Legal Insights

The ‘AI’ label is not a licence to disclaim liability: What customers need to know about AI liability exclusions

• 19 November 2026 • 11 min read
https://assets.maddocks.com.au/files/HEADER-The-AI-label-is-not-a-licence-to-disclaim-liability-eAlert-Nov-2026.jpg

AI liability exclusions have become a fixture of enterprise software contracts, and they are not always easy to spot. Some are obvious, such as blanket carve-outs that exclude all liability for AI-related failures, regardless of the cause. Others are less visible, such as human oversight obligations that require customers to review, verify and assume responsibility for AI outputs, shifting liability away from the supplier without being called an exclusion. 

Yet the risk of AI ‘hallucinations’, often cited to justify these exclusions, may have little relevance to the functionality being supplied. Many AI features are not generative AI, and the ‘AI’ label alone should not justify removing the warranties and other protections that would ordinarily apply to software.

This article explains how to identify these provisions, why they can go further than the underlying risks justify, and what customers can do about it.

Introduction

Suppliers are marketing AI as transformative, but their contracts often tell a different story.

Across enterprise software agreements, AI-specific provisions are now routine and tend to shift risk in one direction: towards the customer. Many suppliers exclude liability for AI-enabled functionality, carve it out of ordinary software performance warranties and require customers to review, verify and assume responsibility for all AI outputs. Some contracts go further, requiring the customer to indemnify the supplier against all losses arising from the customer’s use of AI features that the supplier designs, controls and profits from.

These provisions are often presented as ‘market standard’, but that does not make them commercially justified. A blunt, ‘one-size-fits-all’ approach overlooks the questions that should determine how the risk is allocated: what type of AI is being used, whether the underlying risk is any greater than it was before the feature was labelled ‘AI’ and which party is actually in control of the factors that give rise to that risk.

This article examines three related problems: 

  • why the risk of AI ‘hallucinations’ does not always justify the exclusions sought;
  • how human oversight obligations can indirectly transfer liability to the customer; and
  • why blanket AI exclusions can be commercially unreasonable, even where genuine AI risk exists.

How AI liability exclusions typically appear in contracts

AI liability exclusions take several forms. Their effect is often to remove or limit the supplier’s responsibility for losses arising from AI-enabled functionality, even where the supplier caused or contributed to those losses. Common examples include:

  • Blanket AI carve-outs in limitation of liability clauses

    Suppliers may exclude liability for losses arising from AI features, outputs, recommendations, predictions, classifications or decisions generated by the software. These exclusions may apply even where the functionality is embedded in the core product and entirely within the supplier’s control.

  • No-warranty and ‘as-is’ disclaimers

    Suppliers may disclaim all warranties for AI features, providing them on an ‘as is’ or ‘as available’ basis. In practice, this may leave the customer with no contractual remedy if the feature does not work.

  • Carve-outs from performance warranties

    Suppliers may exclude AI functionality from standard warranty commitments that the software will meet specifications or operate as documented. 

  • Customer indemnities for AI-related losses

    Suppliers sometimes require customers to indemnify them against all losses arising from the customer’s use or operation of AI features. Depending on their wording, these indemnities may shift responsibility to the customer for matters the supplier controls, including third-party IP infringement, security breaches in the supplier’s systems and misuse of customer data. Customers should strongly resist indemnities that transfer these risks.

  • AI-specific schedules and addenda

    AI-specific terms may appear in separate schedules or product annexures, often near the back of the contract and presented as standard product terms. These can contain significant disclaimers, oversight obligations and liability exclusions that override protections negotiated in the main agreement yet may not receive the scrutiny they warrant.

Not all risk transfers are as obvious as exclusions or disclaimers. For example, customer oversight obligations can also shift liability to the customer, as discussed below.

Together, these provisions can create a significant imbalance in risk allocation: the supplier markets and sells AI capabilities as core product features, while limiting its contractual responsibility when those features fail to perform. The ‘AI’ label becomes a basis for withdrawing protections that would otherwise apply.

Why the hallucination justification can be a red herring

The most common justification for broad AI exclusions is the risk of hallucinations. That concern is sometimes real, but it does not arise in the same way across all AI-enabled functionality.

Hallucinations are primarily associated with generative AI and large language models (LLMs), where a system produces confident but fabricated outputs. This is a genuine and well-documented risk for generative AI applications, particularly where a system is asked to generate free-form text, answer questions or produce content based on incomplete context. 

That association is understandable. Most people’s day-to-day experience of AI (whether as lawyers, procurement professionals, executives or end users) has been through LLM-based tools such as ChatGPT or Microsoft’s Copilot, and the risks of those tools (hallucinations, bias, inconsistency) are widely reported.

Not all AI is the same, and the distinction between different types of AI matters but is often overlooked.

Deterministic AI produces consistent, predictable outputs for a given input (the same data in, the same result out). 

Non-deterministic AI, of which generative AI is the primary commercial example, produces outputs that can vary for the same input. Generative AI creates new content by predicting likely outputs rather than retrieving or calculating a verified answer, which is what makes hallucinations possible. 

Many enterprise AI tools are deterministic: they perform defined tasks, their outputs are measurable, and their errors are classifiable. Examples of deterministic AI include: 

  • a fraud detection model scores transactions against patterns learned from historical data and flags those that exceed a defined risk threshold. Once trained and deployed, it will score the same transaction in the same way, and its accuracy can be measured against a defined false positive rate; 
  • a document classification tool assigns documents to predefined categories based on their content, again consistently and measurably; and
  • an optimisation algorithm calculates the most efficient solution to a constrained problem, such as the fastest delivery route given a set of variables, and will produce the same result for the same inputs. 

These are not systems prone to fabricating information or producing unpredictable outputs. They are more akin to conventional software, and the liability frameworks that apply to conventional software should apply to them.

It is worth noting that the same function can be implemented using deterministic or generative AI depending on the supplier’s design choices. Customers should ask how each feature actually works rather than accepting the product description at face value. When errors occur in deterministic systems, they are performance or classification failures, not hallucinations. Suppliers have long accepted liability for defects of this kind, and they should continue to do so regardless of whether the feature is now labelled ‘AI-powered’.

The hallucination justification is therefore often a red herring. It imports generative AI risk logic into enterprise software products where that risk may have little or no relevance.

The human oversight problem: an AI liability exclusion in sheep’s clothing?

AI liability exclusions are often coupled with obligations requiring customers to review, verify or supervise AI outputs before using or relying on them. At first glance, this may appear reasonable, as human oversight is an important part of responsible AI governance. However, these clauses often operate less as a genuine safety measure and more as a mechanism for transferring liability to the customer.

The typical structure is:

  • the supplier excludes liability for AI outputs;
  • the customer is required to review and verify those outputs; and
  • if an issue arises, the supplier argues that the customer either accepted responsibility by approving the output or failed to carry out the required review.

Oversight obligations can therefore reinforce an express liability exclusion or operate as an indirect risk transfer in their own right. Rather than addressing whether the supplier’s product performed as promised, the dispute shifts to whether the customer should have identified and prevented the problem.

There are several reasons why customer oversight obligations can be particularly problematic:

  • Suppliers control the AI

    In many technology contracts, the supplier retains broad discretion to modify, update and enhance the software, including by introducing new AI capabilities, changing models, updating features or using third-party AI systems. Customers may not know which features use AI, what type of AI is being used, or when new AI functionality has been introduced. It is difficult to see how a customer can feasibly oversee AI outputs when it does not even know which outputs are AI-generated. Despite that limited visibility, oversight obligations are frequently imposed without qualification, meaning customers often cannot realistically comply with them.

  • Customers often cannot meaningfully oversee AI.

    Customers may lack the technical expertise, system access and contextual information needed to evaluate AI outputs. The supplier designed and deployed the model and should understand its parameters, limitations and failure modes. In some cases, the customer never receives the output at all: certain AI processes operate entirely on the supplier’s back end, producing decisions that are never exposed to the customer.

  • Supplier-side oversight may be more appropriate

    If human oversight is genuinely required, the supplier will often be better placed to perform it. The supplier understands the model, controls its configuration, and has access to the testing, monitoring and validation processes needed to identify whether outputs are within expected parameters. Customer review of an output’s suitability for a particular business use should not substitute for supplier-side quality assurance.

  • Oversight can defeat the purpose of the solution

    Many AI-enabled tools are procured for the purpose of automating processes at scale. Requiring the customer to manually verify every AI-generated output can undermine the commercial rationale for the product. 

  • Oversight is a matter for the customer to determine

    Whether to review an AI output, and to what degree, will depend on the customer’s operational context, risk appetite and internal governance framework. These are judgement calls that sit within the customer’s own domain. Where oversight is framed as a contractual obligation owed to the supplier, the customer may be exposed to allegations of breach even where it has made a considered, risk-based decision not to review a particular output. Any agreed obligation should specify what review is required and in what circumstances, rather than treating every unreviewed output as a customer failure.

Why blanket AI liability exclusions can be unreasonable

Blanket AI liability exclusions are commercially problematic for several reasons:

  • They reclassify existing defects

    Many AI exclusions treat any error associated with AI functionality as a special category of risk. This can have the effect of reclassifying ordinary software defects as ‘AI risk’. If, for example, a software tool misclassifies a document, misroutes a workflow, fails to detect an anomaly or produces a flawed prediction, the key question should be whether the product performed in accordance with the contract, the documentation and any applicable service levels. The use of AI should not, by itself, change the allocation of risk.

  • They create an expanding and uncertain carve-out

    As more enterprise software incorporates machine learning techniques and AI functionality, a broad AI exclusion can progressively erode core supplier accountability. This is particularly concerning where suppliers can update, modify or enhance the software, including by introducing new AI features, without meaningful customer approval. Each new AI-enabled feature may reduce the supplier’s contractual responsibility without any renegotiation. The problem is compounded where ‘AI’ or ‘AI-related functionality’ is broadly defined or not defined at all. On some definitions, technologies that have existed for decades could fall within the scope of the exclusion. A broadly drafted AI carve-out may therefore capture far more than the parties intended, removing liability for functionality that has nothing to do with the hallucination risk that supposedly justifies the exclusion.

  • They shift performance risk to the customer for matters the supplier controls

    Customers will often have limited or no control over matters such as model selection and architecture, testing, validation, guardrails and release management. Yet a broad AI exclusion may require the customer to bear the consequences of failures arising from those matters. The supplier is the party best placed to prevent, detect and remediate performance issues, and liability should follow accordingly.

  • Suppliers charge for AI while disclaiming liability for it

    Suppliers market AI features as a reason to purchase their products, may charge additional fees for AI-powered functionality, and use AI as a competitive differentiator. Having chosen to monetise those AI features, it is arguably unreasonable for a supplier to then disclaim liability for the very capabilities it is selling. Other categories of commercial software do not typically attract this treatment; a supplier should not be able to market AI as a core product feature while simultaneously refusing to stand behind its performance.

It is also worth remembering that technology contracts already contain well-established mechanisms for limiting supplier exposure, including liability caps, exclusions of indirect or consequential loss, and limits on the categories of recoverable damage. Those provisions are designed to keep a supplier’s exposure commercially reasonable. In many cases, there is no need for an additional AI-specific carve-out on top of those protections, because the existing framework already does the work.

Even for Generative AI, blanket exclusions may be unreasonable

Generative AI does present genuine risks, but those risks do not, of themselves, justify a blanket exclusion of liability.

For generative AI, output quality is influenced by a range of factors, many of which remain within the supplier’s control. These may include model selection, quality of training data, fine-tuning, prompt engineering, grounding techniques, content filtering, testing and deployment. A supplier that offers generative AI as part of a commercial product should not be able to disclaim responsibility for poor outputs that result from inadequate design, testing or safeguards.

By selling generative AI as part of a paid enterprise solution, suppliers present the technology as suitable for commercial use. Customers can therefore reasonably expect an appropriate level of maturity, reliability and risk management. A supplier should not market AI as a valuable capability, then characterise it as too experimental or unpredictable to stand behind when liability arises.

Accuracy disclaimers raise a related issue. There is a difference between occasional imperfect outputs, which may be inherent in generative AI, and outputs so unreliable that the product is commercially unusable. An accuracy disclaimer should not operate as a licence to deliver a solution that does not work.

Even where a customer accepts that individual AI outputs may occasionally be inaccurate, that concession should not displace any fitness-for-purpose or performance warranty that would otherwise apply to the product. There is a meaningful difference between accepting that outputs may not always be perfect and accepting that the product may never reliably perform. Any inaccuracy acknowledgement should be narrowly scoped to inherent output variability rather than operating as a blanket waiver of supplier accountability for overall performance.

Drawing that line is more difficult for generative AI than for deterministic AI, because output quality is harder to measure in a standardised way. However, that does not make contractual standards impossible. For AI functions that are central to the value proposition of the solution, customers should consider output quality obligations, such as minimum thresholds for accuracy, relevance, completeness or coherence, measured against defined benchmarks, datasets or use cases. These mechanisms can establish a contractual floor below which performance must not fall, while still recognising that generative AI outputs may not be perfect in every instance.

Practical Guidance for Customers

Customers should resist blanket AI liability exclusions and instead seek a proportionate, risk-based approach to AI liability. Practical recommendations include:

  • Review the full contract, including schedules and product terms

    AI-specific provisions are often contained in schedules, addenda or product-specific terms. Check whether those provisions override liability, warranty, security or other protections negotiated in the main agreement.

  • Require transparency about AI use. Ask the supplier to identify:
    • which features use AI;
    • what type of AI is used;
    • whether generative AI or an LLM is involved;
    • whether third-party AI models or services are used; and
    • what testing, validation and monitoring processes apply.
  • Conduct a risk assessment before deploying AI features

    Before deploying any AI functionality, customers should assess the risks associated with those features in their specific use context. The assessment should inform decisions about deployment, required safeguards and which outputs warrant human review. For high-risk use cases (such as automated decision-making), the risk assessment should be documented and reviewed periodically. Organisations subject to the Privacy Act 1988 (Cth) should also note that, from 10 December 2026, privacy policies must include specified information about the use of personal information in certain automated decision-making processes, including where a computer program is used to make, or to do a thing that is substantially and directly related to making, decisions that could reasonably be expected to significantly affect individuals’ rights or interests. This may require organisations to understand how AI and automated decision-making features provided by their technology suppliers operate.

  • Distinguish between generative and deterministic AI

    Different risk profiles warrant different treatment. Deterministic AI should remain subject to ordinary software warranties and performance commitments. For generative AI, customers may accept tailored disclaimers for inherent output variability, but should preserve supplier accountability for design, safeguards, security, confidentiality and IP infringement.

  • Calibrate liability

    Where the risk relates to outputs of generative AI, a reduced liability position may be appropriate, but only if the supplier has implemented reasonable safeguards such as content filtering, grounding, testing and supplier-side human oversight. Any liability carve-out should be limited to specific AI features and defined categories of loss, not a blanket exclusion.

  • Resist customer-side indemnities for AI-related losses

    Be alert to provisions that require the customer to indemnify the supplier for losses arising from the customer’s use of AI features. These clauses can operate as reverse indemnities, shifting to the customer risks that the supplier controls, and should generally be rejected or confined to losses caused by the customer’s own misuse or breach.

  • Preserve existing contractual obligations

    Include a provision making clear that the supplier must comply with all of its obligations under the contract notwithstanding the use or incorporation of AI. The fact that a feature is AI-enabled should not excuse non-compliance with obligations that would otherwise apply, including privacy, security, data access, location and transfer requirements, confidentiality, and service level commitments.

  • Use targeted performance commitments and service levels where appropriate

    Where AI functionality is central to the solution, customers should consider measurable requirements and associated remedies appropriate to the use case, such as minimum accuracy thresholds, relevance or completeness benchmarks, maximum error rates, remediation obligations, escalation rights, service credits or termination rights for persistent failure.

  • Make oversight obligations proportionate

    Resist broad obligations to review every AI output. If oversight is accepted, it should be limited to specified high-risk outputs, proportionate to the customer’s expertise and operating model, conditional on the supplier providing sufficient information and access, and matched by supplier-side monitoring and quality assurance obligations. Customers should also be mindful of the distinction between oversight they elect to apply for their own risk management purposes and oversight framed as an obligation owed to the supplier; the former is an exercise of the customer’s own governance judgement, and the two should not be conflated in the drafting.

  • Require supplier-side governance

    Suppliers should be contractually obliged to maintain appropriate AI governance processes, including documented AI policies, model testing and validation, monitoring for performance degradation, incident management, security and privacy controls, human review where appropriate, and audit or reporting rights for high-risk use cases.

  • Require supplier accountability for training and design

    Where the supplier controls the AI model or training data, it should be contractually required to follow industry-standard practices for data handling, model development and testing. Obligations should cover data quality and appropriateness of training datasets, periodic validation, accuracy testing, and ongoing monitoring for performance degradation, bias or drift.

Final thoughts

  • The label ‘AI’ should not be a licence to disclaim liability.
     
  • The core principle of technology contracting (i.e. that the party best placed to control a risk should bear responsibility for it) does not change because a product uses machine learning. Suppliers who design, deploy and commercially benefit from AI functionality should remain accountable for it. 
     
  • Customers are entitled to expect that AI features will perform as described, that existing contractual protections will not be eroded simply because AI is involved, and that oversight obligations will not be used as a mechanism to shift liability onto a party that had neither the visibility nor the means to prevent the problem.
     
  • That does not mean AI risk is not real, or that tailored liability positions are not appropriate in certain contexts. It means that AI-specific terms should be scrutinised as carefully as any other commercial provision. A supplier may claim a provision is ‘market standard’, but that does not make the resulting allocation of risk appropriate. Customers should independently consider and assess whether the particular risk justifies different treatment, which party is best placed to manage it, and whether the proposed drafting reflects that allocation.

Maddocks has deep expertise in drafting and negotiating AI-related provisions in commercial contracts

This includes liability exclusions, human oversight obligations, indemnities, warranty frameworks and supplier governance requirements. If you would like assistance reviewing or negotiating AI terms in your commercial contracts, please contact any member of our Technology Procurement team.

Jeff Goodall

Jeff has deep expertise and extensive experience advising corporate and government clients on a broad range of complex technology and general commercial transactions.

View profile

Jack Evans

Jack specialises in commercial and technology matters including outsourced solutions, technology licensing, hardware acquisition, general procurement and subcontracts and privacy law.

View profile

Bernadette Lim Lo Suy

Bernadette has experience advising on a broad range of commercial matters with a focus on technology procurement, telecommunications, intellectual property, privacy and general commercial matters..

View profile

Recent articles

Online Access