AI Copyright: What to Audit in Your Product Before They Audit You

July 28, 202613 minAI, Legal, SaaS, Risk, Copyright

Short answer (60 seconds): on July 9, 2026 the New York Times + 16 media outlets asked a federal judge to sanction OpenAI for allegedly lying for ~2 years about its technical ability to search its systems for evidence of copyrighted material use. On July 22, the U.S. Treasury left sanctions "on the table" against Moonshot AI for the alleged distillation of Anthropic's Fable. On July 24, the Delhi High Court ruled in favor of OpenAI in India (interim). Three jurisdictions, three different signals. If your SaaS has an AI feature — even a thin API wrapper — you're exposed to the same argument. What you have to audit this week: (1) your ToS and indemnification clause, (2) data lineage of your feature, (3) your provider's training data policy, (4) your E&O policy with an IP/AI rider. Checklist at the end.

When an executive committee calls me about an AI project in 2026, the legal risk is no longer hypothetical. What changed this week is that the risk went from "theoretical and future" to "concrete and this week." This is not an alarmist post. It's an operational post: four audits a SaaS founder can run solo, without a lawyer, in 1-2 afternoons. The formal legal audit is your lawyer's job; this post gives you the framework to know what to ask them.

Why it matters this week

Three cases in 15 days, three jurisdictions, three different signals. Worth understanding concretely because together they sketch the risk map:

Case 1 — NYT + outlets vs OpenAI (U.S., July 9, 2026)

The New York Times, NY Daily News, and 15 other outlets filed a motion before a federal judge asking that OpenAI be sanctioned. What they allege: OpenAI deliberately concealed its ability to search its training data and output logs for evidence of copyrighted article use, for approximately two years. The key phrase in the court filings is that OpenAI allegedly deployed a "deliberate and systemic effort to obstruct discovery."

This matters for the precedent, not the specific case. If the judge rules in favor of the publishers, any other publisher can use the same argument (and the same discovery obstruction complaint) against any SaaS with an AI feature trained on scraped content. You don't need to be OpenAI. You just need to have a feature that uses a model trained on copyrighted content and be unable to show where that content came from.

Sources: NYT, AP News, Bloomberg Law.

Case 2 — Treasury vs Moonshot AI (U.S.–China, July 22-23, 2026)

On July 22, TechCrunch reported that White House Science Advisor Michael Kratsios accused Moonshot AI — creators of the Chinese Kimi K3 model — of having built their model by distilling Anthropic's "Fable" LLM, using chips not authorized for export to China. Treasury Secretary Scott Bessent said sanctions remain "on the table." On July 23, a TechCrunch follow-up noted that several technical experts dispute the narrative: they argue exploiting Fable wouldn't explain Kimi K3's performance.

What does this have to do with your SaaS? Two things. First: if the U.S. starts sanctioning companies for cross-border use of Western models, your AI supply chain (provider + sub-provider + country of hosting) becomes a real legal risk. Second: things move fast. It went from "public accusation" to "sanctions threat" in 48 hours.

Sources: TechCrunch July 22, TechCrunch July 23.

Case 3 — Delhi High Court vs OpenAI (India, July 24, 2026)

On July 24, the Delhi High Court rejected ANI Media's request for an injunction against OpenAI. Reason: training ChatGPT on news content falls prima facie under Section 52 of the Indian Copyright Act (fair dealing). The judge further reasoned that restricting training would be detrimental to AI growth and public interest.

Caveat: this is an interim ruling on injunctive relief, not a final decision on whether infringement occurred. The underlying case remains open. And it only applies in India. If your SaaS serves customers in the U.S., EU, UK, or Latam, it doesn't protect you. But it does tell you something important: there's no global rule. Each jurisdiction decides differently, and that means the legal audit you run has to assume the worst case, not the best.

Source: Economic Times, Bar and Bench.

Audit 1 · ToS and indemnification clause

The question you need to answer today: what happens if a customer or a third party sues you because your AI feature reproduced copyrighted material? Are you covered by your contract with your model provider (OpenAI, Anthropic, Google), or by your own policy?

What to look at concretely:

• Open your model provider's ToS. Look for "indemnify", "indemnification", "defend", "hold harmless". Does your provider defend you against third-party claims for use of the model? Or is that defense limited to their own acts (not yours)?
• Look for "training data", "outputs", "intellectual property". Does the provider claim ownership over the outputs? Do they transfer rights to you? Or do they give you a license?
• Look for "warranty disclaimer" clauses. Almost all providers disclaim any warranty about non-infringement of outputs. If they do, know that your risk isn't covered by anyone.
• If your provider is OpenAI or Anthropic via API, they historically offer limited indemnification for cases where the output infringes and you didn't modify the prompt in a way that produced the infringement. Verify the exact current terms — they change often.

What I've seen in Latam projects: 70% of SaaS with an AI feature have not read the IP section of their model provider's ToS. The 30% who have read it have a false sense of security because they read the indemnification clause without reading the carve-outs (exceptions that void it).

Audit 2 · Data lineage

The question: if a judge asks you to prove what data your AI feature used, can you do it in 48 hours?

What "data lineage" means in practice for your feature:

• RAG corpus: what documents did you index? Where do they come from? Do you have a record (with timestamps) of when each was indexed? If tomorrow you were asked to list the 10,000 documents that form your knowledge base, could you?
• Fine-tuning data: if you fine-tuned a base model, what data did you train it on? Where is it stored? Who has access?
• Embeddings: embeddings technically encode fragments of the original text. If you're sued over your RAG corpus and asked to preserve the embeddings, can you identify which embeddings correspond to which original document?
• System prompt and few-shot examples: are they documented and versioned? Or do they only live in production code without a changelog?

Anonymized case: B2B SaaS for legal documents (~30 people) built an AI copilot to summarize contracts. RAG over their internal knowledge base (all proprietary, licensed, or public). They passed the origin audit. But when I asked for the log of which embeddings corresponded to which document, they didn't have it. It took them 3 weeks to reconstruct it with their vector DB provider. If a judge had asked for that information in 48 hours, they wouldn't have had it. That's insufficient data lineage.

The minimum that must exist: a table (even a simple CSV) that maps each chunk in your vector DB → original document → ingestion date → source (URL, customer upload, feed, etc.). If your vector DB provider gives it to you in a query, great. If not, write a script to generate it. It's one day of work.

Audit 3 · Provider's training data policy

The question: does your model provider openly declare what type of data they used to train? And can you get that declaration in writing?

The problem: most frontier providers (OpenAI, Anthropic, Google, Meta with Llama) have training data policies that are deliberately ambiguous. They say things like "we use a mix of public data, licensed data, and human-generated data" without specifying proportions, sources, or opt-out mechanisms.

What to ask your provider concretely (by email, not chat):

1. Sworn statement of training data origin for the specific model you use (not for "our family of models"). Different models in the same family can have different policies.
2. Confirmation of right to sublicense outputs — can you give your customers a license over the outputs without the provider claiming anything?
3. Opt-out mechanism for publishers — is there a formal process for a publisher to ask that their content be excluded from training? If so, are there known publishers who have used that process against this model?
4. Change history — if the model was re-trained or updated, does the new policy apply retroactively to prior outputs?

Why it matters: the legal argument "I didn't know what data was used to train" is sympathetic but weak. The argument "I asked, I got a written statement that said X, I acted based on that" is much stronger in discovery.

Audit 4 · Insurance: E&O + IP/AI rider

The question: does your Errors & Omissions (E&O) policy cover an IP claim for content generated by your AI feature?

The reality I see: most startup SaaS E&O policies cover service defects, professional negligence, data breaches. They do not cover by default intellectual property claims on outputs. Much less if those outputs were generated by an AI model that the provider trained on scraped content.

What to ask your broker (or a new one, if your current doesn't know AI):

• Written confirmation that the policy covers IP/AI claims. Email or written rider, not verbal.
• Specific coverage for "AI-generated content" or "generative AI outputs". Some brokers (Coalition, At-Bay, Vouch in the U.S.) already have specific riders.
• Sub-limit per incident and annual aggregate. Realistic range for a Series A-B startup: USD 1-5M per incident.
• Exclusions to check: does the policy exclude claims based on "model provider's training data"? Does it exclude copyright claims specifically? If the answer is yes to either, you have a hole.
• Expected rider cost: USD 5,000-15,000/year on top of the existing E&O premium for a USD 5-20M raised startup. Affordable. Unaffordable is not having it and finding out when the claim arrives.

What this point does NOT cover: legal advice on your specific policy. This is so you know what to ask your broker, not so you decide alone. A serious broker specialized in tech should be able to answer the 5 questions above in a 30-min call.

The cross-border angle: why it matters even if you don't operate in China or India

The Moonshot/Fable case (July 22) and the Delhi HC ruling (July 24) are signals that AI copyright risk is not just "U.S. vs OpenAI." It's a problem being played out in multiple jurisdictions with contradictory signals.

What this means for your SaaS:

• If you serve U.S. customers, your worst case is the NYT scenario (sanctions, aggressive discovery, possible damages).
• If you serve Indian customers, the Delhi HC ruling favors you for now, but the underlying case remains open.
• If you use a provider with infrastructure in Asia (even as a sub-provider), the sanctions/export angle applies.
• If your feature lets users generate content that is then distributed globally, you're technically exposed to every jurisdiction where that content is consumed.

Operational implication: don't assume that being in Latam makes you safe. Your AI supply chain matters. Your model provider matters. Your ToS with your customers matters. The full legal audit requires assuming the worst case, not the best.

Actionable checklist for this week

Prioritized. If you can only do five things, do the first five. The rest can be delegated.

• High priority · day 1-2: Read your model provider's ToS (OpenAI, Anthropic, Google, etc.). Look for "indemnification", "IP", "training data". Document what you find in an internal note.
• High priority · day 1-2: List all providers in your AI supply chain (model, embeddings, vector DB, hosting, GPU provider). Is any of them in a jurisdiction with sanctions risk?
• High priority · day 2-3: Confirm with your model provider, by email, their training data policy specific to the model you use. Save the email.
• High priority · day 3: Contact your insurance broker. Ask specifically if your E&O covers IP/AI. Request written confirmation.
• High priority · day 3-4: Map your data lineage. Table: each chunk in vector DB → original document → date → source. If you don't have it, write a script to generate it.
• Medium priority · week 2: Talk to your lawyer about drafting your ToS toward the end customer. Does your ToS protect you if the output infringes?
• Medium priority · week 2: Evaluate alternatives to your current provider. Is there a model with explicitly licensed training data (e.g. AI2's open models, Mistral with enterprise license) you could use for critical flows?
• Medium priority · week 2: If you serve enterprise customers, let them know you're running the audit. A proactive "compliance in progress" note can be the difference between renewing the contract or not.
• Low priority · month 1: Consider a second provider (Claude ↔ OpenAI are usually interchangeable with minor changes) as fallback if the primary enters systemic litigation.
• Low priority · month 1: If you do fine-tuning, audit the origin of fine-tuning data with the same rigor as RAG.
• Low priority · month 1: Document who on your team has access to training datasets and embeddings. If no one has formal access, you have a security risk in addition to the legal one.

What this post does NOT cover

To be clear:

• This is not legal advice. If you need specific legal advice on your contract with your model provider or on how to draft your ToS toward your customers, hire an IP/AI lawyer. In Latam there are specialized firms; in the U.S., lawyers with experience in generative AI litigation.
• It doesn't cover pre-LLM classical ML. Copyright risk on models trained with proprietary datasets is different and more manageable. This post assumes you use frontier LLMs or recent open models.
• It doesn't cover the provider side. If you're building a model from scratch and training on scraped data, this post isn't for you — you need specific legal advice well before training.
• It doesn't cover deepfakes or image generation. The NYT/Moonshot/Delhi cases are about text. The jurisprudence on AI-generated images has its own precedents.

If you want help working through this

If your SaaS has an AI feature and you want help running this operational audit (the 4 above), or want me to review your specific case with your model provider, book a free 30-minute call. I run this kind of audit for Latam SaaS founders and it usually uncovers something the client hadn't seen.

Frequently asked questions

Does this apply to my SaaS if I only use RAG on my own documents?

Less than if you fine-tuned or used scraped training data, but not zero. If your embedding or LLM provider was trained on copyrighted material and you can't show your output is sufficiently transformative, you're exposed to the same argument the publishers are making against OpenAI. The minimum audit is still necessary.

How much does adding IP/AI coverage to my E&O policy cost?

For a Series A-B SaaS startup (USD 5-20M raised), an IP/AI rider on an existing E&O policy typically costs USD 5,000-15,000/year extra, depending on coverage (USD 1-5M per incident) and deductible. Brokers specialized in tech (Coalition, At-Bay, Vouch) have specific products. If your current E&O doesn't mention 'AI' or 'generative', it almost certainly doesn't cover you.

What do I do if my model provider won't give me training data details?

Document the refusal in writing (email is enough). That protects you in discovery if you ever get sued: it shows you attempted due diligence. Meanwhile, assume the worst case in your internal risk assessment: that the model was trained on scraped copyrighted content. If that changes your decision to use it (e.g. it's replaceable by an openly licensed model), consider migrating.

Does the Delhi HC ruling in favor of OpenAI mean the risk is low?

No. It's an interim ruling on injunctive relief, not a final decision on infringement. It also only applies in India under the Indian Copyright Act. If your SaaS serves customers in the US, EU, UK, or Latam, that ruling doesn't protect you. What it does tell you is that there is no global rule yet — and that's part of the problem: each jurisdiction decides differently.

Do I do this audit myself or hire a lawyer?

The 4 audits in this post you can do yourself or with your CTO this week (they're operational: read ToS, list providers, request statements, talk to your broker). What you CANNOT do alone is the legal advice on your specific contract with your model provider or the drafting of your ToS toward your customers. That's a lawyer, not this post.