A business owner in Phoenix scans a QR code on a dog's vest. An app chirps. A green checkmark appears. The handler is waved through. Everyone feels confident. And the entire interaction may have just violated federal law.
The ADA two-question rule is one of the most misunderstood provisions in American disability law, and the explosion of third-party verification apps has made that misunderstanding worse, not better. In 2026, the collision between technological convenience and statutory rights is producing real harm to real people. This article unpacks the legal framework, the DOJ's clear guidance, and the question that every AI engineer building in this space must answer honestly: does your product help businesses comply, or does it help them violate the law while feeling compliant?
What the Law Actually Says
Under Title II and Title III of the Americans with Disabilities Act, staff at public accommodations may ask exactly two questions when it is not obvious that a dog is a service animal. First: is this a service animal required because of a disability? Second: what work or task has the dog been trained to perform?
That is the complete legal toolkit available to a business. Not three questions. Not a request for identification. Not a visual inspection of a certificate, vest, or digital credential. The DOJ regulation at 28 C.F.R. Part 36 is unambiguous on this point. Businesses cannot require documentation that a dog is a service animal, require the dog to demonstrate its task, or inquire about the nature of the person's disability.
The rule reflects a deliberate policy choice, not an oversight. Congress and the DOJ understood that service animal certification does not exist in any standardized federal form. Requiring documentation would effectively create a credentialing system through the back door, one that would disproportionately burden people with disabilities who lack access to documentation resources, money, or technical literacy.
The ADA two-question rule is, in short, a floor for handler dignity. Verification apps that attempt to raise that floor are not innovating on top of the law. They are eroding it.
The Verification App Proliferation
Since approximately 2019 the market for service dog verification apps has grown significantly. These products vary widely in their claims and architectures. Some are simple QR code registries that generate a scannable badge for a fee. Others layer in photo verification, handler attestation forms, and certificate generation. The most technologically ambitious claim to use AI-based behavioral assessment or computer vision to evaluate whether a dog meets some threshold of task competency.
What unites nearly all of them is a business model that profits from the legal confusion they simultaneously create and exploit. By charging handlers for badges and businesses for scanning infrastructure, they monetize a vacuum. That vacuum exists specifically because the ADA does not create a documentation system. The apps fill the vacuum by implying one should exist.
From an engineering standpoint, many of these systems are architecturally thin. A QR code linked to a database record with handler-supplied information and no independent verification is not meaningful authentication. It is form theater. Even systems that use photo capture or video submission rarely employ models with the sensitivity and specificity required to distinguish a well-trained service dog from a well-behaved pet, and none have published peer-reviewed validation data against standardized public access benchmarks like the Assistance Dogs International Public Access Test.
The proliferation is also self-reinforcing. When a large retailer adopts a scanning system, handlers report being turned away if their dog is not registered with that specific vendor. That de facto documentation requirement is exactly what federal law prohibits, regardless of how the requirement is implemented technically.
DOJ Guidance and the Documentation Prohibition
The Department of Justice has addressed the documentation question directly and repeatedly. In its ADA.gov FAQ guidance, the DOJ states explicitly that covered entities cannot require persons with disabilities to provide documentation, such as proof that the animal has been certified, trained, or licensed as a service animal. The DOJ has further clarified that this prohibition extends to third-party registrations, ID cards and vests as a condition of access.
This is not an interpretation that requires nuanced legal analysis. It is the plain reading of the regulation. A business that tells a handler "please scan your dog's registration so I can let you in" has imposed a documentation requirement. It does not matter that the documentation is digital, that it costs only a small fee, or that the app's interface is friendly and the green checkmark feels reassuring.
Enforcement has historically been light, which is part of why the app market has grown. The DOJ's civil rights division is not resourced to investigate every access denial. Private right of action under Title III is available but litigation is expensive and emotionally costly for handlers. State attorneys general have varying levels of engagement. The result is a legal prohibition with significant enforcement gaps, and technology companies have learned to operate in those gaps.
AI engineers building in this space should understand that light enforcement does not equal legal permission. The DOJ has not issued guidance carving out an exception for AI-based verification. No such exception exists. The gap between what the law permits and what the market is doing is a liability exposure question, not a sign that the law has shifted.
State Registry Legislation and Federal Preemption
A number of states have passed or introduced legislation creating voluntary or mandatory service dog registries. These laws vary considerably in their structure. Some create genuine voluntary registries that offer no access implications. Others are more problematic, suggesting that registration confers some special status or that unregistered animals may be questioned more aggressively.
The federal preemption question is important here. Where state law conflicts with federal ADA requirements, federal law controls under the Supremacy Clause. A state cannot legislatively create a documentation requirement that federal law prohibits, even if the state frames that requirement as a consumer protection measure or a business assistance tool. Any state registry scheme that has the practical effect of conditioning public access on documentation is preempted by the ADA regardless of the state's stated intent.
In practice, state registry laws have become marketing material for verification app companies. An app that is referenced in state guidance, even non-binding guidance, gains a veneer of official sanction. Disability advocates at organizations including the International Association of Assistance Dog Partners have consistently challenged this framing, noting that official-looking materials do not change the federal legal baseline.
Engineers building compliance tools for multi-state businesses need to build for federal floor requirements, not state-level variations that may themselves be legally vulnerable. A tool that helps a business "comply" with a state registry scheme may simultaneously put that business in violation of Title III. That is not a compliance tool. That is a liability transfer mechanism.
What AI Can and Cannot Do Legally
This is the question that ServiceDog.AI is built to engage directly. AI has genuine and powerful applications in the service dog space. Computer vision models trained on large canine behavioral datasets can evaluate task performance with meaningful accuracy. Gait analysis pipelines can detect learned postural responses that correlate with trained interruption behaviors. Biometric pairing models can authenticate handler-dog team identity over time. These are real capabilities with real value.
None of them are legally permissible as gatekeeping tools at the point of public access under current ADA framework.
The legal constraint does not make the technology worthless. It redirects where that technology should sit in the workflow. AI-based behavioral assessment is appropriate for training programs, where trainers use video analysis to evaluate task acquisition. It is appropriate for certification bodies like Assistance Dogs International, where independent evaluation of trained task performance is part of a voluntary accreditation standard. It is appropriate for handler self-assessment tools that build handler confidence and training accountability without creating a documentation artifact that businesses then leverage as a condition of access.
The TheraPetic® Training Plus program, accessible through officialservicedog.com, approaches this correctly. AI-assisted task analysis supports the training pipeline. It does not generate a certificate that a CVS or a Delta gate agent can demand. The distinction is architecturally and legally significant. The output of the model serves the handler's training journey, not a business's desire to offload access decision liability onto a third-party app vendor.
Engineers should also understand what AI cannot do reliably in this context. No currently published model can determine from a brief visual observation whether a dog has been trained to perform a specific task for a specific person's specific disability-related need. The behavioral markers of trained task work are context-dependent, handler-specific and often invisible to brief observation. A model that claims otherwise is overstating its capability. The ADA's two-question framework exists precisely because this assessment is genuinely difficult, and Congress placed the determination of necessity on the handler, not on a third party.
The Disability Community Position
The disability community's position on verification apps is not uniform, but the predominant voice from established advocacy organizations is one of strong skepticism verging on opposition. That position deserves careful engagement rather than dismissal.
Handlers who have lived experience with access denials understand something that engineers sometimes miss: documentation requirements do not actually solve the problem of fake service dogs. Someone willing to fraudulently pass off a pet as a service dog is equally willing to pay fifteen dollars for a fake certification badge. Verification apps that charge handlers for registration do not screen for trained task work. They screen for willingness to pay. The handler with a fully trained psychiatric service dog who refuses on principle to pay for a registry they know is legally meaningless gets the same green checkmark as the person with a golden retriever in an Amazon vest.
The IAADP and Assistance Dogs International have both published positions emphasizing that the only meaningful verification of service dog status is behavioral observation of trained task work, and that this observation is appropriately conducted by trainers and certification bodies, not by retail staff using a smartphone app during a ten-second interaction.
Disability advocates also raise a civil rights framing that technologists should internalize. Requiring a person with a disability to produce documentation before accessing public space places the burden of proof on the person with the disability. That inversion of burden is not a minor administrative inconvenience. For people with PTSD, anxiety disorders, mobility impairments, and other conditions that make public confrontation acutely difficult, a demand for documentation at a store entrance can be a significant harm. The ADA places that burden on the business, which must take the handler's word absent specific indicators of direct threat or disruption. Verification apps systematically shift that burden back.
Building Compliant Verification Tools
If AI cannot function as a public access gatekeeper without violating the ADA, what can it do? The answer involves reframing the goal from access control to compliance support.
For businesses, a compliant AI tool assists staff in understanding and applying the two-question rule correctly. Natural language processing systems can provide real-time guidance during edge cases, helping a manager understand what questions are permitted, what follow-up is appropriate when a dog is behaving in a way that poses a genuine direct threat, and how to document incidents without creating illegal documentation requirements. This is ADA compliance assistance, not access gatekeeping.
For trainers, AI-based behavioral assessment tools provide genuine value in the training pipeline. Computer vision models like those built on top of DeepLabCut or SLEAP for multi-point canine pose estimation can evaluate whether a dog's response latency, task completion rate and distraction recovery meet threshold criteria. These outputs inform training decisions. They do not generate certificates for business scanning systems.
For handlers, AI-assisted self-evaluation tools support accountability and confidence. A handler who has used video analysis to document their dog's task acquisition over months of training has something genuinely meaningful, not a QR code that proves nothing, but a longitudinal behavioral record that demonstrates training investment and task reliability. That record has value in housing contexts under the Fair Housing Act, in employment contexts, and in the handler's own understanding of their dog's readiness.
Building for these use cases is both legally defensible and technically honest about what AI can and cannot assess reliably. ServiceDog.AI's technical work sits in this space. The goal is to advance the state of canine behavioral AI in ways that serve handlers and trainers without creating systems that courts will eventually hold to have violated the civil rights of people with disabilities.
The ADA two-question rule is not a bug in the system that technology should route around. It is a carefully considered protection. The engineers who understand that will build tools that last. The ones who do not will build a compliance liability that eventually finds its way to a federal courthouse.