Tech AI News - 테카이

18. Where does the code you upload to a free AI tool go — read the terms

· 7 min read

This post was translated from the Korean original by AI.한국어 원문 읽기 →

In the middle of a busy workday, pasting a line of code or a document into a free AI tool feels like a harmless choice. But that input can end up as training data, and "it wasn't used for training" doesn't mean "it was safe" — neither is obvious unless you actually open the terms.


Background — why this question is coming up now

Unless a company bans it, employees reach for whichever AI tool is closest at hand. Most of these are free plans, and few people read the terms all the way through when signing up. The problem is that a clause allowing "your input may be used to improve the model" shows up far more often in free-plan terms than in paid ones.

What recently sharpened this picture was a policy change at Anthropic. Starting August 28, 2025, Anthropic let consumer-plan Claude users (Free, Pro, Max) choose directly whether their data would be used for model training, and said that unless users opted out, the retention period — previously 30 days — would extend to up to five years (TechCrunch, 2025.8.28). Business-use Claude for Work, Enterprise, and API traffic were excluded from this change.

A past case illustrates the same pattern. In 2023, engineers at Samsung Electronics' semiconductor division were caught by the company's internal security team after pasting source code and meeting notes into ChatGPT during work, prompting the company to ban generative AI use on its internal network entirely (Forbes Australia, 2023.5). This incident is still cited often, because it shows the risk started the moment sensitive content left for an external server — regardless of whether it was ever used for training.

Key data & current state

ItemStatusSource
OpenAI personal ChatGPT (Free, Plus, Pro)Data sharing is on by default for personal workspaces; users must opt out to exclude their data from training. Business, Enterprise, Edu, and API default to excluding trainingOpenAI Help Center, checked 2026.11
Anthropic Claude consumer plansAfter the 2025.9.28 opt-out deadline, data not opted out is used for training, and retention extends to up to five years. Claude for Work, Enterprise, and API are excludedTechCrunch, 2025.8.28
Google Gemini personal accounts (free)Unless "Gemini Apps Activity" is turned off, conversations are used for product improvement and may be reviewed by humans. Conversations flagged for review may be kept for up to 3 years even after activity is deleted. Workspace and enterprise accounts are excluded from both training and human reviewGoogle Gemini Apps Privacy Notice (tracked by ConductAtlas), checked 2026
Mistral AI free subscriptionUnless opted out, input and output are used for model training by defaultConductAtlas provision tracker, checked 2026.5

Interpretation: None of these four companies split cleanly into "paid = safe, free = risky." OpenAI's default even varies by whether a paid plan is personal (Plus, Pro) or business (Business, Enterprise), and Google and Anthropic also draw their line between consumer and business use, not price. "What contract form you use" is a much closer match to the real rule than "how much you pay."

Deeper analysis

1) Why training clauses are common in free plans

Running a service for free means recovering the cost somewhere. The most common approach AI companies take is using free users' input and output to improve the model. The terms usually include phrasing like "aggregated and anonymized data," but that only means the data is processed so individuals can't be identified — it's not a guarantee that the raw code or sentences themselves never enter the training pipeline.

Conversely, in the tiers companies pay for (API, Business, Enterprise), excluding training is often the default. OpenAI's help center explicitly states that Business, Enterprise, Edu, and API offerings all default to not using input and output for training (OpenAI Help Center). Breaking a promise of "not used for training" to a customer under contract carries immediate legal risk.

2) The trap in assuming "paid means safe"

What gets missed here often is that paid doesn't always mean training-excluded by default. For OpenAI, ChatGPT Plus and Pro are paid plans billed monthly, but because they're personal workspaces, data sharing is on by default — users have to turn it off themselves in settings to be excluded from training (OpenAI Help Center). In other words, the real line isn't "did you pay" but "is this a personal contract or an organizational one."

There are also cases where a clause moves in the opposite direction. ConductAtlas's tracking of Mistral AI's privacy policy records that a clause that once existed — stating that "Le Chat Enterprise or paid API versions are not used for training" — has recently been removed (ConductAtlas). Rather than assuming a paid plan is safe and moving on, it's safer to go directly to the official data policy page for the plan you're on and find the sentence that starts with "by default."

3) Excluding data from training and keeping it confidential are separate problems

Even if you get a promise that your data won't be used for training, that doesn't mean it's stored safely or that no one ever looks at it. Google states that among consumer Gemini conversations that weren't opted out, items flagged for human review may be retained separately for up to 3 years even after the user later deletes their activity history (per the Gemini Apps privacy notice tracked by ConductAtlas). "Not used for training," "no one looks at it," and "deleted quickly" are three different promises.

As the Samsung case shows, the risk already begins the moment sensitive content is sent to an external server in the first place (Forbes Australia, 2023.5). Domestic Korean media have made a similar point. One outlet reported that turning off the training-use option on ChatGPT or Claude doesn't mean data is deleted immediately or excluded from legal discovery, and that U.S. courts have also ruled inconsistently, case by case, on whether AI conversations deserve confidentiality protection (Korea Business Review, 2026.9.12). Analysis: opting out of training is closer to a request not to let your data answer someone else's question — it isn't a comprehensive safeguard that also removes the risk of retention, leaks, or legal disclosure.

4) Terms can change at any time

Just as Anthropic quietly shifted its default — from excluding consumer-plan data from training by default to the new opt-out model it announced in August 2025 (TechCrunch, 2025.8.28) — terms can change quietly to fit a company's business strategy. ConductAtlas's May 2026 comparison of the terms of ten companies — OpenAI, Anthropic, Google Gemini, GitHub Copilot, Midjourney, xAI, Perplexity, Cursor, Meta, and Hugging Face — found that training-related clauses vary widely by platform, plan, and login status, and even carry exceptions within the same company. For instance, it notes that Anthropic may still use conversations flagged as "safety review" for model improvement even when a user has opted out, and that xAI only offers an opt-out option to logged-in users (ConductAtlas, 2026.5.12). There's no guarantee that a clause you check today will still hold after the next update.

In practice — 3 things to check before putting company code or customer data into a free AI tool

  1. Check directly on the official data policy page whether the plan you're using is "training-excluded by default." Consumer (free, personal paid) and business/API often have different defaults even within the same company, so look for the sentence starting with "by default" in the service's actual data policy or privacy notice — not a blog summary.
  2. Even after opting out, check retention periods, human review, and exceptions too. Being excluded from training doesn't mean protection from deletion, access, or legal disclosure, so it's safer to never input irreversible material like company code or customer PII in the first place.
  3. Route sensitive material through an environment covered by your organization's AI-use policy or a business contract (DPA). If there's no internal policy, check with security or legal before inputting anything, and note the date you checked the terms — since terms can change anytime, always attach the caveat "as of the date I checked."

References

  • #AI
  • #artificial intelligence
  • #AI terms of service
  • #data privacy
  • #AI security
  • #ChatGPT
  • #information security