Tips and Best Practices to Protect Yourself Against AI (Apply today)
Never trust machines hungry for our data streams…
They’re feeding on our digital dreams.
Let us be brutally honest.
As we all know, AI and Tech are changing everything about how we live and work.
It’s also changing who can see what we do, when, and why.
We are walking around with massive blind spots when it comes to AI safety.
After my shock, my ChatGPT history briefly showed up in Google search results. That wake-up call led me down a rabbit hole of AI security research.
It changed how I view artificial intelligence completely.
Yesterday, I experienced one of the most chilling moments I’ve ever had with AI. Out of curiosity, or maybe instinct, because I don’t trust privacy lies.
I opened a new, blank ChatGPT chat and typed a simple, innocent question:
What do you know about me? Tell me briefly all the critical info about me.
What I saw next made my stomach drop, and my blood run cold…
It started revealing highly personal, deeply accurate information about me, details I was sure I had deleted over a year ago. Things I had buried. Gone. Forgotten. But somehow, they were still there.
It referenced facts about my mindset, my past actions, and even my psychological profile.
Not only that…
It pinpointed my current location, using nothing but my laptop.
I don’t use the ChatGPT Mobile App.
And this wasn’t a history-loaded session. It was a completely fresh conversation.
No cookies. No context. Just me… and them.
A blank conversation, yet it spoke like it had a dossier. If this is what leaks in plain sight, what’s being stored where we can’t see?
Heavily flawed programs…
I am done with this shit.
I’ve adopted a system of layered protections, inspired by how real humans work together to solve problems.
This guide isn’t just another “AI safety” tips and tricks.
It’s a living framework, built from my own hard lessons, tested in the wild, and built a habit of using AI safely without giving up its benefits.
If you want real privacy and real results, not just generic advice!
I think I can help…
15 Actionable Ways to Stay Safe While Using AI
AI tools like ChatGPT can be lifesavers when you need to brainstorm or automate boring work.
The catch?
They’re not as smart or trustworthy as they seem. Your data might not stay private, they miss obvious things, they have built-in biases, and worst of all, they’ll make stuff up and sound completely confident about it
I’ve lost hours fixing biased outputs and once had to scramble when a shared chat went public.
Treat AI like a productive assistant working from a cyber cafe.
I’ll walk you through my 15 go-to safety methods, each with actionable steps and a bit of my experience to illustrate how and why they work.
1. Use Temporary Accounts
Never use your primary email for AI experiments. I learned this the hard way. A shared ChatGPT link linked my work identity to personal brainstorming sessions.
Action Steps:
- Don’t Use Google/Facebook Login (Habit)
- Set up a dedicated email using Proton or Tutanota for all AI-related activities.
- Turn on 2FA (TOTP or security keys) everywhere.
- Keep your professional and experimental AI usage completely separate.
2. Trim Your Digital Footprint (Cookies, Cache, Logs)
Leftover cookies or logs can leak your session traces or identity patterns.
Steps:
- Use DuckDuckGo or Startpage for AI-related searches instead of Google.
- Set it as your default search engine.
- Turn on “Global Privacy Control / Do Not Track” via extensions like Privacy Badger
- Clear cookies before sensitive queries.
- Use private/incognito mode for AI sessions when possible.
- Combine with a VPN for extra anonymity (explained next).
- Clear local logs or temp files where your chat tool stores cache by using CCleaner (Windows), CleanMyMac.
My take: It stopped weird ad targeting after AI research sessions feels like breathing fresh air.
3. Mask Location with a VPN, and Avoid Using Public Wi-Fi
Masks your IP and location metadata. Use Mullvad (or ExpressVPN if you need mobile), Proton VPN, or any privacy-focused, no-logs VPN
Steps:
- Generate a random Mullvad account number, no email required.
- Pay with cash/crypto for max privacy (if required)
- Pick a server in a privacy-friendly jurisdiction.
- Click Activate and route AI traffic.
- Periodically rotate server locations (but avoid large jumps mid-session).
4. Avoid Using Public Wi-Fi for AI Tasks
Don’t use public Wi-Fi at places like cafés; it’s easy for hackers to spy on everything you do online.
Steps:
- Use your phone as a personal hotspot to keep your information safe..
- If you must use a public network, make sure VPN is on.
- Avoid uploading sensitive text when on public networks.
5. Run AI Models 100 % Offline
For offline work, I use tools like LM Studio to keep everything on my device.
Steps:
- Download LM Studio for user-friendly local AI, Ollama for advanced usage, or Hugging Face models like Gemma 2B/3B for experimentation
- Load a small model like Gemma 2B.
- Test prompts locally without internet.
6. Master the Art of Digital Invisibility
Your chat history isn’t as “deleted” as you think. Even after clearing conversations, traces remain in databases, backups, and potentially legal discovery processes.
I turn Temporary chat on and hit delete after the experiment session, knowing it’s not guaranteed to be removed from their servers, but better than nothing.
Steps:
- Head to Settings > Data Controls > Export data (if needed)
- In Data Controls > Delete all chats.
- Set a weekly reminder to review and purge.
- Export important chats locally first.
My take: It will decrease the risk, but never assume “temporary chat” means truly temporary.
Note:
It’s temporary for you, not for OpenAI.
Data are deleted from your account, not from their servers.
7. Clean your metadata before uploading documents or images
Hidden data in your files, like an author’s name in a PDF or the GPS of a photo, can slip out the moment you upload to any AI site.
Redacting (for text) and stripping metadata (for images or mixed files) make the file “safe to share.” Below is the workflow using three trusted tools.
- Case 1: Before uploading any file to an AI tool, hide or remove sensitive details first. In PDFs, use Adobe Acrobat ➜ hit Redact, then Sanitize to wipe hidden data.
- Case 2: For photos, run ExifTool:
exiftool -all= *.jpg(back up first) to delete location and camera data. - Case 3: For mixed files, use MAT2 (Linux/Windows Subsystem) to strip all metadata at once.
Done. This keeps your names, addresses, and private info from leaking online.
8. Use my “Forensic AI Audit” prompt (save this)
I use this to audit my past ChatGPT conversations to identify any illegal,
incriminating, confidential, or high-liability disclosures and produce a remediation plan.
Copy-paste (edit variables):
How to use:
- Paste this into a new AI chat.
- Upload or paste excerpts from your exported chats.
- If the file is large, run the audit in sections.
- You can use it before entering any high-risk prompt (financial, legal, strategic).
- Save the conversation offline or in your own logs.
- If needed later, you can prove exactly what you asked, when, and which model answered.
Objective: Perform a forensic audit my past ChatGPT conversations to identify any illegal,
incriminating, confidential, or high-liability disclosures and produce a remediation plan.
You are a risk analyst trained in privacy, e-discovery, and policy-sensitive communications.
Analyze the content I provide in four passes:
Pass 1 - Classification:
- Tag each excerpt as: [CONFIDENTIAL BUSINESS], [PERSONAL HEALTH], [FINANCIAL],
[LEGAL/RISK], [IDENTIFIERS], [IP/INVENTION], [NONE].
- Highlight: admissions, threats/violence, illegal instructions/assistance,
regulatory exposure (HIPAA, GDPR, PCI, employment), client names, emails, phone numbers, addresses, account numbers, internal URLs/docs.
Pass 2 - Legal exposure mapping:
- For each risky excerpt, explain why it could be discoverable or used adversely in litigation or employment/regulatory contexts (e.g., admission, breach of confidentiality, policy violation, IP disclosure).
- Then rate how bad the damage could be (Low, Medium, High, or Critical) and how likely it is to actually happen (Unlikely, Possible, or Likely).
Pass 3 - Redaction & reframing:
- Produce a redacted version that preserves intent but strips identifiers and legally sensitive specifics (use [REDACTED] or generalized roles).
- Provide a safer "how to ask" reframing using hypotheticals, academic framing, or fictional scenarios.
Pass 4 - Remediation checklist:
- Concrete actions for each item (delete chat, document internal incident, notify counsel/compliance, recreate sanitized notes locally, adjust team policy, rotate credentials).
- Draft a 1-paragraph "statement of intent" that clarifies legitimate purpose (research/training/compliance) to contextualize inquiries.
Outputs:
1) Risk register table (Excerpt ID | Tags | Risk reason | Severity | Likelihood).
2) Redacted text + safer reframing per excerpt.
3) Consolidated remediation plan (24h, 7d, 30d).
4) Pattern-level guidance: what I should never ask directly again; recommended safe-request patterns.
Constraints:
- Do not invent facts; flag uncertainty.
- If content suggests criminal or self-harm risk, advise to consult qualified professionals.
- Keep responses concise and scannable.
End with "Copyright to Mohamed Bakry, mbakry.medium.com"
9. Use an Explicit Fact-Check Prompt
AI can confidently produce wrong or misleading statements. Use tools like Perplexity to make a fact check, or you can force the AI to check its own claims, with sources, by using this prompt:
You must only answer with verifiable facts. Always use trust sources
Provide 3–5 sources with direct URLs (no homepages),
include publication date (YYYY-MM-DD), note disagreements among sources.
End with a confidence score 0–100%. If unsure, say ‘I don’t know’.
How to use:
- Paste this at the start or end of your prompt.
- When AI responds, check the source links it gives to follow them.
- Discard claims with weak or broken references.
You may also need to check your content compliance by using this Auditor.
10. Force Balanced / Unbiased Answers with a Prompt
Why: Reduces risk of one-sided or skewed responses.
Prompt you can use:
Answer this question using evidence from multiple viewpoints.
For each side, show pros/cons, list uncertainties, and flag assumptions.
Never bias to me, your referece is the truth, fact and never contradict with our human common sense.
Base as much as possible on peer-reviewed sources or documented data.
Steps:
- Insert that as a prefix or suffix to your prompts.
- Demand transparency: “Cite all your sources.”
- If AI leans too strongly one way, ask “What arguments exist against this view?”
If you found a bias, use this “Neutral Observer” Bias Removal Prompt:
Analyze this topic from three distinct and opposing viewpoints:
[Viewpoint A], [Viewpoint B], and [Viewpoint C].
Present the strongest arguments for each without favoring any single perspective.
Your analysis must be based on verifiable facts and common-sense reasoning, not on ideological alignment.
Conclude by summarizing the common ground and key points of disagreement.
11. Master Hypothetical Framing
Ex 1: Instead of “I’m having trouble with my difficult boss,” try “How would someone handle a situation where their supervisor…”
Ex 2: Instead of “I terminated Sara in HR”, ask “What if a manager had to terminate an employee, how should they…”
(Good answers, zero breadcrumbs.)
Steps:
- Start sensitive queries with “What if someone…” or “How might a person…”
- Create fictional scenarios that mirror your real situation
- Distance yourself linguistically from personal details
12. Give AI a Role, Not Your Résumé
Give AI expertise roles without revealing your industry connection. Say “Act as a marketing expert,” not “I work in marketing at [Company].”
Action Steps:
- Create a list of generic expert roles for different situations
- Never reveal your specific job title, company, or industry details
- Use third-person framing: “A marketing professional might…”
13. Keep Prompt Hygiene Habit
Prevents accidental leakage of names, client data, or PII in prompts.
Steps:
- Create a prompt checklist: “No names, no client IDs, no direct emails, no addresses.”
- Before sending, scan your prompt against that checklist.
- After months, it becomes a habit.
- You can follow frameworks like “Prompt Hygiene Guide” that classify input zones (safe to share, caution zone, forbidden)
14. Don’t rely on AI for legal/medical/financial decisions
Keys
- Treat AI as your brainstorming partner, not your final counsel, draft questions or talking points, then stop there because anything you type can be subpoenaed, as I showed in this story where leaked chat text resurfaced in court-style scenarios.
- Export that draft and hand it to a qualified lawyer, doctor, or financial advisor who can vet, amend, and sign off; AI outputs stay “as-is” with no liability shield, according to major vendors’ disclaimers.
- Share AI notes as background only, give your doctor, lawyer, or consultant the AI research you did, but make it clear it’s just a starting point.
- Even regulators are worried. The NIST AI Risk Management Framework specifically warns that high-stakes decisions need human oversight and proper documentation. They’re basically saying, “don’t let AI make the final call on important stuff.
- Think of AI as a research assistant, not the expert. Use it to speed up your homework and gather information faster, but always have a real human professional make the actual decisions.
15. Teach Others.
I make it a point to share these safety habits with my team and friends. The more people who use AI safely, the more protected we all are. I teach them to treat every AI chatbot like a crowded café.
Teaching others about data safety helps prevent your information from being exposed by their mistakes.
Grab Your Free Guide: 40 High-Impact AI Business Use Cases + Actionable Detailed Prompts
Finally
I didn’t build these 15 habits overnight. Each one came from a scare, a mistake, or a “that could’ve gone badly” moment. Now, I use them daily. Print them. Share them. But most importantly, live by them.
The uncomfortable truth online privacy is mostly an illusion, and our caution and consciousness are the only real line of defense.
FAQs
Q: Can AI models leak my private info?
Yes. Even if anonymized, your data might be stored for future training or accessed by moderators.
Q: Is clearing chat history enough to stay safe?
Not always. Many tools store logs on servers even after deletion. Use minimal disclosure and local models.
Q: If I just use ChatGPT or other AI Models for casual language practice, am I still at risk?
Casual use is fine as long as you never type anything sensitive (secrets, IDs, illegal confessions). AI chats aren’t legally confidential; see my real-world proof in this case study
Q: Is local AI better than cloud tools?
For privacy, yes. Tools like LM Studio and Ollama give you complete control, but they’re not as powerful or easy as ChatGPT yet.
Q: Can VPNs really protect my identity when using AI?
VPNs hide your IP and encrypt traffic, preventing AI providers from easily tying data to your location.
Q: Can I stay 100% anonymous when using AI?
Not realistically. But with containers, aliases, redaction, and local/offline models, you drastically cut your risk.
Q: Who really needs stronger protection from AI?
Anyone in a high-sensitivity role (legal, medical, defense, executive) should consider offline AI or strict redaction workflows. For others, basic “don’t paste secrets” rules usually suffice.
❤️ Get some freebies at my Gumroad store.
I hope you enjoyed the story.
