Surveys show AI workplace adoption outpacing firm safeguards
The warning came in the September 2026 decision in Clinco v. Commissioner, according to Sarah Webber, an associate professor of accounting at the University of Dayton, and Kaitlin Newkirk, an assistant professor of accounting at Xavier University, both certified public accountants who research tax accounting. The two analyzed the case in the September 2026 issue of the Journal of Accountancy.
The case involved a restaurant owner accused of underreporting more than $2 million in business receipts and claiming unsupported depreciation deductions on rental properties. The taxpayers’ attorney, Abraham Wagner, cited four cases to support his argument that an IRS notice assessing additional taxes was invalid. The court found that three of the citations appeared to be fabricated — AI-generated hallucinations pointing to nonexistent Tax Court cases — and rejected the taxpayer’s argument. The court did not sanction Wagner in the case but warned that future submissions of fake citations could lead to financial penalties or disciplinary action against the attorney.
Courts and tribunals have also held employers and AI developers accountable for AI errors delivered directly to clients and customers. Deloitte produced a 237-page report in 2025 for the Australian government examining that country’s automated system for penalizing welfare recipients who failed to meet job-seeking requirements; the report contained nonexistent academic citations and a fabricated quotation attributed to a federal court judgment. Deloitte later issued a revised version, attributed the errors to generative AI and agreed to partially refund its fees.
Air Canada’s website chatbot told a customer in November 2022 that he could purchase a ticket under the airline’s bereavement policy and seek reimbursement later — a policy the carrier said did not exist. Air Canada refused to honor what the chatbot had told the customer. The customer sued, and the British Columbia Civil Resolution Tribunal awarded 812 Canadian dollars, ruling that Air Canada had failed to take reasonable care to ensure its chatbot provided accurate information and had negligently misrepresented the bereavement policy.
The courts have also tested the liability of AI platforms directly. A Georgia court in 2025 dismissed Walters v. OpenAI, in which the radio host Mark Walters alleged that ChatGPT falsely accused him of embezzlement; the court granted summary judgment for OpenAI, finding Walters had not established fault and damages. A similar suit, Keene v. Google LLC, brought by a TV producer after Google’s AI summaries allegedly published false information about his criminal history, was allowed by the court in September 2026 to proceed in part, without deciding Google’s liability.
The terminology regulators use for such failures is still being refined. The National Institute of Standards and Technology, in a July 2024 report on AI risk management, labeled AI mistakes and hallucinations “confabulations” — the confident presentation of erroneous or false information. Some researchers draw a sharper distinction, reserving “hallucinations” for fabrications from whole cloth and “confabulations” for plausible distortions of real information; both forms can mislead users who do not verify AI outputs independently.
The volume of unreliable output remains substantial. A Stanford University 2024 study of specialized AI tools for legal research found erroneous results in 17% to 33% of responses tested. Yet, as Webber and Newkirk note, those findings have not stopped many law firms and other employers from embracing AI tools for the time savings and efficiency gains they promise.
Workplace safeguards lag adoption. A 2026 Nationwide survey found 36% of surveyed businesses have written AI policies and 37% provide any training on responsible AI use. Webber and Newkirk write that businesses cannot rely on telling staff to verify AI outputs and must instead set clear expectations about when AI is appropriate, which outputs need independent review, and who bears responsibility for the final work product.
AI may change how professional work gets done, the authors write, but it does not change where the buck stops — which they place with the professional or firm using the tool, not with the developer.