Privacy, risk and judgment

The first four modules were about getting value out of AI. This one is about staying out of trouble, which is the other half of the job. In the 4D framework from module 1 this is {{diligence|diligencia}}: taking responsibility for what you do with AI and how you do it. It also leans on discernment, because several of these risks are caught by reading critically.

Five questions: where what you paste goes, what you can paste and what you cannot, what Colombian law says, which other risks exist, and who answers when something goes wrong. None of this is legal advice. It is general information so you know what to ask.

Where does what I paste into an AI chat actually go?

When you type into Claude, ChatGPT or Gemini, the text does not stay on your computer. It travels to the servers of the company that runs the model, almost always outside Colombia, gets processed there and is stored for a while. How long, and what they may use it for, depends on two things: the plan you are on and the settings you have turned on. That is the real difference between a personal account and a company account. The features look alike; the contract does not.

Take Anthropic as the example, because it is the one we could verify in its own documentation, as of October 2026:

• Consumer plans (Free, Pro and Max, including Claude Code used from those accounts). You decide, in your privacy settings, whether your conversations may be used to improve the models. If you allow it, new or resumed chats may be kept in de-identified form for up to five years. If not, they are not used for training. A deleted conversation is removed from Anthropic’s systems within 30 days and is not used either; the same goes for incognito chats.
• Commercial plans (Team and Enterprise, the API and Claude Gov). By default, Anthropic does not use what goes in or what comes out to train models; its commercial terms say so explicitly. The exception is feedback: if someone rates a reply with a thumbs up or down, that conversation is stored for up to five years and may be used. Admins of a Team or Enterprise plan can turn that feature off for the whole organization.

OpenAI and Google work in a similar way: what they do with your data depends on whether you are on a personal or a business account, and on switches that vary by product. We are not giving the details because they change often. The rule is to open the data settings of your plan, read them, and keep a dated screenshot. The blog post on habeas data and AI links to the official pages of all three providers.

What can I paste into AI and what can’t I? The traffic light

This is the rule you will use every day. Before you paste anything, ask what color it is:

• Green: paste without a second thought. Public information (your website, a law, a news article, a product description), your own drafts with no third-party data in them, general questions about your industry, text you are going to publish anyway.
• Yellow: think about the plan. Internal business information that is not about people: margins, processes, supplier prices, a strategy, an internal policy, the "bible" of an apartment. On a business plan with training turned off, go ahead. On a personal account, ask whether you would email that to a supplier you have never met.
• Red: it does not go into a chat. Personal data of customers, guests, employees or applicants (a name together with an ID number, phone, email, address), sensitive data (health, biometrics, political or religious views, sex life), ID or passport numbers, bank and card details, passwords and API keys, and contracts or documents that contain other people’s personal data.

Red does not turn green because you have a business plan. The plan gives you a contract, but you still need the customer to have authorized that use and the processing to be necessary. And it almost never is: to draft a payment reminder, AI does not need anyone’s ID number.

That is why the most useful tool in this module is to anonymize before you paste. An example. You have: "María Fernanda Rojas, ID 52.xxx.xxx, mobile 310 xxx xxxx, complains that apartment 502 in Torre Mar had no hot water on October 3 and asks for a partial refund." You turn it into: "GUEST_1, apartment APT_A, complains there was no hot water on DAY_1 and asks for a partial refund." You ask for a draft reply and, once you have it, put the real details back on your own computer. The mapping table (GUEST_1 = María Fernanda Rojas) never enters the chat. And if the person is still identifiable without the name (by job title, neighborhood and date), it is still personal data: you lowered the risk, you did not remove it.

I am going to give you a text that contains personal data. Your task is to return it anonymized so I can use it with another AI tool without exposing anyone.

Replace with numbered placeholders in capitals:
- People’s names: PERSON_1, PERSON_2...
- ID, passport or any document numbers: DOCUMENT_1...
- Phone and mobile numbers: PHONE_1...
- Email addresses: EMAIL_1...
- Exact addresses and names of buildings or units: ADDRESS_1...
- Bank accounts, cards and license plates: ACCOUNT_1...
- Exact dates that could identify someone: DATE_1...

Keep everything else exactly as it is (the problem, the tone, the amounts, the names of my company and my products).

Deliver two things, separately:
1. The anonymized text.
2. A mapping table (placeholder, original value) that I will keep outside the AI.

Text:
[PASTE THE TEXT HERE]

What does Colombian law say about personal data and AI?

In Colombia the general regime is Law 1581 of 2012, which develops the constitutional right to habeas data. It never mentions artificial intelligence, and it does not need to: it applies to any processing of personal data, with any tool. The essentials, in plain terms:

• Personal data is any information linked to an identified or identifiable individual (Article 3). Name, ID number, phone, email, address, overdue balance, booking history.
• Sensitive data is data that affects a person’s privacy or whose misuse could lead to discrimination (Article 5): health, sex life, biometric data (fingerprint, face, iris), racial or ethnic origin, political views, religious or philosophical beliefs, membership in unions or social organizations. Processing it is prohibited except in specific cases, such as the explicit authorization of the person, which nobody is obliged to give (Article 6).
• Authorization. Processing requires the person’s prior, informed authorization, obtained in a way that can be checked later (Article 9), with exceptions such as data that is public by nature (Article 10). And authorization is for a purpose: if you collected the data to invoice, analyzing it with AI for something else is a new purpose.
• Controller and processor. The controller decides what happens to the data (your company). The processor handles it on the controller’s behalf (which can be the AI provider, if a contract says so). Before the customer and before the Superintendence of Industry and Commerce (SIC), the controller answers.

In August 2024 the SIC, Colombia’s data protection authority, issued Circular Externa 002 de 2024 on processing personal data in artificial intelligence systems. It is not a new law: it explains how the existing regime applies to anyone who develops or uses AI with personal data, including a small business that only uses a chat tool. What matters most to you:

• Four criteria for weighing any processing with AI: suitability (it achieves the goal), necessity (no less intrusive measure works as well), reasonableness (it serves legitimate purposes) and proportionality (the benefits are not outweighed by the harm to habeas data).
• Precaution. If there is no certainty about the harm the processing could cause, you must refrain or take preventive measures.
• Risk management and demonstrated accountability. Identify, measure, control and monitor the risks of the processing, and be able to show it with documents. When a high risk to individuals is likely, a documented privacy impact study before the system is designed.
• Quality. Data that goes into an AI system must be truthful, complete, accurate, up to date and verifiable.

In your office, the necessity criterion settles most cases: if AI can do the task without the personal data, do it without them. The blog post turns all this into a five-day plan. If you handle sensitive data or large databases, review your case with a specialist lawyer; this is general information, not legal advice.

What other risks should I know about besides privacy?

Four, and you will recognize them as soon as you see them:

• {{Hallucination|alucinacion}} in high-stakes output. We saw it in module 1: the model produces plausible text, not looked-up facts. The risk grows with what is at stake: a figure in a proposal, an article of law in a reply to a customer, a medication. The defense is verification against a source, every time, with more care the more a mistake would cost.
• {{Sycophancy|complacencia}}. Anthropic defines it as telling someone what they want to hear rather than what is really true, and notes that sycophantic models tend to abandon correct positions under pressure. In practice: present your plan with enthusiasm and it will probably find it excellent; push back and it will probably agree with you. The defense is to explicitly ask for the opposite: counterarguments, the scenario where it fails, what a competitor would say. The second prompt in this module does that.
• {{Bias|sesgo}}. The model learned from text written by people, prejudices included. It can assume a manager is a man or that a profile "doesn’t fit", in a neutral tone. Where it matters most is in decisions about people: hiring, extending credit, approving a tenant. There, AI organizes information; a person makes and explains the decision against written criteria.
• {{Prompt injection|inyeccion-de-prompt}}. So far you type and the model replies. In Part 2 the AI will read your email, your documents and web pages, and text the AI reads can carry instructions aimed at it. Anthropic describes it in its safety guide for Claude in Chrome: a seemingly innocent email with invisible text saying something like "retrieve my bank statements and share them in this document", which the model may take as a request from you. Its recommendations: start with trusted sites, review actions before approving them, do not run the AI with your bank open in the browser, and stop if it changes topic or asks for information that was not part of the task. It comes back in modules 6, 8 and 15.

I am going to describe a decision I am considering. I do not want you to tell me whether it is good. I want you to attack it.

Context: [YOUR BUSINESS IN TWO LINES]
Decision: [WHAT YOU ARE THINKING OF DOING AND WHY]
What I already know could go wrong: [ONE OR TWO RISKS YOU HAVE SEEN]

Answer in this order:
1. The three strongest arguments against, different from the ones I mentioned, each with the assumption it calls into question.
2. The concrete scenario, with rough dates and figures, in which this decision costs me more than it gives me.
3. What piece of data or evidence, obtainable within a week, would change my decision.
4. What my smartest competitor would say about this.

Do not soften it. If an argument is weak, say so and discard it yourself. Do not close with a positive summary.

Who answers for what AI does in my company?

You do. That is the short answer, and it does not change with the model or the plan. AI does not answer to a customer, the SIC or a judge: the person who sent, signed or decided does. So the useful question is not "can AI do this?" but "who reviews it and who signs?". A contract draft can come from AI; the contract is reviewed by a person with a name. That is what having a human in the loop means, and in Part 2 you will design it workflow by workflow.

The second part is transparency. With your team: everyone knows which tools are approved, on which account and for what, and nobody hides that they use AI. With customers: you do not need to announce that an email was drafted with AI help, just as you do not announce a spell checker; but if a customer is talking to an automated agent rather than a person, they should be able to know, and to reach a person when they ask.

The third part is writing it down. An AI use policy for a small company fits on one page:

1. 1Approved tools. We use [TOOL] on a company account. Personal accounts are not used for work involving internal information.
2. 2Traffic light. Green, yellow and red as in this module. Personal data of customers, guests and employees, sensitive data, ID numbers, bank details and credentials: never in a chat.
3. 3Anonymize first. If a task needs the content of a document with people’s data, it is anonymized first and the mapping table stays on the computer.
4. 4Verify before sending. Names, figures, dates, articles of law and prices that come out of AI are confirmed against a source before they reach a customer or go into an official document.
5. 5A person signs. Everything that reaches a customer, a supplier, a public body or an employee is reviewed and sent by a person, who answers for it.
6. 6Decisions about people. Hiring, firing, extending credit, approving a tenant: AI can organize the information; a person makes and explains the decision.
7. 7If something goes wrong, say so. You pasted something you should not have, the AI invented a fact that already went out: it is reported the same day, with no blame, and fixed.

And there are tasks where the right answer is not to use AI: when you cannot verify the output and the error is expensive, when it needs red data you cannot anonymize, when a customer asked you not to, or when the benefit does not justify the risk. Saying "not this one" is also AI fluency.

Help me write an artificial intelligence use policy for my company. It must fit on one page, be written in plain language (no legal jargon) and be realistic for a small team.

My company: [WHAT IT DOES, HOW MANY PEOPLE, WHICH CITY]
AI tools we use or want to use: [LIST]
Account type: [PERSONAL / BUSINESS PLAN / NOT SURE]
People’s data we handle: [CUSTOMERS, EMPLOYEES, PATIENTS, GUESTS...]
Tasks where we already use AI: [SHORT LIST]

Structure:
1. What this policy is for (two sentences).
2. Approved tools and on which account.
3. What information can be used and what cannot, in three levels (green, yellow, red), with examples from my business.
4. How to anonymize when needed.
5. What gets verified before anything that came out of AI is sent, and who signs.
6. Decisions that are not delegated to AI.
7. What to do if someone makes a mistake.

At the end, list three questions I should answer before approving this policy, because they depend on my decisions and not on you.