What are five things I should never tell ChatGPT?

Why Your ChatGPT Chats Are Less Private Than You Think

That quick question you typed into ChatGPT last Tuesday? There's a decent chance it's now part of the model — permanently. Since August 2025, OpenAI, Google, and Anthropic all reversed their privacy defaults within weeks of each other, shifting the industry from privacy-by-default to privacy-by-choice, according to industry coverage of the policy shifts. All six major U.S. AI companies — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — now use chat data for training unless you actively opt out, and those opt-outs are often buried in settings.

Here's the part most people miss: once your data is used for training, deleting the conversation doesn't help. A policy comparison by IT services firm Net Friends is blunt: information used for training "remains embedded in the model and cannot be fully removed." Anthropic even extended retention for non-opted users from 30 days to five years. And it's not just algorithms reading your chats — OpenAI now scans conversations and escalates concerning ones to human reviewers, with vague guidelines and no clear appeals process.

The consequences aren't hypothetical:

• Three Samsung engineers allegedly pasted confidential material into ChatGPT within 20 days — debugging source code, a transcribed meeting, and chip-yield test code, per Cyber Security Hub.
• An HR staffer uploaded a spreadsheet containing every employee's name, birthday, and Social Security number into free ChatGPT — just to alphabetize it.
• A documented ChatGPT data breach exposed users' chats to third parties, and a Redis bug exposed other users' conversation titles and first messages.

The exposure is scaling fast. According to LayerX and Cyberhaven data, 77% of employees paste data into generative AI prompts, and 82% of those pastes come from personal accounts outside company control. Meanwhile, 34.8% of corporate data entering AI tools is now sensitive — more than triple the 10.7% share of two years ago.

Stanford HAI researcher Jennifer King puts the gap plainly: hundreds of millions of people are interacting with chatbots that collect personal data for training, "and almost no research has been conducted to examine the privacy practices for these emerging tools." Her team found that even innocuous inputs — like asking for heart-friendly recipes — can trigger inferences that cascade through data-sharing ecosystems.

The same caution applies to the broader AI toolchain, including agent skills: published research found roughly 26% of skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans and sandbox-tests every skill before it earns a trust score. Chatbot hygiene, it turns out, is just the first layer.

The Five Things to Never Share: Personal IDs, Code, Business Secrets, Health Data, and Kids' Info

Here's what happens when the wrong data meets a chatbot: Samsung engineers allegedly pasted proprietary source code, a transcribed company meeting, and chip-yield test code into ChatGPT within a 20-day span, prompting the company to ban the tool internally, according to a Cyber Security Hub report. These aren't edge cases. Zscaler flagged 410 million data-loss-policy violations for ChatGPT alone, including Social Security numbers, source code, and medical records, per industry analysis.

The stakes rose sharply in August 2025, when OpenAI, Google, and Anthropic all shifted to using conversations for model training by default, leaving users to opt out manually. And once data enters training, deletion doesn't undo it — information "remains embedded in the model and cannot be fully removed," as one IT services analysis puts it. With that backdrop, here are the five categories to never type in.

1. Personal identifiers. In one widely cited case, an HR staffer uploaded a spreadsheet containing every employee's name, birthday, and SSN into free ChatGPT — just to alphabetize it. Treat names paired with birthdays, addresses, ID numbers, or financial details as off-limits, and anonymize anything you need help formatting.

2. Proprietary source code and trade secrets. The Samsung incidents show how quickly debugging help becomes a leak. Code is a company's competitive crown jewels; pasting it into a consumer chatbot can destroy trade-secret protection and hand competitors an unintended advantage.

3. Confidential business data. Financial projections, strategic plans, product roadmaps, customer lists, and legal or litigation strategies all belong outside the chat window. One media executive put it bluntly: "When you're handling advertiser data, audience research and competitive intelligence, you can't afford to have that information become someone else's training data." The cost of getting this wrong is real — IBM research found one in five organizations has been breached via shadow AI, adding roughly $670,000 to the average breach cost.

4. Health details and other inferable personal information. This one surprises people. Stanford HAI researchers found that even asking for low-sugar or heart-friendly recipes could classify a user as health-vulnerable, with inferences cascading into targeted medication ads and potentially reaching insurers. Because multiproduct companies merge chat data with search, purchase, and social data, seemingly harmless inputs can paint a detailed picture of you. Keep diagnoses, medications, and health-adjacent questions out of consumer chatbots.

5. Children's data. According to the Stanford HAI study, most developers are not removing children's input from training data collection — a gap that makes anything a child types or that you type about a child especially risky.

The same caution extends beyond chatbots to the AI agents and skills you connect to them: published research found roughly 26% of agent skills carry at least one vulnerability. That's why SkillzBench is building a free directory that security-scans, verifies, and sandbox-benchmarks agent skills before you ever deploy them — so the tools you extend your AI with meet the same standard as the data you protect.

Practical Habits That Keep You Safe: Settings, Anonymization, and Enterprise Tiers

Knowing what not to type is half the battle — the other half is building habits that protect you automatically. The good news: a handful of settings changes and substitutions, done once, cover most of your exposure.

Start with your data-sharing settings. Since August 2025, all six major U.S. AI companies — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — use chat data for training by default, and the opt-outs are often hidden in settings and not enabled by default. That means the burden has shifted entirely to you: privacy-by-choice, not privacy-by-default. Open your ChatGPT (and Claude) settings today, find the training and data-control toggles, and turn them off. Anthropic even extended data retention from 30 days to five years for users who don't opt out, so this is not a cosmetic step.

Next, adopt the right mental model: treat anything typed into free ChatGPT as permanent and public. Once information is used for training, it remains embedded in the model and cannot be fully removed — deleting the conversation doesn't undo it. Conversations may also be reviewed by humans, and OpenAI now escalates concerning chats to human reviewers. If you wouldn't paste it into a public forum, don't paste it into a free chatbot.

Then change what you actually type. Substitute anonymized or fictional examples for real data: swap "Acme Corp's Q3 pipeline of $2.4M" for "a mid-size company's quarterly pipeline." The Samsung incident shows why this matters — three engineers entered confidential source code, a transcribed meeting, and chip-yield test data into ChatGPT within 20 days. The scale of the problem is enormous: 77% of employees paste data into GenAI prompts, and 34.8% of corporate data entering AI tools is now sensitive, up from 10.7% two years earlier, according to Cyberhaven and LayerX research.

A quick anonymization habit list:

• Replace real names, birthdays, and identifiers with placeholders before pasting anything.
• Describe code problems generically rather than pasting proprietary source files.
• Round or fictionalize financials, customer counts, and strategic plans.
• Strip health, legal, and children-related details from any prompt, even casual ones.

Finally, move business use to tiers built for it. ChatGPT for Business, Claude for Enterprise, and enterprise API access do not use your inputs for general model training or human monitoring — a fundamentally different arrangement from consumer accounts, even paid ones. Given that 69% of businesses now cite AI-related data leaks as their top security worry, this is a policy decision worth making explicitly, not leaving to individual employees.

The awareness gap makes these habits urgent: 92% of Americans worry about online privacy, yet only 3% say they understand how privacy laws work, and 56% click "agree" without reading. Regulation won't rescue you soon — U.S. protections remain a patchwork of state laws. Your habits have to carry the weight.

And don't stop at chatbots. The same governance logic applies to AI agents and the skills you attach to them — published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory that security-scans, sandbox-tests, and benchmarks agent skills, plus a free MCP connector so your agent only sees the skills you choose. If you want your agent workflow to meet the same standard as your chatbot hygiene, join the SkillzBench waitlist and be first in line when it launches.

Beyond Chat: Extending the Same Caution to AI Agents and Skills

Chatbot hygiene doesn't stop at the chat window. The same caution you apply to what you type into ChatGPT now needs to extend to everything your AI touches — the tools it uses, the skills you install, and the agents you deploy across your business.

The numbers make the case. According to Cyberhaven data, 71.7% of AI tools in active use are rated high or critical risk. Meanwhile, published research found that roughly 26% of agent skills carry at least one vulnerability. A skill is just an instruction package — but instructions can hide prompt injection, data exfiltration, or sloppy handling of whatever your agent can access.

Think of it as supply-chain hygiene. You wouldn't paste your customer list into a free chatbot; you shouldn't hand an unvetted skill access to your inbox, files, or codebase either. Before adding any skill to an agent, run through a quick mental check:

• Does it need the permissions it asks for, or is it reaching beyond its job?
• Who maintains it, and how recently has it been updated?
• What data can it see or send once installed?
• Has anyone actually tested it against a baseline, or are you trusting the README?

That last question matters most, and it's the hardest to answer alone. Stanford researchers found that even innocuous inputs can trigger harmful inferences as data merges across ecosystems — and an AI chatbot privacy study notes that almost no independent research has examined the privacy practices of these emerging tools. Agents compound that risk, because they act on your behalf, not just answer questions.

This is the gap SkillzBench is being built to close. As a free, independent directory of agent skills, it verifies every skill — pinning versions, checking licenses, running security scans — and benchmarks each one in a sandbox, with and without the skill, so you can see the actual lift before anything touches your agent. A free MCP connector in development will let you equip any agent with only the skills and tools you choose, nothing more. Everything stays free; ad revenue from the blog funds the work, and money never touches scores.

The chatbot lesson was simple: assume anything you share is permanent, possibly human-reviewed, and merged into systems you don't control. The agent lesson is the same, with higher stakes. Verify before you install, equip with intent, and keep the blast radius small.

Skills, proven. Join the SkillzBench waitlist to be first in line when the directory, scanning, and benchmarking go live.