Fact-Checking AI Content: Why Companies Need It by 2027

Companies are accountable for what their AI says, and nobody else is checking it. Courts and regulators in 2024 to 2026 held publishers responsible for AI errors, EU transparency law covers labelling but not accuracy, Google judges AI content on purpose and value, and about 9% of AI-generated facts about brands in one 2026 dataset were wrong. A claim-level check before publishing is the missing control.

Published 31 October 2026. This article compiles 40 sources with dates, sample sizes and the exact scope of each figure. One figure in my brief, a US$290,000 Deloitte refund, did not match the sources and is corrected below. See Methodology.

#The short version

Four things are true at the same time, and the sources support all of them.

1. The legal position is settled in the direction of the publisher. A Canadian tribunal (2024), the FTC (2024 to 2026), the SEC (2024) and UK regulators (October 2026) have each treated a false claim as the company's claim, regardless of whether a person or a machine wrote it.
2. The new transparency laws do not check accuracy. The EU AI Act's Article 50 (applying from 2 August 2026) and India's synthetic content rules (in force 20 February 2026) are about labelling and provenance. A correctly labelled false claim is still a false claim. A UK regulator said a label is not a "get out of jail free card".
3. The volume is high and the review is thin in places. Graphite estimates about half of newly published online articles are primarily AI-generated. In Ahrefs' 2026 survey, 7% of AI-using marketers said more than 75% of what they publish is pure AI, up from 4% in 2025.
4. The error rate is low per sentence and high per company. AI platforms were 95.9% accurate when they answered about one brand (Seer), yet in Searchable's dataset 75% of brands had at least one false fact. Mistakes concentrate in the facts only the company knows.

The rest of this article sets out the evidence, where the datasets disagree, and then my own analysis, labelled as such, of what a sound pre-publish check looks like.

#1. What AI errors have cost so far

Documented costs are mostly reputational and procedural, with a few hard numbers: a C$812.02 award against Air Canada, a refunded final instalment on a contract of nearly A$440,000 at Deloitte Australia, about $1.6 million paid for a Deloitte report in Newfoundland and Labrador that contained false citations, and at least 155 monetary sanctions against US filers of AI-invented citations through July 2026.

Deloitte Australia, 2025

In July 2025 Deloitte delivered a 237-page targeted compliance framework assurance review to Australia's Department of Employment and Workplace Relations (DEWR). Fortune reported on 7 October 2025 that the report contained citations to non-existent academic research papers, a fabricated quote attributed to a federal court judgment, references to non-existent reports attributed to law and software engineering experts, and a non-existent book attributed to a Sydney University professor. University of Sydney researcher Chris Rudge flagged the problems and the Australian Financial Review first reported them in late August 2025. A revised version, with an update dated 26 September 2025, removed the fabricated material and disclosed that Azure OpenAI was used to create the report.

On money, the sources need care. City AM, 6 October 2025, put the contract at AU$440,000 and said Deloitte would repay the final instalment, without stating that instalment's amount. YourStory likewise reports a refund of "a portion" of a payment of nearly A$440,000. The figure of $290,000 that appears in several headlines, including Fortune's URL and JOPR, is attached to the value of the report in the text I could read, not to a stated refund. Its currency is not stated on those pages. It is consistent with a US dollar conversion of roughly A$440,000, but that is my inference. I therefore do not state a refund amount. What the sources do support is that the refund was partial, that it was the final instalment, and that Deloitte said in the revised report that the corrections did not affect the substantive content, findings or recommendations.

Deloitte Canada, Newfoundland and Labrador, 2025

On 25 November 2025 HR Reporter reported on the 526-page Health Human Resources Plan commissioned by the Newfoundland and Labrador government. The contract was worth $1,598,485 over two years, March 2023 to March 2025, according to an access-to-information request, and The Independent's investigation found at least four false citations. The province's health department said Deloitte acknowledged the four were incorrect and committed to a full review of every citation.

Reporting in March 2026 by CP24 (Canadian Press) added that Deloitte says the report was not written with AI and that it used AI "to support a small number of research citations". The contract had no provisions on whether or how AI could be used. The province said it was reviewing options related to compensation, and from December 2025 changed procurement so that vendors must state whether and how they intend to use AI, with the province able to audit that use. A separate 10-year education plan from the Education Accord team also had citations to apparently non-existent sources, with no contract value given in the source.

The pattern in both Deloitte cases is the same. The substance may have been sound, and the failure was in unverifiable detail that nobody checked before delivery.

Moffatt v Air Canada, 2024

In Moffatt v. Air Canada, 2024 BCCRT 149, decided 14 February 2024, the airline's website chatbot told a customer that reduced bereavement fares were available and that a ticket already flown could be submitted for a reduced rate within 90 days of issue. That was not the airline's policy. The tribunal rejected Air Canada's argument that it was not responsible for its chatbot, writing that "It should be obvious to Air Canada that it is responsible for all the information on its website" and that a chatbot "is still just a part of Air Canada's website". The award was $650.88 in damages, $36.14 in interest and $125 in fees, a total of $812.02.

The amount is small. The principle is not: the company owned the sentence the machine wrote.

Lawyers and invented citations

The best running count is Damien Charlotin's AI Hallucination Cases database, which has tracked court filings with fabricated material since April 2023. AFP reporting republished by Malay Mail on 16 June 2026 put it at 1,600 examples from 35 countries, with the US at 1,116, Canada 173, Australia 74 and Britain 59, and said such filings rose eightfold over the prior year compared with the 12 months before. The database is crowd-sourced and not definitive.

For the US, a Maryland State Bar Association article dated 27 July 2026, citing the same database, gave this breakdown:

Year	US instances	Pro se litigants	Lawyers	Monetary sanctions	Disciplinary referrals
2023	11	not stated	not stated	not stated	not stated
2024	37	not stated	not stated	not stated	not stated
2025	524	308	208	97	59
2026 (to late July)	668	405	256	58	21

The 2026 role figures sum to 667, one short of the stated 668, and the source does not explain the gap. By the source's count, 2026 had already passed the full 2025 total by 144 instances (about 27%) by late July. Fines in 2025 included 4 above $50,000 and in 2026 so far 2 above $50,000. Recent examples in the AFP report: a Louisiana judge fined a lawyer US$1,000 on 18 May 2026 after finding seven fabricated or misrepresented quotes, and a Mississippi federal judge fined four lawyers US$8,000 in total and barred two from the court for two years.

Roughly four in ten of the US instances in each year involved a lawyer (208 of 524 is 39.7%, and 256 of 668 is 38.3%), which is my arithmetic from the table, not a figure the source reports.

Other 2026 incidents

In March 2026 the OECD AI incidents monitor recorded that Ars Technica fired a senior AI reporter after an article he co-authored carried fabricated quotes attributed to a real person, generated by an AI tool. The article was retracted and the publication apologised. This is an editorial case, not a marketing one, but it shows the same failure: a plausible sentence, attributed to a real person, nobody confirmed.

On 28 May 2026 the Regional Court of Munich I (case 26 O 869/26) granted a temporary injunction against Google in a dispute with two Munich publishers, after Google's AI Overviews allegedly linked them to scams and subscription traps. According to that report, the court treated the AI Overviews as Google's own content, "entirely new, independent statements", rejected Google's argument that users could check the linked sources, and rejected its Digital Services Act host-provider defence. Google said it will appeal. This is a first-instance, interim ruling reported by a secondary outlet, so it is a signal, not settled law. It matters here because it is about false statements regarding companies made by an AI answer, the other direction from your own publishing.

#2. Regulation: what it covers and what it does not

The main rules in force by October 2026 target disclosure and deception, not accuracy. The EU's Article 50 (from 2 August 2026) is about labelling and marking. The FTC and SEC cases are about unsupported claims regardless of how they were produced. UK regulators say brand owners remain responsible. India's rules, in force 20 February 2026, are aimed at platforms and synthetic media.

EU AI Act, Article 50

Per Goodwin's August 2026 alert, Article 50 became generally applicable and enforceable by national authorities on 2 August 2026, whether or not a system is high-risk. The European Commission's AI Act Service Desk summary of the text sets out four duties:

• Providers must make sure people know they are interacting with an AI system, unless that is obvious.
• Providers of systems generating synthetic audio, image, video or text must mark outputs in a machine-readable format so they are detectable as artificially generated. Assistive editing functions and systems that do not substantially alter input are outside this.
• Deployers using emotion recognition or biometric categorisation must inform the people exposed.
• Deployers must disclose deepfakes, and must disclose AI-generated or manipulated text "published to inform the public on matters of public interest", unless the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication.

That last exception matters for companies. The Commission's page on the guidelines says deployers must inform people when exposed to "text publications on matters of public interest without human review or editorial control". Burges Salmon, 28 July 2026, reports that the Commission reads public interest broadly (politics, public health, consumer safety, economic or cultural developments) and that spelling or grammar correction does not meet the exception. The Commission expects "substantive review by relevant human decision makers" or an editorial entity able to approve, alter or reject text on substantive grounds. The same article says it does not specifically address marketing text, so whether a company's blog post or LinkedIn article falls within "public interest" will turn on its subject.

Three precise points:

• It is labelling, not accuracy. Nothing in Article 50 as summarised requires a statement to be true.
• There is a grace period for marking. Generative systems already on the market before 2 August 2026 have until 2 December 2026 for the Article 50(2) marking duty (Goodwin; Burges Salmon).
• Fines are material. Article 50 non-compliance can draw up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Goodwin, citing Article 99(4)). Goodwin is a law firm summary, so check the Official Journal text before relying on the figure.

The date that is relevant to a "2027" planning horizon is a different one. The AI Omnibus, Regulation (EU) 2026/1744, in force 27 July 2026, moved the standalone high-risk obligations to 2 December 2027 and product-embedded ones to 2 August 2028, and left Article 50 on its existing timetable (Goodwin). Most marketing content is not high-risk. The relevance is that enforcement attention, guidance and national authority capacity will all be maturing across 2027.

United States: FTC and SEC

On 25 September 2024 the FTC announced Operation AI Comply, five cases involving deceptive AI claims. The chair at the time said "Using AI tools to trick, mislead, or defraud people is illegal." The cases:

Company	Allegation	Amount stated
DoNotPay	Promoted an AI "robot lawyer" without testing it against a human lawyer	$193,000 settlement
Ascend Ecom	Falsely promised AI tools would generate passive income	At least $25 million alleged consumer losses
Ecommerce Empire Builders	Promised an "AI-powered Ecommerce Empire" with no evidence for income claims	Programs near $2,000, storefronts tens of thousands
Rytr	AI writing tool generated detailed reviews unrelated to user input, likely false	None stated
FBA Machine	Guaranteed income from AI-powered storefronts	More than $15.9 million alleged consumer losses

Two later FTC matters are directly about claims a company makes. In the Workado order, approved 28 August 2025, the FTC required the maker of an AI content detector to stop making accuracy claims without "competent and reliable evidence", to keep that evidence, to notify customers and to file annual compliance reports for four years. The tool was reportedly trained almost exclusively on academic content and performed poorly outside it. Per DLA Piper, on 27 January 2026 the FTC announced a resolution with Growth Cave over its claim that its GrowthBox "AI software" would automate nearly 100% of setting up and running an online course; the proposed orders bar misrepresenting that AI will improve profitability, effectiveness or efficiency.

The SEC acted earlier. On 18 March 2024 it announced civil penalties of $225,000 against Delphia (USA) Inc. and $175,000 against Global Predictions Inc., $400,000 in total, over false claims about using AI. The SEC chair said investment advisers "should not mislead the public by saying they are using an AI model when they are not."

On reviews, the FTC's rule banning fake and misleading reviews and testimonials was announced on 14 August 2024. TechCrunch's summary says it covers fake reviews including AI-generated ones, reviews from people without real experience, and fake engagement such as bot-generated followers. Morgan Lewis reports publication in the Federal Register on 22 August 2024, an effective date of 21 October 2024 and a civil penalty of up to $51,744 per violation, adjusted annually. By January 2026, FoodNavigator reported that the FTC had sent warning letters to 10 undisclosed companies during the 2025 holiday season and cited penalties of up to $53,088 per violation. I did not find a first enforcement case under the rule in the sources checked.

United Kingdom

The ASA reported in November 2024 that it identified 16,000 ads using the term "AI" in July to September 2024, that its AI monitoring scans over 3 million ads a month, and that advertisers making specific AI performance claims must be able to back them with evidence. In an October 2026 write-up of a panel, Lewis Silkin reported that the ASA's chief executive said the CAP and BCAP Codes are principles-based and need no AI-specific provisions, that an AI label is not a "get out of jail free card", and that a CMA director said whether a claim is made by a human or AI "is largely irrelevant". The CMA had issued consumer protection and AI guidance earlier in 2026. One ASA ruling cited, BCCR BeliefCoding, involved an ad claiming 83% of users saw improvement, where AI reportedly produced an inaccurate summary of the underlying paper. I could not locate the ruling text itself, so treat that detail as reported.

India

India's Ministry of Electronics and Information Technology amended the IT Rules, 2021 on 10 February 2026, effective 20 February 2026 (Hogan Lovells). Permitted synthetic content must be prominently labelled and carry lasting metadata or origin markers, significant social media intermediaries (more than 5 million registered users) must require users to declare synthetic content and verify declarations, and intermediaries must act on government or court orders within three hours, down from 36. The obligations fall on intermediaries and synthetic media tools. The source does not describe a duty on an ordinary company to fact-check its own marketing, and it does not specify fines.

#3. Platforms and search: what they reward and what they label

Google says it judges content on purpose and usefulness, not on whether AI wrote it: its March 2024 scaled content abuse policy applies "whether automation, humans or a combination are involved". LinkedIn and others show C2PA Content Credentials on images, but the standard records how media was made, not whether a claim is true, and platform pipelines often strip it.

Google

Google's February 2023 post, Google Search's guidance about AI-generated content, says it explains how AI content fits into its "long-standing approach to show helpful content to people on Search". I could read only the page's metadata, so for the wording of the policy I rely on PPC Land's account. It reports that on 5 March 2024 Google introduced the scaled content abuse policy alongside expired domain abuse and site reputation abuse, that Elizabeth Tucker, a Google director of product management, said the policy applies "whether automation, humans or a combination are involved", and that Google projected a 40% reduction in low-quality, unoriginal content in results, then reported 45% on 26 April 2024 after the rollout ended on 19 April. That 45% comes from Google's own evaluations with no published method or outside audit.

The policy, in PPC Land's summary, names generating many pages with generative AI without adding value as one pattern, next to scraping, stitching content from different pages, spreading output across sites, and keyword-heavy pages that make little sense. In May 2026 Google clarified that its spam policies also apply to generative AI responses in Google Search, including AI Overviews and AI Mode, in a documentation update published 15 May 2026, adding no new rules.

The practical reading: AI authorship is not the trigger. Scale without value is. Factual accuracy is not named as a ranking test in these sources. It matters through other routes: user trust, regulators, and the AI answer engines that quote you.

C2PA Content Credentials

Fortune, 26 November 2024, described LinkedIn adding a visible "CR" label to uploaded content carrying C2PA credentials, which users can click to see the tool used, the signer and the signing timestamp. Limits it lists: it depends on creators applying credentials, some platforms' video encoders strip them, and users misunderstand the symbol. A 2026 adoption review by SoftwareSeni says fewer than 1% of news images and videos published globally include C2PA metadata (citing the Reuters Institute), and that C2PA "records what was declared at signing" and does not detect AI-generated content. It also notes TikTok adopted Content Credentials and that Cloudflare supports them when the site owner opts in. I did not verify the Reuters Institute figure at its source.

Provenance for text is a different problem. C2PA is mostly an image and video standard in these sources. A sentence claiming "we cut onboarding time by 40%" has no credential that says whether it is true.

How AI answer engines repeat claims about companies

Three datasets bear on this, and they measure different things.

• News, not brands. The EBU and BBC study, published 22 October 2025, assessed more than 3,000 AI assistant responses across 18 countries and 14 languages with 22 public service media organisations. 45% had at least one significant issue, 31% had sourcing problems and 20% had major accuracy problems. Gemini had significant issues in 76% of responses.
• Brand facts reviewed by the brands themselves. Searchable's data, with participation to 29 July 2026 and updated 30 July, covers 1,704 brands and 32,556 AI-generated facts. Brands marked 3,005 (9.2%) incorrect. Among brands that reviewed 20 or more facts, 75% had at least one false fact. False-fact rates by engine: Perplexity 11.5%, Google AI Overviews 11.4%, ChatGPT 8.1%. The page groups errors as outdated details, mistaken identity (one brand confused with another) and fabrications. The sample is Searchable's customers, not random, and an employee grades each fact.
• One brand, many prompts. Seer Interactive's study, 27 August 2026, ran 1,562 prompts on six platforms from 1 May to 9 June 2026 and analysed 28,123 responses against 87 attributes of Seer itself. When the platforms answered, they were 95.9% accurate; 2.8% of responses were inaccurate and 31.2% were no-answer. Comparative prompts were 18.8% accurate with 80.3% no-answer. It is self-published, about one company.

For how often AI tools cite companies at all, see my own data in Who AI Cites for B2B Buyer Questions and the compiled AI search statistics for 2026.

#4. Volume: how much is AI-written, and who reviews it

Graphite estimates about half of new online articles are primarily AI-generated, with the share flat near 50% since Q1 2025. Ahrefs found 93% of surveyed marketers use AI for content and 7% say more than 75% of what they publish is unedited AI. McKinsey's 2025 survey, as reported, had 23% of organisations scaling agentic systems.

Graphite, 15 May 2026. Graphite's analysis drew 55.4 thousand English-language articles at random from Common Crawl, published January 2020 to March 2026, and classified them with three detectors (Pangram, Copyleaks, GPTZero), averaging the results. The AI share was 49.6% in Q1 2025, 50.9% in Q4 2025 and 49.9% in Q1 2026. Reported false positive rates on 15.7 thousand pre-ChatGPT articles were 1.355% to 1.844% per detector. Limits stated: Common Crawl is a proxy for the web, AI-assisted articles with human editing were not evaluated, and detection accuracy is disputed. The authors hypothesise, without direct proof, that the plateau reflects practitioners finding AI-written articles performed poorly in search.

Ahrefs, September 2026. In Ahrefs' survey of 301 marketers (published 30 September 2026, modified 2 October), 93% (280 of 301) use AI to create or help create content. Among AI users, 7% say more than 75% of their published content is pure AI, up from 4% in 2025. 69% of all respondents (74% of AI users) publish more AI content than last year. 66% trust AI-generated content only if a human reviews it first, and 2.7% trust it with no review. The survey dates are not stated and the sample is small.

McKinsey, 2025. As reported by CX Today, McKinsey's State of AI in 2025 found 88% of respondents' organisations regularly use AI in at least one function, 62% are at least experimenting with agents, and 23% are scaling agentic systems somewhere. 51% had experienced at least one negative consequence, and inaccuracy was the most common at 30%. I read this second-hand and did not verify it on McKinsey's site.

Two inferences, labelled as such. First, I found no survey measuring how many agents publish to external channels with no human approval step, so the share of unreviewed machine publishing is unmeasured. Ahrefs' 7% is the nearest proxy and covers only marketers who responded. Second, the 23% scaling agents is across business functions, not publishing.

For the LinkedIn-specific picture, including detector-based prevalence estimates and what disclosure does to trust, see How Much of LinkedIn Is AI-Written.

#5. What the data disagrees on

Four disagreements matter: whether newer models hallucinate less (benchmarks are mixed), whether brand answers are accurate (95.9% when answered versus 75% of brands with a false fact), whether AI content volume is still rising (flat per Graphite, rising per Ahrefs), and whether a label helps (regulators say it does not cure a false claim).

Newer models versus fewer errors. The Vectara HHEM leaderboard (as relayed by an aggregator; results updated 6 May 2026; 102 models) shows the lowest hallucination rate at 1.8% (antgroup/finix_s1_32b) and GPT-5.4 Nano at 3.1%, with the highest at 24.2% (Ministral 3B). Within the table, Gemini 3 Pro (13.6%) is higher than Gemini 2.5 Pro (7.0%), and GPT-5.2 at high effort (10.8%) is worse than at low effort (8.4%). These are summarisation tasks on provided documents, and the page does not draw the conclusion that newer is worse. So the honest reading is that rates have fallen for some models and not others, and the benchmark does not predict accuracy on facts outside the supplied text.

Why company-specific facts are different. OpenAI's 5 September 2025 paper explanation says arbitrary, low-frequency facts cannot be inferred from patterns, so models hallucinate on them, and that evaluations reward confident guessing over abstaining. Its example: on SimpleQA, gpt-5-thinking-mini abstained 52% of the time with 26% error, while o4-mini abstained 1% with 75% error and slightly higher accuracy (24% versus 22%). Your headcount, pricing tier, customer list, certifications and the date of your last funding round are exactly this kind of fact. A model that has not seen your approved version can only guess, abstain, or repeat an outdated page.

Brand accuracy: 96% or 91%, depending on what you count. Seer found 95.9% accuracy when answered; Searchable found 9.2% of facts false. Both can be true: Seer counts responses, Searchable counts facts reviewed by brands, Seer reports accuracy only among answered prompts while Searchable counts individual facts, and the samples differ (one brand versus 1,704 mostly customer brands). The shared signal is that errors are a minority of statements but a majority of brands carry at least one. Searchable classes its false facts as outdated details, mistaken identity or fabrication. Neither study can say how much a false fact costs.

Is AI volume still rising? Graphite shows a plateau in share near 50% since Q1 2025. Ahrefs shows the share of marketers publishing mostly unedited AI rising from 4% to 7% and 69% publishing more than last year. These are different populations (random web articles versus surveyed marketers) and different measures (share versus change in volume per marketer). Both can hold if the total number of articles keeps growing at a constant AI share.

Do labels solve it? Article 50 and India's rules rely on labelling and marking. UK regulators, per Lewis Silkin, say a label does not cure a misleading claim. The Commission's own reading of the human review exception implies that substantive review, not disclosure, is what removes the duty to label public-interest text. Labelling and accuracy are separate controls.

#6. Shruti's analysis

This section is my reasoning, not sourced fact. My view is that the risk has moved from how well AI writes to whether what it writes about your company is true. A sound pre-publish check has four parts: saved facts, claim-level checking, a named approver and a record. I founded Klype AI, which builds a check of this kind, so I have a stake in this argument.

The risk moved. For two years the common worry about AI content was quality: generic prose, sameness, detector scores. The sources above say the exposure now sits elsewhere. Google's policy asks about purpose and value. LinkedIn's stated test, in the dataset I compiled earlier, is perspective. The cases that cost money or standing (Deloitte, Air Canada, the filings, the quotes at Ars Technica) were fluent. They failed on a specific, checkable statement. Fluency is now cheap, and a confident wrong sentence reads the same as a right one.

Why it concentrates on "facts about the company". A model can write a reasonable paragraph about onboarding. It cannot know your current onboarding time, which customers you may name, or that last quarter's number was restated. OpenAI's explanation of hallucination and Searchable's finding on outdated details point the same way: the facts that are least likely to be in a model are the ones about you. The same applies in reverse when AI engines describe you, which is the reason the Munich ruling matters, though the remedy there falls on the platform, not on you.

Why 2027 rather than now. None of the sources identifies 2027 as a cliff. It is a planning judgement based on three trends in them: the weight of precedent is accumulating (the lawyer database grew from 37 US instances in 2024 to 524 in 2025 and 668 by late July 2026), EU and UK regulators are building enforcement capacity from 2026, and AI-agent adoption is moving from experiments (62%) to scale (23%) in McKinsey's numbers. If agents publish at the rate people draft today, the number of unchecked claims per month grows faster than any editor can read.

What a sound pre-publish check looks like. I would separate four layers.

1. Saved facts. One maintained set of approved statements about the company: legal name, founding year, headcount range, product list and what each does, customer names you may use, certifications with dates, figures with their source and as-of date, and statements you never make. Every check compares against this, not against the writer's memory or the model's. Each fact has an owner and a review date.
2. Claim-level checking. Split a draft into claims: numbers, names, dates, comparisons ("faster than"), superlatives ("only", "first"), customer results, legal and compliance statements, and any quote or citation. Each claim gets one of three outcomes: matches a saved fact, supported by a linked primary source, or unsupported. Unsupported claims are cut or sent to a person. This is what the Workado order and the ASA's evidence standard describe in regulatory terms: a claim needs "competent and reliable evidence" before it runs.
3. A named approver. Article 50's exception for public-interest text turns on a natural or legal person holding editorial responsibility, and the Commission expects substantive review. Newfoundland and Labrador's procurement changes (vendors declare how they use AI, the province can audit it) point the same way. A named person, with authority to reject a draft, is both a control and the thing a regulator or court will look for.
4. A record. Keep the draft, the claims extracted, the fact or source each was matched to, who approved, and when. The Deloitte cases were flagged by outside readers after delivery. A record lets you show what was checked, correct quickly, and answer a regulator or client.

What it does not do. A claim-level check cannot prove an insight is good, cannot catch a true statement that is misleading in context without a person reading it, and does not replace legal review of regulated claims. It is also only as good as the saved facts. Stale facts produce confidently approved errors.

Where I could be wrong. If model providers reliably abstain on unknown company facts (Claude's 46.3% no-answer rate in Seer's data shows one such pattern), the main exposure shifts from invented facts to omission. If regulators keep to labelling and leave accuracy to existing advertising law, the pressure will come from clients and procurement rather than law, as in Newfoundland and Labrador. Either way a check is still cheap relative to a retraction.

I founded Klype AI, which builds a pre-publish check of the kind described above. I have no results or user figures to report here and I have not tested any claims in this article with it.

#What this means for a company publishing with AI

1. Decide who owns accuracy. Name one person per channel with authority to reject a draft. Do this before choosing tools.
2. Write the fact set once. Start with the 30 to 50 statements most often repeated about the company, with owner and as-of date. The number is a practical starting suggestion, not a benchmark from a source.
3. Check claims, not tone. Extract numbers, names, comparisons, quotes and citations from every AI draft and match each one.
4. Audit what AI says about you quarterly. Ask the major AI tools the same 20 to 30 buyer questions and compare with your fact set. Searchable's 75% suggests most companies will find at least one error. The method in How to Get Your Brand Mentioned by ChatGPT covers the visibility side, and my AI visibility check is one way to start.
5. Keep a log. Who approved what, against which fact, on what date.
6. Check vendors too. Ask agencies and consultants whether and how they use AI on your work, as Newfoundland and Labrador now does.

#Methodology

This is a literature and source compilation. I ran no experiments and no new analysis of content. For every source I recorded publisher, date, sample size where stated, and whether it is a primary document, a law firm summary, a news report or a vendor study. Work was done on 31 October 2026.

Primary sources I read directly: the British Columbia Civil Resolution Tribunal decision; the FTC press release of 25 September 2024; the SEC press release of 18 March 2024; the EBU press release; the Searchable and Seer Interactive pages; Graphite's report; Ahrefs' article; OpenAI's article; the European Commission AI Act Service Desk and digital strategy pages. Law firm and trade summaries stand in for primary text where I could not retrieve it, and the text names each: Goodwin and Burges Salmon for the AI Act and Omnibus dates and penalties; Hogan Lovells for India's rules; Morgan Lewis and TechCrunch for the FTC reviews rule; Lewis Silkin and Mishcon de Reya for the ASA and CMA; PPC Land for Google's policy wording.

Corrections and limits:

• Deloitte refund. My brief described a partial refund of about US$290,000. The sources show a contract of nearly A$440,000, a refund of the final instalment of unstated size, and a $290,000 figure attached to the report's value without a stated currency. I do not state a refund amount.
• Currency. The Newfoundland and Labrador figure and the Air Canada award are given as the sources print them. They are in Canadian dollars by context, which I infer.
• Google. I could not retrieve the body of Google's 2023 and 2024 Search Central posts, so the policy quotations come from PPC Land and are second-hand. The 45% result is Google-reported.
• Munich ruling. Reported by a secondary outlet, interim, and under appeal.
• EU text. I did not read the Official Journal text. The Article 50 description rests on the Commission's service desk summary and law firm alerts, which say the summary is non-binding.
• FTC penalty figure. $51,744 is the figure at publication of the rule and $53,088 the figure cited in January 2026 after inflation adjustment.
• McKinsey and Vectara. Read via secondary pages. Not verified at origin.
• Counts. Charlotin's database is crowd-sourced. The AFP count (1,600, global, 16 June 2026) and the MSBA count (1,240, US only, 27 July 2026) are different scopes and dates and should not be added.
• Not found. I found no study measuring the share of AI agents that publish externally without human review, no first FTC enforcement case under the fake reviews rule in the sources checked, and no figure for the cost of a false AI-generated fact to a brand.
• Not legal advice. This summarises public sources.

#Sources

1. "Deloitte to partially refund Australian government for report with apparent AI-generated errors" (via Fortune), 7 October 2025. https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund
2. "Deloitte refunds Australian government after AI 'made up citations' in report," City AM, 6 October 2025. https://www.cityam.com/?p=2367690
3. "Deloitte to partially refund Australian govt after using AI in $440,000 report," YourStory. https://yourstory.com/ai-story/deloitte-repay-australia-ai-report-errors
4. "When AI Hallucinates: Deloitte's $290K Refund After Faulty Report," JOPR, 8 October 2025. https://jopr.org/news/detail/when-ai-hallucinates-deloittes-290k-refund-after-faulty-report
5. "AI errors: Province grapples with Deloitte report marred by false citations," HR Reporter, 25 November 2025. https://www.hrreporter.com/focus-areas/automation-ai/ai-errors-province-grapples-with-deloitte-report-marred-by-false-citations/393782
6. "Newfoundland and Labrador overhauls procurement after report errors thought to be AI," CP24 (Canadian Press), 11 March 2026. https://www.cp24.com/news/canada/2026/03/11/newfoundland-and-labrador-overhauls-procurement-after-report-errors-thought-to-be-ai/
7. Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia, 14 February 2024. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
8. "Do your job: Global courts face rising tide of AI-generated fake cases and quotes," AFP via Malay Mail, 16 June 2026. https://www.malaymail.com/news/tech-gadgets/2026/06/16/do-your-job-global-courts-face-rising-tide-of-ai-generated-fake-cases-and-quotes/223865
9. "Confronting the Hard Numbers Behind Lawyers, AI and Phantom Precedents," Maryland State Bar Association, 27 July 2026. https://www.msba.org/site/content/News-and-Publications/News/General-News/Confronting_the_Hard%20Numbers_Behind_Lawyers_AI_and_Phantom_Precedents.aspx
10. "Ars Technica Fires Reporter After AI-Generated Fabricated Quotes Published," OECD.AI incidents monitor, 3 March 2026. https://oecd.ai/en/incidents/2026-03-03-880a
11. "German Court Finds Google Liable for AI Summaries," Let's Data Science, 2026 (Regional Court of Munich I, 26 O 869/26, 28 May 2026). https://letsdatascience.com/news/german-court-finds-google-liable-for-ai-summaries-54c1ff70
12. "Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force," Goodwin, August 2026. https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force
13. "Article 50: Transparency obligations for providers and deployers of certain AI systems," European Commission AI Act Service Desk. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
14. "Guidelines on transparency of AI-generated content," European Commission. https://digital-strategy.ec.europa.eu/policies/guidelines-transparency-ai-generated-content
15. "EU AI Act transparency guidance lands," Burges Salmon, 28 July 2026. https://www.burges-salmon.com/articles/102nekf/eu-ai-act-transparency-guidance-lands-what-providers-and-deployers-need-to-do-no
16. "FTC Announces Crackdown on Deceptive AI Claims and Schemes," Federal Trade Commission, 25 September 2024. https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes?os=f
17. "FTC's AI crackdown continues: Workado ordered to drop unsupported accuracy claims," Frankfurt Kurnit Klein & Selz, 4 September 2025. https://advertisinglaw.fkks.com/post/102l358/ftcs-ai-crackdown-continues-workado-ordered-to-drop-unsupported-accuracy-claims
18. "FTC resolves another case involving AI washing," DLA Piper, February 2026. https://www.dlapiper.com/en/insights/publications/2026/02/ftc-resolves-another-case-involving-ai-washing
19. "SEC charges two investment advisers with making false and misleading statements about their use of AI," U.S. Securities and Exchange Commission, 18 March 2024. https://www.sec.gov/newsroom/press-releases/2024-36
20. "FTC finalizes rule banning fake reviews," TechCrunch, 14 August 2024. https://techcrunch.com/2024/08/14/ftc-finalizes-rule-banning-fake-reviews
21. "FTC Issues Final Rule on Consumer Reviews and Testimonials," Morgan Lewis, August 2024. https://www.morganlewis.com/pubs/2024/08/ftc-issues-final-rule-on-consumer-reviews-and-testimonials
22. "FTC warns about fake reviews in letters that could portend enforcement push," FoodNavigator, 7 January 2026. https://www.foodnavigator.com/Article/2026/01/07/ftc-signals-crackdown-on-fake-reviews-and-testimonials/
23. "ASA's approach to AI in advertising regulation," Mishcon de Reya, November 2024. https://www.mishcon.com/news/asas-approach-to-ai-in-advertising-regulation
24. "AI in advertising: regulators signal that existing rules still apply," Lewis Silkin, 2 October 2026. https://www.lewissilkin.com/insights/2026/10/02/ai-in-advertising-regulators-signal-that-existing-rules-still-apply
25. "India introduces mandatory labelling for AI and 3-hour takedown for illegal content," Hogan Lovells, 2026. https://www.hlc.com/en/publications/india-introduces-mandatory-labelling-for-ai-and-3hour-takedown-for-illegal-content
26. "Google Search's guidance about AI-generated content," Google Search Central Blog, February 2023. https://developers.google.com/search/blog/2023/02/google-search-and-ai-content
27. "Scaled content abuse," PPC Land (reporting Google's spam policies, March 2024 announcement). https://ppc.land/scaled-content-abuse/
28. "Google spam policies now officially cover AI Overviews and AI Mode in Search," PPC Land, 16 May 2026. https://ppc.land/google-spam-policies-now-officially-cover-ai-overviews-and-ai-mode-in-search
29. "LinkedIn content credentials and labeling challenges," Fortune, 26 November 2024. https://fortune.com/2024/11/26/linkedin-content-credentials-labeling-challenges-c2pa
30. "C2PA Adoption in 2026: Hardware, Platforms and Verification Reality," SoftwareSeni. https://softwareseni.com/c2pa-adoption-in-2026-hardware-platforms-and-verification-reality
31. "AI's systemic distortion of news is consistent across languages and territories," European Broadcasting Union, 22 October 2025. https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste
32. "The AI Brand Misinformation Report," Searchable, updated 30 July 2026. https://www.searchable.com/data/ai-brand-misinformation-report
33. "Brand Accuracy in AI 2026," Seer Interactive, 27 August 2026. https://www.seerinteractive.com/insights/ai-brand-accuracy-study
34. "AI Now Writes as Many Online Articles as Humans," Graphite, 15 May 2026. https://graphite.io/five-percent/ai-now-writes-as-many-online-articles-as-humans-do
35. "69% of Marketers Are Publishing More AI Content Than Last Year," Ahrefs, 30 September 2026. https://ahrefs.com/blog/marketers-ai-content-creation
36. "McKinsey's State of AI: The Scaling Gap Is Now CX's Problem," CX Today (reporting McKinsey, State of AI in 2025). https://cxtoday.com/mckinseys-state-of-ai-the-scaling-gap-is-now-cxs-problem
37. "Why language models hallucinate," OpenAI, 5 September 2025. https://openai.com/index/why-language-models-hallucinate/
38. "Vectara HHEM Hallucination Leaderboard," Benchmarklist (results updated 6 May 2026). https://benchmarklist.com/benchmarks/vectara_hhem/
39. "Who AI Cites for B2B Buyer Questions," shrutikaushik.com. https://shrutikaushik.com/writing/who-ai-cites-b2b-buyer-questions-2026
40. "How Much of LinkedIn Is AI-Written? Reach and Trust Data (2026)," shrutikaushik.com. https://shrutikaushik.com/writing/linkedin-ai-written-posts-reach-trust-data-2026

By Shruti Kaushik