Sharktech Case Studies – AI, Automation & Growth in NSW, Australia

SharkTech Global - AI Data Privacy Case Study

SHARKTECH GLOBAL

Case Study

Drafted by the Research and Analysis wing of Sharktech Global | October 2025

Let's Talk About What Nobody Tells You

Righto, here's something that'll make you sit up: When you hit delete on that ChatGPT conversation or clear your Claude history, you're only deleting your ability to see it—the data itself? Still sitting on their servers.

Unknown Fact #1: "Delete" Doesn't Mean Gone

AI companies keep multiple copies of your data for model training, safety checks, and system improvements. Your "deleted" conversations live on in:

Training data archives - Cleaned up but still there, helping train future models
Safety monitoring databases - Kept forever for abuse detection
The AI's actual brain - Your data literally becomes part of how the AI thinks
Backup systems - Standard disaster recovery stuff, usually 90+ days minimum

Here's the kicker: Once your data gets baked into a large language model's training, there's no way to fully extract it. It's encoded across billions of tiny calculations throughout the entire neural network. Think of it like trying to un-bake a cake—good luck with that.

How Your Chats Actually Train Their Systems

The Training Loop They Don't Advertise

Second 1: You type something with your business data

Second 2: Hits their servers, gets logged with timestamp, your IP, and heaps of metadata

Hour 1: Automated systems check if it's good quality training material

Day 1: Quality stuff gets flagged for reinforcement learning

Week 1-4: Your data enters the fine-tuning queue

Forever: Patterns from your chat influence how the AI responds to everyone

Unknown Fact #2: You're Leaving a Digital Fingerprint

Every business using AI leaves what we call a "semantic fingerprint" in the system. If your team keeps asking similar questions about your secret sauce, the AI starts recognising the pattern—and those patterns can influence what it tells OTHER users asking related questions.

Imagine if your mate borrowed your Netflix and started getting recommendations based on what you've been watching. Same idea, but with your business intelligence.

"We've seen cases where competitors accidentally benefited from strategic insights another company fed into the same public AI months earlier. They had no idea—it just came through in the AI's suggestions."
— Data Privacy Researcher, 2025

It's Not Just What You Type—It's Everything Else Too

Unknown Fact #3: The Metadata Tells Your Whole Story

They're not just storing your prompts, mate. AI systems are collecting:

When you use it: Your work patterns, business hours, crunch times
What topics you focus on: Reveals your priorities and concerns
How many goes you take: Shows complexity of your challenges
Your industry lingo: Pinpoints exactly what market you're in
Question sequences: Exposes how your team actually works
Device info: Maps out your organisation's structure

Put it all together and you've got a comprehensive profile of how your business operates—and you never even had to share an official document.

Hypothetical Scenarios: Understanding the Risks

Disclaimer

The following scenarios are hypothetical illustrations based on common industry concerns and potential risk patterns. These are not documented real cases but rather educational examples to demonstrate how data exposure risks could manifest in practice.

Hypothetical Scenario 1: The Marketing Calendar Situation

Illustrative example based on common AI usage patterns:

Imagine a Sydney retailer uses ChatGPT to fine-tune their Christmas promotional calendar. Three months later, a competitor rolls out a suspiciously similar campaign structure. After some digging, they discover both marketing teams were using the same AI tool. The first company's detailed planning prompts could have subtly influenced the suggestions given to the second.

This scenario illustrates potential risks when multiple businesses in the same market use shared AI platforms.

Hypothetical Scenario 2: The Manufacturing Process Example

Illustrative example of potential IP exposure:

Consider a Melbourne manufacturer who has spent years perfecting their quality control process. Someone on the team might use AI to help write training manuals, detailing the whole procedure. Within months, industry forums could start discussing similar methods—potentially traced back to AI-generated content that had absorbed and recombined bits from multiple sources, including their confidential process.

Hypothetical Scenario 3: The Customer Data Risk

Illustrative example of compliance risks:

An employee might think they're being clever by pasting customer feedback into AI for sentiment analysis. Problem: it could include email addresses and phone numbers. That personally identifiable information could then:

Be stored on overseas servers (potential data sovereignty violation)
Be used to train sentiment models (potential compliance breach)
Become potentially accessible to system administrators (security risk)
Be difficult or impossible to fully retract (permanent exposure risk)

This scenario illustrates how the business would carry liability, not the AI provider.

Unknown Fact #4: AI Models Play Telephone With Your Data

AI trained on billions of conversations develops what researchers call emergent knowledge structures. Your industry-specific data mixes with everything else, creating unexpected knowledge leakage. The AI might figure out connections between your business and market trends that you never explicitly mentioned—then share those insights with your competitors.

87%

of businesses unaware AI keeps deleted conversations (est.)

650M+

business prompts processed monthly (est.)

chance of guaranteed data removal

18mo

average time your data influences models (est.)