gartner predicts that by 2030, AI will train on more synthetic data than real data. They also predict that by 2027, 60% of data leaders will face critical failures in managing synthetic data.

We are moving from Data Mining ⛏️ (digging for a finite resource) to Data Foundries 🏭 (engineering an infinite one).

5️⃣ The One Big Risk: “Model Collapse”
🧨 If AI trains only on synthetic data, it becomes like a photocopy of a photocopy – getting blurrier every generation. This is scientifically proven as Model Collapse (Nature, 2024).

The Solution for Risk Leaders:
Maintain a “Human Anchor”: You must preserve a Golden Ratio (often ~25-30%) of real human data to ground the model.

💡 The Bottom Line
In the AI era, competitive advantage won’t come from how much data you collect. It will come from how intelligently you can manufacture and govern the data you need.