Insights · 6 min read
What Makes Data Worth Less to AI Labs: Seven Value Killers
Not every archive is an asset. These are the reasons buyers mark data down, and which ones you can fix.
Most writing about selling data to AI companies is optimistic. This is the other half. Some archives are worth little, and some are worth little only because nobody prepared them. Telling those two apart is the most useful thing a data owner can do before talking to a buyer.
Seven things that lower value
| Problem | Why buyers mark it down | Can you fix it? |
|---|---|---|
| Duplicates | Repeated records teach nothing new and inflate cost | Yes, by deduplication |
| Wrong difficulty | Tasks every model passes, or none does, give no learning signal | Partly, by selecting and grading |
| No way to check | If correctness cannot be confirmed automatically, it is hard to use for agent training | Sometimes, if tests or states can be added |
| Already public | Models may have seen it, so it adds little and cannot serve as a clean test | No |
| Unclear rights | Buyers need warranties; without them the data is risky | Often, with documents and a lawyer |
| Personal data | Must be removed or de-identified, at a cost, and some context is lost | Yes, at a cost |
| Thin or generic | Small, shallow or boilerplate records give few useful tasks | No, but bundling can help |
Duplicates
Researchers who studied common text datasets found many near-duplicate examples, including a single 61-word sentence repeated over 60,000 times in one widely used corpus. Removing duplicates let models emit memorized text about ten times less often and reach the same accuracy in fewer training steps. The lesson for owners: buyers deduplicate, so repeated content, copied templates and exported reports that restate each other will not count twice.
Tasks that teach nothing
Epoch’s interviews with reinforcement learning practitioners describe a narrow useful band. A pass rate of 0 percent or 100 percent means the model will not learn from the task. Some buyers want at least roughly a 2 to 3 percent success rate, meaning at least one success in 64 or 128 attempts, and tasks near 70 percent may be thrown out in favor of harder ones. Work that frontier models already do easily, like routine boilerplate, sits at the wrong end of that range.
No way to check the result
The same interviews name robustness against reward hacking as the top quality criterion, and note that many tasks lack clean verification beyond passing unit tests. A discussion thread with no recorded outcome is interesting to read but hard to grade. A bug report linked to a fix with a failing test is easy to grade. If your records never captured outcomes, expect a lower offer.
Material models have already seen
A 2025 study found that leading models could name the files to edit for tasks from a popular public coding benchmark far more often than for repositories outside it, a sign of memorization. Public material tells a buyer little they cannot get elsewhere. We cover this in /insights/never-published-data-and-benchmark-contamination.
Rights and personal data
Lawyers quoted in press coverage of the data resale market describe unvetted workplace data as legally risky, and report that suppliers scrub personal details and leave out HR records while buyers scrub again. Each step removes content and adds cost. Two companion articles go deeper: /insights/data-ownership-rights-before-selling-data-to-ai-companies and /insights/personal-data-lowers-data-value.
Thin archives and honest uncertainty
Reporting on wind-down data sales put typical payouts at $10,000 to $100,000 per company, with a few reported larger. Size is not the whole story either, since a large archive of duplicated or generic material yields few useful tasks. The reporting also notes that no company has published tests comparing real and synthetic training data, so buyers quote widely different numbers. When the buyer is unsure of the benefit, the offer shrinks.
Why cheap tasks can backfire
Epoch’s report notes that cheap, low-quality tasks may waste compute, citing an estimate of about $2,400 of training compute spent per task. That is one estimate, not a market rule. But it points to a real effect: when a lab pays to train on each task, a weak task is a double loss, wasted compute on top of the purchase price. Buyers therefore screen hard, and what they screen out is worth nothing to them.
Mixed archives get averaged down
Few archives are wholly good or wholly bad. A typical company archive contains a core of rich material, such as real fixes with tests and linked discussion, surrounded by filler: generated reports, copied libraries, social chat. A buyer does not pay for the core and ignore the rest. The filler adds review time and storage and can carry its own rights or privacy problems. Separating the core from the filler before a conversation starts usually helps, even though it makes the archive look smaller, because the buyer then sees what is actually worth reviewing.
What you can fix before you ask
- Deduplicate. Remove copied exports, vendored libraries and repeated templates.
- Keep outcomes. Preserve tests, build results and the state of tickets, not just conversation.
- Document rights. A short record of who made the data and under what terms saves weeks later.
- Plan the privacy work. Decide what to exclude outright, such as HR and health material.
- Combine sources. A small archive linked to tickets and chat may beat a larger isolated one. See /insights/linked-data-code-tickets-chat.
The free estimate at /estimate already reflects most of these: it asks how many sources you hold, whether personal data is present and how clear your rights are, and it lowers the figure when the answers are weak. It will tell you plainly if an archive looks small.
Questions we get
- Why would an AI lab not want my data?
- Usually because it is duplicated, already public, impossible to verify, tangled up with personal data or unclear rights, or too small to produce many useful tasks. Each problem can often be reduced with preparation.
- Is old or messy company data worth anything?
- Sometimes. Messy data is not the issue by itself; unfixable problems are. Duplicates and missing links can be cleaned. Missing rights or missing records cannot be recreated.
- Does the size of my codebase set its value?
- Not by itself. A large codebase full of duplicated, generic or untestable code yields few useful tasks, while a smaller one with real problems, tests and linked context can yield more.
- Can I sell data that models can already find online?
- You can try, but public material adds little for a buyer, and it cannot serve as a clean test because models may have seen it.
What is your data worth?
Sources
- Lee et al., “Deduplicating Training Data Makes Language Models Better” (2021, ACL 2022)
- Epoch AI, “An FAQ on reinforcement learning environments” (2026)
- Liang, Garg and Zilouchian Moghaddam, “The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason” (2025)
- MIT Technology Review, report on AI labs buying data from dead startups (October 13, 2026)
- Gizmodo, report on failed companies selling old workplace chat and email archives to train AI (April 17, 2026)