AI has changed the value of your data

By Angus Lewington, Alien IT · Last updated: 9 September 2026 · Facts checked: 9 September 2026

For years the simplest explanation for data collection was advertising: a company watches what you do so it can guess what you might buy. That explanation is now too small, because ordinary records have become instructional data.

An archive of email, chat, code, documents and support tickets holds more than topics and keywords. It holds examples of how people explain, decide, negotiate, diagnose, fix mistakes and get work done. Those examples can be used to train, test and ground AI systems.

An airline’s internal memory went to auction

In August 2026, Google won a bankruptcy auction with a US$10 million bid for a de-identified copy of internal data from Spirit Aviation Holdings, the parent of Spirit Airlines. The proposed transfer was unusually broad. Reporting and court materials described email, Microsoft Teams items, documents, spreadsheets, calendars, source code and operational data.

Reporting on the auction said the dataset could help Google improve its products and AI models. Mercor, another AI-focused company, was named backup bidder at US$7.5 million.

This was not a purchase of the airline. Google won an auction for defined data assets, and the sale still needed court approval. Spirit said the copy would be de-identified and would exclude customer and personally identifiable information. The Association of Flight Attendants-CWA objected that linked records could still let someone reconstruct information about individuals or small groups. The court postponed the approval hearing, and the proposed transfer remained contested in the latest docket material available when this article was checked on 9 September 2026 (Reuters report via Investing.com, union court filing, court docket index).

The dispute matters because removing names does not settle every privacy, confidentiality or re-identification question, especially when the usefulness of a dataset depends on keeping the relationships between its records.

What makes workplace data different from public text?

Much of the public internet is duplicated, promotional, incomplete, machine-generated, or stripped of the context around a decision.

Internal work records can hold something else:

• the original problem
• the incomplete information people had at the time
• discussion between people in different roles
• attempts that failed
• the decision that was made
• what happened next
• the correction or escalation

A support ticket can connect a symptom to a diagnosis and a fix. A code review can show a proposal, an objection and a safer version. A calendar plus an email thread can map the sequence that produced a decision. Operational records can connect behaviour to outcomes.

None of this means the data is clean, fair, accurate or lawful to reuse. It explains why authentic workflow records may be sought for improving and testing models.

Six different things people call “AI training”

The phrase is used loosely. Keep these processes separate.

1. Pre-training

A model learns broad statistical patterns from a very large body of text. This is what most people picture when they hear “training data”.

2. Fine-tuning

An already trained model is adjusted with a smaller, task-specific dataset. The goal might be a style, a subject area, an output format or a behaviour.

3. Human feedback and preference work

People compare or rate outputs, and the results help shape how the model behaves. A thumbs-up, a correction or a reviewed conversation can be useful here.

4. Evaluation

A dataset is used to test whether a model succeeds, fails, gets worse or behaves safely. Evaluation material may never change the base model and can still be valuable.

5. Retrieval-augmented generation

A search step finds passages in your documents and pastes them into the prompt at answer time (this is RAG). Standard RAG does not normally change the model itself, but the model service still processes the passages it is sent unless the whole system runs locally.

6. Service operation and safety review

A provider can keep or review material to show you your chat history, investigate abuse, meet legal duties, fix bugs or protect users. “Not used to train the foundation model” does not mean “never stored, logged, reviewed or processed”.

A useful privacy explanation names the actual process.

What data is involved?

For consumers, AI services may receive prompts, responses, uploaded files, images, voice, feedback, account information, device details, content from connected apps and usage metadata.

For businesses, the potential scope is much wider:

• email and attachments
• Teams or Slack messages
• documents and spreadsheets
• calendars and meeting records
• customer and CRM history
• support tickets and call summaries
• source code and repositories
• internal policies and procedures
• pricing, inventory and operational records
• model inputs, outputs and ratings

Some enterprise products ground their answers in your existing data stores, using the permissions people already have. That is powerful, and it makes old access mistakes easier to exploit: if a staff member can already open a forgotten SharePoint folder they should never have seen, an AI assistant may make its contents much easier to find.

How does information reach an AI system?

The obvious path is a prompt. Other paths are easier to miss.

Uploads. A document, image or spreadsheet becomes input, even if you only asked for a short summary.

Connected applications. Email, storage, code repositories, calendars and web search can be queried through connectors or agents.

Feedback. Rating an answer can attach the surrounding conversation under a different retention or review rule.

System logs. Prompts and responses may be kept for history, audit, safety or debugging even when they are not used to improve the base model.

Retrieval indexes. A private document collection can be copied, cut into pieces and stored again inside a search index. Deleting the original does not prove it has gone from every derived store.

Third-party tools. A browser extension or AI front end may send data to its own provider as well as to the model API.

The privacy statement for the model alone is not a map of the product.

Consumer, business and API policies are not interchangeable

A line like "ChatGPT uses your data" or "Claude does not train on chats" means nothing until it names the product, the setting and the date.

Policies checked 9 September 2026; recheck the linked pages before relying on them.

OpenAI

OpenAI says content submitted to services for individuals may be used to improve model performance depending on user settings (OpenAI consumer-data FAQ). It says content from the API, ChatGPT Business and ChatGPT Enterprise is not used for training by default unless the customer explicitly opts in (OpenAI enterprise privacy). Its controls allow users to turn off model improvement, and Temporary Chat has separate history, memory, training and limited safety-retention behaviour (OpenAI Data Controls).

Anthropic

Anthropic says consumer users choose whether chats and coding sessions may be used for model improvement; permitting that use changes the published retention period (Anthropic consumer update). It says Claude for Work and API chats are not used for training unless the organisation joins its Development Partner Program or uses another explicit route (Anthropic commercial training). Default API retention and approved zero-data-retention arrangements have product, safety and legal qualifications (Anthropic commercial retention).

Google Gemini

Google’s Gemini Apps notice describes prompts, files, screens, connected-app content and other information. It allows activity and temporary-chat controls, while documenting separate periods and purposes for ordinary activity, temporary chats, feedback and human-reviewed conversations (Gemini Apps Privacy Hub).

Microsoft Copilot

Microsoft says prompts, responses and Microsoft Graph data in Microsoft 365 Copilot are not used to train foundation models. It also says prompt/response interactions can be stored for user history and organisational audit, eDiscovery and retention controls (Microsoft). Agents and external connectors can have their own terms.

These are summaries, not a substitute for reading the contract. Policies change, and enterprise agreements, regions and the features you switch on can alter the answer.

What this means in Australia

For an Australian organisation covered by the Privacy Act, a provider’s “no training” statement is not the end of the analysis. Personal information can sit in the input, output, chat history, connector data or administrator records even when it never changes a foundation model.

The Office of the Australian Information Commissioner tells organisations choosing commercial AI to do their due diligence, understand who can access the inputs and the generated output. It also tells them to build privacy in from the start and keep reviewing the product over its life. As a matter of best practice, it recommends not entering personal information, and especially sensitive information, into publicly available generative AI tools (OAIC commercial-AI guidance).

That is regulator guidance, not a blanket ban and not legal advice for your situation. Coverage and obligations depend on the entity, the information and the purpose. The practical point is clear: choose an approved product tier, send as little as the task needs, write down the full data path, and get legal advice for high-stakes uses.

Who can receive the data?

Depending on how the system is built, the recipients can include:

• the AI provider
• cloud infrastructure providers and their subprocessors
• human safety or quality reviewers
• the organisation that administers a business account
• connected search, storage or application services
• the maker of a third-party interface or plug-in
• a local managed-service provider
• legal authorities where valid legal process applies

Local hosting removes some recipients only if every relevant component is local. It can add a different recipient if a support provider has remote access. Ask which people and systems can reach which layer, rather than whether the system is cloud or local.

Why the data is valuable

Data can support at least five kinds of value:

1. Capability: teaching or grounding systems in a field or a workflow.
2. Quality: finding failure cases, corrections and better responses.
3. Safety: recognising prohibited or risky behaviour.
4. Product design: learning which tasks users attempt and where they struggle.
5. Competitive knowledge: understanding processes, markets and operating patterns.

Not every company pursues all five. A privacy notice or contract may restrict use. The point is that advertising is no longer the only plausible reason an organisation might want authentic human data.

What can be inferred when datasets are combined?

An email body can have the sender’s name removed and still keep dates, job role, location, the reply chain and references to distinctive events. Payroll, calendars and team structures narrow the possibilities further. Code commits and issue threads can connect a pseudonymous author across systems.

Combining records can also reveal facts about the organisation: who approves unusual refunds, how maintenance problems escalate, which routes or products fail, how long decisions take, what staff disagree about, and which internal controls exist mainly on paper.

This is why de-identification has to be judged against the whole released dataset, the links kept within it and the other information reasonably available, not just whether one name field was removed.

When the company changes, the data can remain

23andMe shows the consumer side of the same custody problem. The genetic-testing company entered Chapter 11 bankruptcy in March 2025. A court later approved a US$305 million transaction with TTAM Research Institute, an affiliate led by co-founder Anne Wojcicki, and the acquisition closed on 14 July 2025.

The sale covered substantially all of the operating assets and drew objections and regulator scrutiny over genetic and other personal information. The final arrangements kept customers’ deletion and research opt-out rights and added restrictions on later transfers. The FTC had warned the bankruptcy trustee that 23andMe’s existing privacy promises should travel with the data (FTC letter, SEC closing filing, filed privacy safeguards).

Bankruptcy does not cancel privacy law, but data custody can become a central issue in insolvency and asset sales. Read what a privacy policy says about mergers, acquisitions, bankruptcy and onward transfer before the crisis, not during it.

What this means for an ordinary person

You do not need to avoid AI. Match the tool to the information.

Low-stakes brainstorming in a consumer service you have configured is different from uploading a medical history, privileged legal advice, a client database or your employer’s confidential code. A business account with contractual controls is different from a staff member’s personal free account. A local model is different from a local-looking interface that quietly calls an external API.

Use one simple rule: if disclosing the input, the output or even the fact that you did the task would cause real harm, do not proceed until you know the exact data path and the approved product.

What you can do about it

For a business, the useful controls start before a prompt is sent.

Classify data before choosing tools

Create categories people can actually use: public, internal, confidential, highly restricted. Give examples. State which AI products and features are approved for each.

Separate consumer and business accounts

Do not assume a personal subscription inherits enterprise terms. Use centrally managed identity, settings, contracts and offboarding.

Inventory connectors

Record which tools can reach email, drives, code, calendars, web search and third-party agents. Review connector permissions and subprocessors, not only the base model.

Minimise the input

Remove names, client identifiers, secrets and irrelevant pages when the task does not need them. “The provider has good security” is not a reason to send information the task does not need.

Configure model-improvement and retention controls

Record the setting, who can change it and the date you checked. A policy page without the actual workspace configuration is incomplete evidence.

Keep selected workloads local

Local inference and private RAG can keep routine processing on infrastructure you control. Audit the model runtime, interface, embeddings, index, logs, backups and any support access. Local hosting also moves patching and recovery onto you. The local AI and private RAG guide walks through that decision.

Preserve an exit path

Export prompts, approved outputs, evaluation sets and configuration where it makes sense. Know how to remove a connector, revoke credentials and delete derived indexes.

Limits of those controls

• Policy settings can change or be misconfigured.
• A business contract does not stop an authorised employee pasting data into a different tool.
• Redaction can fail when the surrounding context re-identifies people.
• Local systems can be compromised, stolen or neglected.
• Deletion may be subject to backups, security retention or the law, so it is not instant erasure from every system.
• RAG can expose documents if permissions are wrong.
• The most private option may not produce good enough results for the task.

Controls reduce exposure without removing risk.

The practical conclusion

AI has made authentic work records more useful because they capture how people reason and act, not merely what topic they discussed.

It calls for custody: know where each record can go and who can reach it there.

Know which data enters which product. Distinguish training from retrieval, evaluation, safety review and storage. Keep selected sensitive workloads in-house where practical, and use hosted services under the right tier and terms when they are the better tool.

Your data does not become unimportant because it looks ordinary. Sometimes the ordinary record is exactly what teaches a machine how real work gets done. Alien IT’s version of the in-house approach is the Alien Private AI Server.

Not sure which jobs can go to a hosted AI service and which should stay in-house? Talk it through with Alien IT. Call 02 9707 0999 or use the contact page.

Frequently asked questions

Related reading