Last updated: August 2026
The marketing page tells you what an AI tool can do. The terms of service tell you who owns the outputs, who is liable when an output is wrong, and whether the vendor will stand behind you if a generated result draws a legal claim. For an engineer or a firm, those are not fine-print details; they decide whether a tool is safe to build real work on. And the recurring theme across every major vendor’s terms is the same: the outputs come with no warranty, the vendor caps its own liability, and the real protections live only in the business tiers, on conditions.
This guide walks the clauses that actually matter for engineering use and how they differ between consumer and business terms. It sits under our roundup of the best AI tools for engineers and complements our pillar on AI tool data security and engineering IP.
Short answer: Read six clauses before you adopt an AI tool. You generally own the outputs, but only to the extent the law allows, and purely AI-generated content may not be copyrightable at all. The services are provided “as is” with no warranty of accuracy or fitness, which is the clause engineers must internalize, because the tool disclaims responsibility for the answers you might build on. The vendor caps its liability, sometimes to as little as the greater of your fees or a fixed small amount. An IP indemnity that defends you against copyright claims exists, but only on business or enterprise tiers and only if you meet its conditions. Training on your inputs is the consumer default and off by default on business tiers. And an acceptable-use violation can quietly void the protections you were counting on. None of this transfers your professional duty of care to the vendor.
Do you own the outputs?
Mostly yes, with an asterisk. Anthropic’s commercial terms state that the customer “owns its Outputs” and retains rights to its inputs (Anthropic). OpenAI’s published terms similarly assign to the user its rights in the output, and note that identical outputs can be produced for other users (per OpenAI’s terms of use). The asterisk is copyright law: in the United States, purely AI-generated content without human authorship is not eligible for copyright protection, a position the Copyright Office and the courts have maintained in the Thaler line of decisions. So “you own the output” means you own whatever rights exist, which for a fully machine-generated artifact may be none. For engineering deliverables, that matters less than the liability clauses below, but it is worth knowing before you treat a generated document as protectable IP.
The “as is” disclaimer, and why it matters most
This is the clause engineers should read first. Anthropic’s commercial terms state that the services and outputs are provided “as is” and “as available” without warranty of any kind, disclaiming merchantability, fitness for a particular purpose, and non-infringement. OpenAI’s terms carry the same disclaimer and, per its published terms, advise that output should not be relied on as a sole source of truth or a substitute for professional advice. In plain terms: the vendor does not promise the answer is correct, and it explicitly tells you not to depend on it. For an engineer who might build a calculation, a specification, or a design decision on an AI output, that disclaimer is the whole ballgame. It is why every reliable workflow keeps a qualified human verifying the result, a point we make throughout our coverage, including AI hallucination in engineering.
Limitation of liability: the vendor’s cap
If an output causes a loss, how much can you recover from the vendor? Very little. Anthropic’s commercial terms cap aggregate liability at the fees paid in the prior twelve months and exclude indirect and consequential damages (Anthropic). OpenAI’s consumer terms are reported to cap liability at the greater of the fees you paid or a small fixed amount on the order of one hundred dollars (per OpenAI’s published consumer terms). Business terms are negotiated and typically track the twelve-month-fees model. The practical reading is that the financial risk of a bad output sits with you, not the vendor, which reinforces why an AI output is an input to your judgment, never a substitute for it.
IP indemnity: who defends you if an output infringes
A newer and genuinely useful protection is the IP indemnity: the vendor agrees to defend you if a generated output is claimed to infringe someone’s copyright. The catch is that these exist only on business or enterprise tiers, and each is conditioned.
| Vendor | Business-tier IP indemnity | Key condition |
|---|---|---|
| OpenAI | Copyright Shield (Enterprise and API) | Not offered on free ChatGPT or Plus |
| Microsoft | Customer Copyright Commitment | You must use the built-in guardrails and content filters |
| Google Cloud | Generative AI indemnity (training data and output) | Not for output you knew or should have known was infringing |
| Anthropic | Defense under the commercial terms | Authorized, paid use of the services or outputs |
The specifics, from the vendors: Microsoft’s Customer Copyright Commitment obligates it to defend customers against certain third-party IP claims relating to output, on the condition that the customer used the required guardrails and content filters (Microsoft); as Microsoft put it, it will defend the customer and pay adverse judgments “as long as the customer used the guardrails and content filters we have built into our products” (Microsoft). Google Cloud offers both a training-data and a generated-output indemnity for covered generative AI services, excluding output the customer knew or should have known was likely infringing (Google Cloud). Anthropic’s commercial terms include a defense against third-party claims that authorized paid use of the services or outputs infringes IP rights (Anthropic). OpenAI’s Copyright Shield extends a defense to ChatGPT Enterprise and API customers and, by its terms, not to free ChatGPT or Plus (as analyzed by Proskauer). Every one of these is conditional, and none of them touches professional negligence: an indemnity defends against a copyright claim, not against a design that was wrong.
Training, input license, and the consumer-versus-business line
Two more clauses split cleanly along the tier line. Whether the vendor trains on your inputs is the consumer default on several tiers and off by default on business tiers, a distinction we cover in does ChatGPT train on my drawings. Relatedly, the input license you grant differs: business terms such as Anthropic’s state the vendor may not train on customer content, while consumer tiers grant a broader license to improve the service. The pattern is consistent across the whole terms comparison: the protections an engineering team actually wants, no training, an IP indemnity, and a real support relationship, are the ones that appear only in the business and enterprise terms. If confidential client material is involved, pair this with our guide on whether it is safe to use AI with client drawings.
A terms-of-service checklist before you adopt
- Output ownership. Confirm you own the outputs, and remember that purely AI-generated content may not be copyrightable.
- The “as is” disclaimer. Assume no warranty of accuracy or fitness, and design a verification step accordingly.
- Liability cap. Check how far the vendor limits its liability, because the loss from a bad output largely sits with you.
- IP indemnity. Confirm whether your tier has one, and read its conditions, because a free or consumer tier usually has none.
- Training and input license. Use a business tier that does not train on your inputs when the work is sensitive.
- Acceptable use. Read the usage policy, because violating it can void your indemnity and your account.
- Consumer versus business. Match the tier to the risk; the protections you need are in the business terms.
The professional bodies are blunt about the limits of any of this. ASCE cautions against warranty and guarantee clauses in engineering service contracts because professional liability insurance typically excludes claims arising from express warranties (ASCE Policy Statement 388), so an engineer who warrants an AI-assisted deliverable may undercut their own coverage. And an AI system has no license, no standing, and no accountability, so the engineer retains the duty of care no matter what the tool’s terms say.
Frequently asked questions
Do I own what an AI tool generates for me?
The major vendors’ terms assign the outputs to you, so in contract terms you own them. The limit is copyright law: in the United States, content generated purely by AI without meaningful human authorship is not eligible for copyright protection. So you own whatever rights exist, which for a fully machine-generated artifact may be none. For most engineering work this matters less than the warranty and liability clauses, but do not assume a generated document is automatically protectable IP.
Does the vendor take responsibility if an AI output is wrong?
No. Major AI terms provide the service and outputs “as is” with no warranty of accuracy or fitness, and they cap the vendor’s liability, sometimes to the greater of the fees you paid or a small fixed amount. OpenAI’s terms even advise not treating output as a sole source of truth or a substitute for professional advice. The financial and professional risk of relying on a wrong output sits with you, which is why a qualified human must verify anything an AI produces before it drives a decision.
What is an AI copyright indemnity, and do I have one?
It is a promise by the vendor to defend you if a generated output is claimed to infringe someone’s copyright. OpenAI’s Copyright Shield, Microsoft’s Customer Copyright Commitment, Google Cloud’s generative AI indemnity, and Anthropic’s commercial defense all offer versions of this, but only on business or enterprise tiers and only if you meet the conditions, such as using the vendor’s content filters or not knowingly generating infringing material. Free and consumer tiers generally have no indemnity, and none of these covers professional negligence.
Are the consumer and business terms really that different?
Yes, and the difference is the whole decision for a firm. Consumer tiers may train on your inputs, carry no IP indemnity, and give you the same “as is” disclaimer and tight liability cap. Business and enterprise tiers add a no-training default, an IP indemnity, a data processing agreement, and a real support relationship. If you are doing professional work, especially with client material, the protections you need exist only in the business terms.
Sources
- Anthropic, Commercial Terms of Service (ownership, no-training, as-is, liability, indemnity)
- Microsoft, Customer Copyright Commitment required mitigations
- Microsoft, Copilot Copyright Commitment announcement
- Google Cloud, Generative AI indemnified services
- Proskauer, analysis of OpenAI Copyright Shield scope
- ASCE Policy Statement 388, warranty and guarantee clauses
Written by the CognitiveFuture editorial team. We build our guidance from vendors’ own terms and policy pages and reputable legal analysis, and we label vendor commitments as such. Terms change and some vendor pages could not be retrieved directly at the time of writing, so confirm the current terms for your tier and jurisdiction before relying on them. This is general information, not legal advice; consult a qualified attorney and your professional liability insurer for your situation.
Richard Johnson
Richard Johnson is an AI specialist at one of the world's largest technology companies, where he has spent the past three years helping organizations adopt AI. CognitiveFuture extends that work publicly: gathering the available evidence on each tool, from vendor documentation to independent reviews and user feedback, and cutting a crowded market down to the right choice for the job in front of you.
