At its core, how does a large language model write an answer?

It predicts likely next words, one piece at a time, from patterns learned in training. A language model generates text by repeatedly predicting the next token from patterns it learned during training. Apps can add search or your documents, but the core is prediction, not lookup.

An AI answer confidently cites a court case, with a name, a date and a quote. What should you do before relying on it?

Check that the case exists and says what the AI claims. Models can invent convincing details, often called hallucinations. In Mata v. Avianca (2023) a judge fined two lawyers $5,000 for citing cases ChatGPT had made up. Asking the same model to confirm isn’t a real check.

Does a chatbot’s model learn from your conversation while you’re chatting?

No: the model is fixed; it uses the conversation so far, plus anything the app saves to memory. The model’s weights don’t change as you chat. Apps can store things about you in a memory feature, and providers may use saved conversations to train future models later, depending on your settings. That’s Level 3.

You type a question into ChatGPT. Where does the model that answers it run?

On OpenAI’s servers, which receive your prompt. ChatGPT is a cloud service: your prompt is sent over the internet to OpenAI’s servers, answered there and sent back. That’s why the provider’s storage and training policies apply to everything you type.

Which of these does OpenAI say its models are trained on?

Publicly available content on the internet / Data from partners / Information users, human trainers and researchers provide. OpenAI lists public internet content, partner data, and information from users, human trainers and researchers. It says it avoids paywalled content and the dark web, and filters out things like hate speech and spam.

What does “open weights” mean?

The model files are published, so you can download and run the model on your own hardware. Open weights means the trained model itself is published, not necessarily the training data. You can run it on your own machines; the license sets the rules, and MIT or Apache 2.0 licenses allow commercial use.

This switch is on by default for individual ChatGPT accounts. What does it allow?

OpenAI may use your conversations to train its models. “Improve the model for everyone” lets OpenAI use your chats for training. You can turn it off, and it then applies to new chats. Rating an answer with a thumbs up or down can still let OpenAI use that conversation.

Which of these does OpenAI not use for training by default?

ChatGPT Business and Enterprise / The API, used by developers. Business, Enterprise, Edu and API data are excluded by default. Free and Plus are individual accounts, used for training unless you opt out.

You allowed a cloud AI to train on your chats, then change your mind and opt out. What happens to the data already used to train a released model?

It stays in models already trained. Anthropic says plainly that opting out stops future training but doesn’t remove data from models that have already been trained. A model can’t simply forget what it learned.

In 2025, shared ChatGPT conversations like these turned up in Google search. How did they get there?

Users had ticked an opt-in “make this chat discoverable” box when sharing. The share dialog had an opt-in box that let search engines index the chat. Many people ticked it without realising what it meant; about 4,500 chats were found and OpenAI removed the feature. (The results pictured are an illustration.)

Which of these prompts would put confidential company data on someone else’s servers?

“Debug this function from our internal codebase: …” / “Summarize these board minutes before we announce the acquisition.”. Source code and unannounced deal papers are exactly what Samsung’s engineers ran into in 2023. General questions are fine anywhere; company secrets belong on a model you run yourself.

You press “share” on a chatbot conversation to send it to a colleague. What’s the safest assumption?

Anyone with the link can read it, and it could end up public or in search results. A share link is a public page for anyone who has it. In 2025 more than 370,000 shared Grok chats turned up in Google without users being warned. Treat a shared chat as published.

In May 2025 a US court ordered OpenAI to keep ChatGPT chats, including deleted ones. What case was it for?

A copyright lawsuit brought by The New York Times. It was a copyright lawsuit. Millions of users who had nothing to do with it had their chats preserved, and in January 2026 a judge upheld an order to hand over a sample of 20 million de-identified logs.

Are your conversations with a consumer AI chatbot confidential, like conversations with your lawyer?

No: courts have ruled they aren’t privileged, so they can be demanded as evidence. In United States v. Heppner (February 2026) a court held a defendant’s own Claude chats weren’t protected by attorney-client privilege. OpenAI’s CEO had already warned that ChatGPT conversations can be required in a lawsuit.

When does OpenAI say it may refer a conversation to law enforcement?

When human reviewers find an imminent threat of serious physical harm to others. That’s OpenAI’s stated threshold, after human review; it says it doesn’t refer self-harm cases. Separately, police can obtain chats with warrants and other legal requests, as OpenAI’s transparency report shows.

Which setup keeps working with the internet unplugged?

An open-weight model running in your office. Once the model is downloaded, it runs entirely on your own hardware. No connection, no outside servers.

A Toronto firm uses a US cloud AI provider that stores its data in a Canadian data centre. Can a US legal order still reach that data?

Yes: the CLOUD Act covers data a US provider controls, wherever it’s stored. The CLOUD Act (2018) requires US providers to disclose data in their possession, custody or control whether it’s stored inside or outside the United States. Where the server sits matters less than who controls it.

Your team runs an open-weight model on a Mac Studio in the office. Who can use your prompts to train a model?

Nobody, unless you decide to. The weights are just a file on your machine: running the model doesn’t send your prompts to its publisher or to other users. Nothing leaves your hardware, so there’s nothing for anyone else to train on unless you choose to.

Test your AI knowledge

How well do you really know AI?

Six quick levels on how AI works and where your prompts really go. Score 80% to earn the AI Essentials & Privacy Professional Certificate, see your rank, and challenge your team to beat you.

6 levels
18 questions
About 10 min
Printable certificate

Start the challenge

Free, no sign-up needed. Your answers stay in your browser.

Faraday Machines

Professional Certificate

AI Essentials & Privacy

This certifies that

Your Name

has earned the AI Essentials & Privacy Professional Certificate by passing the Faraday Machines Test Your AI Knowledge challenge with a score of 100% (18 of 18), demonstrating an understanding of how AI models work and are trained, how cloud AI services handle personal and business data, when conversations can be retained, reviewed or requested by courts, and how open-weight models run privately on your own hardware.

DateYour completion date

Credential IDFM-AIEP-0000000

Issued byFaraday Machines · Toronto

Pass with 80% and download yours as a PDF.

The lessons

Everything the challenge asks about, in one place. Each level of the game opens with one of these. Start the challenge.

Level 1 · What AI is

It writes by predicting what comes next

Chatbots like ChatGPT, Claude and Gemini, and open models like Qwen or DeepSeek, are large language models. Knowing how they produce an answer explains both what they’re great at and where they go wrong.

A large language model learns patterns from enormous amounts of text. To answer, it writes one small piece at a time (a token, roughly a word or part of one), each time predicting what’s most likely to come next. (The percentages in the picture are illustrative.)
It doesn’t look facts up in a database unless the app connects it to search or your documents, so it can give fluent, confident answers that are wrong. In June 2023 a US judge fined two lawyers $5,000 for filing a brief that cited court cases ChatGPT had made up.
Chatting doesn’t retrain the model: its weights are fixed once it’s trained. It works from what’s in the current conversation, plus anything the app saves with features such as ChatGPT’s memory.

Sources: Google for Developers: Introduction to large language models · LawSites: Court sanctions lawyers who filed fake cases from ChatGPT (Mata v. Avianca, June 2023) · OpenAI Help Center: Memory FAQ

Level 2 · How AI is trained

Training makes the model; the weights are the result

Before anyone can chat with it, a model is trained on huge amounts of data. What comes out is a file of billions of numbers called weights, and where that file runs decides who sees your prompts.

OpenAI says its models learn from publicly available internet content, data from partners, and information that users, human trainers and researchers provide. It says it avoids paywalled content and the dark web.
Training happens in stages: pre-training, predicting the next token across vast amounts of text, then post-training, where people rate and correct answers to make the model more helpful and safer.
Cloud AI keeps the weights on the provider’s servers, so every prompt travels there. Open-weight models publish the weights, so you can download and run them yourself; even OpenAI released two, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 license in August 2025. The license decides what you may do with them.

Sources: OpenAI: How ChatGPT and our foundation models are developed · InfoQ: OpenAI releases gpt-oss open-weight models under Apache 2.0 · Faraday Machines: Open-weight AI license tracker

Level 3 · Your chats

Do they learn from what you type?

Cloud providers improve their models with data, and conversations are data. Whether yours are used depends on your plan and your settings, and the defaults aren’t always in your favour.

OpenAI may use conversations from individual ChatGPT accounts to train its models by default. You can switch it off in Settings, Data controls, “Improve the model for everyone”. Business, Enterprise, Edu and API data are not used by default.
In August 2025 Anthropic asked Claude Free, Pro and Max users to choose whether their chats may train its models; the switch in the pop-up was preset to On. If you allow it, chats can be kept for five years instead of 30 days, and opting out later doesn’t remove data from models already trained.
Google says human reviewers, including third parties, read some Gemini chats, and reviewed chats can be kept for up to three years even if you delete your activity.
In May 2026 Canada’s federal Privacy Commissioner and three provincial commissioners found OpenAI’s collection of data to train ChatGPT overly broad and out of line with Canadian privacy law.

Sources: OpenAI Help Center: How your data is used to improve model performance · Anthropic: Updates to consumer terms and privacy policy (Aug 28, 2025) · Google: Gemini Apps Privacy Hub · Office of the Privacy Commissioner of Canada: OpenAI investigation findings (May 6, 2026)

Level 4 · Leaks

Once it’s typed in, it’s out of your hands

Most data exposure isn’t a hack. It’s ordinary features and ordinary people: a share button, a bug, an engineer pasting code to get help.

In July 2025, shared ChatGPT conversations with an opt-in “make this chat discoverable” box ticked showed up in Google. Reporters found about 4,500, and OpenAI removed the feature around August 1, 2025.
In August 2025 Forbes found more than 370,000 Grok conversations indexed by Google after users pressed “share”, without being warned they’d become public.
In May 2023 Samsung temporarily banned generative AI on company devices after engineers uploaded internal source code to ChatGPT. Its worry: the data sat on outside servers, hard to retrieve or delete.

Sources: Search Engine Journal: OpenAI pulls shared ChatGPT chats from Google search · Forbes: xAI published hundreds of thousands of Grok conversations · Bloomberg: Samsung bans ChatGPT use by staff after leak (May 2, 2023)

Level 5 · Courts and police

Your chats are records, and records can be demanded

A cloud conversation is stored by a company, so it can be preserved, subpoenaed or searched like any other business record. That can happen in cases you have nothing to do with.

In May 2025, in a copyright case brought by The New York Times, a US court ordered OpenAI to preserve ChatGPT logs, including chats users had deleted. The general order ended in September 2025, but in January 2026 a judge upheld an order to hand over a sample of 20 million de-identified ChatGPT logs.
There’s no legal privilege: Sam Altman said in July 2025 that ChatGPT chats have no confidentiality like a lawyer or doctor, and in February 2026 a US court ruled a defendant’s own Claude chats weren’t protected by attorney-client privilege.
Police can ask too. OpenAI reported 75 government requests for user content in the second half of 2025 and disclosed data in 62. In October 2025 Forbes reported the first known federal search warrant for OpenAI user data based on prompts.
OpenAI also says conversations in which users plan to harm others go to a human review team, which may refer an imminent threat of serious physical harm to law enforcement.

Sources: OpenAI: How we're responding to The New York Times' data demands · Bloomberg Law: OpenAI must turn over 20 million ChatGPT logs, judge affirms (Jan 5, 2026) · TechCrunch: Sam Altman warns there's no legal confidentiality when using ChatGPT as a therapist · Orrick: Court rules AI conversations are not privileged (United States v. Heppner) · OpenAI: Government requests for user data, July to December 2025 · Stanford CIS: Questions to ask about the ChatGPT warrant (Oct 2025) · Futurism, quoting OpenAI's "Helping people when they need it most" (Aug 2025)

Level 6 · Local AI

Open weights in your own office

Run an open-weight model on your own hardware and the questions above mostly disappear: there’s no provider to train on your prompts, keep them, review them or be ordered to hand them over.

A local model answers on your machine and keeps working with the internet unplugged once it’s downloaded.
Where data sits isn’t the whole story. Under the US CLOUD Act, US providers must disclose data in their possession, custody or control wherever it’s stored, including Canadian data centres.
Canada is changing its own rules too: Bill C-22, the Lawful Access Act, passed the House of Commons in June 2026 and is before the Senate. It would make it easier for police to obtain subscriber information.

Sources: US Department of Justice: The CLOUD Act (18 U.S.C. § 2713) · InfoQ: OpenAI releases gpt-oss open-weight models under Apache 2.0 · Faraday Machines: Open-weight AI license tracker

Frequently Asked Questions

How does AI like ChatGPT actually work?

ChatGPT, Claude, Gemini and open models like Qwen are large language models. They learn patterns from huge amounts of text during training, then write answers one small piece (a token) at a time by predicting what is most likely to come next. They don’t look facts up unless an app connects them to search or documents, so they can sound confident and still be wrong.

Does ChatGPT train on my conversations?

For individual ChatGPT accounts, yes by default: OpenAI may use conversations to train its models unless you turn off “Improve the model for everyone” in Data controls. OpenAI says it does not train on ChatGPT Business, Enterprise or Edu data, or on API data, by default. Temporary Chats are not used for training.

Can police get my ChatGPT conversations?

They can request them. OpenAI reported 75 government requests for user content in the second half of 2025 and disclosed data in 62. Separately, OpenAI says it may refer a conversation to law enforcement when human reviewers find an imminent threat of serious physical harm to others. Courts have also ordered chat logs preserved and handed over in civil lawsuits.

What is the difference between cloud AI and local AI?

Cloud AI runs on a provider’s servers, so every prompt is sent to them and handled under their policies. Local AI runs an open-weight model on your own hardware, so prompts never leave your device or office, nothing is used to train someone else’s model, and it works offline.

Is the AI Essentials & Privacy Professional Certificate an official qualification?

It’s a certificate of achievement from Faraday Machines for passing the Test Your AI Knowledge challenge with a score of 80% or more. It shows you understand how AI models work and are trained, and how cloud and local AI handle your data; it isn’t an accredited qualification.

How do I compare my score with others?

Your result shows your rank: Beginner, Informed, Professional or Expert. Share the challenge with friends or your team and compare ranks, points and best streaks.

Sources

[1] Google for Developers: Introduction to large language models. developers.google.com

[2] LawSites: Court sanctions lawyers who filed fake cases from ChatGPT (Mata v. Avianca, June 2023). www.lawnext.com

[3] OpenAI Help Center: Memory FAQ. help.openai.com

[4] OpenAI: How ChatGPT and our foundation models are developed. help.openai.com

[5] OpenAI Help Center: How your data is used to improve model performance. help.openai.com

[6] OpenAI Help Center: Chat and file retention policies in ChatGPT. help.openai.com

[7] Anthropic: Updates to consumer terms and privacy policy (Aug 28, 2025). www.anthropic.com

[8] Google: Gemini Apps Privacy Hub. support.google.com

[9] Office of the Privacy Commissioner of Canada: OpenAI investigation findings (May 6, 2026). www.priv.gc.ca

[10] Search Engine Journal: OpenAI pulls shared ChatGPT chats from Google search. www.searchenginejournal.com

[11] Forbes: xAI published hundreds of thousands of Grok conversations. www.forbes.com.au

[12] Bloomberg: Samsung bans ChatGPT use by staff after leak (May 2, 2023). www.bloomberg.com

[13] OpenAI: How we're responding to The New York Times' data demands. openai.com

[14] Engadget: OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions. engadget.com

[15] Bloomberg Law: OpenAI must turn over 20 million ChatGPT logs, judge affirms (Jan 5, 2026). news.bloomberglaw.com

[16] TechCrunch: Sam Altman warns there's no legal confidentiality when using ChatGPT as a therapist. techcrunch.com

[17] Orrick: Court rules AI conversations are not privileged (United States v. Heppner). www.orrick.com

[18] Stanford CIS: Questions to ask about the ChatGPT warrant (Oct 2025). cyberlaw.stanford.edu

[19] OpenAI: Government requests for user data, July to December 2025. cdn.openai.com

[20] Futurism, quoting OpenAI's "Helping people when they need it most" (Aug 2025). futurism.com

[21] US Department of Justice: The CLOUD Act (18 U.S.C. § 2713). www.justice.gov

[22] InfoQ: OpenAI releases gpt-oss open-weight models under Apache 2.0. www.infoq.com

[23] Faraday Machines: Open-weight AI license tracker. www.faradaymachines.com

Keep Your Prompts in Your Own Office

Faraday Machines installs open-weight models on Mac Studio hardware in Toronto, so your team gets capable AI with no prompts leaving the building, no training on your data and no per-question bill.

Talk to Us

Free private AI assessment