AI is now saying things about your company that are not true. What a pre-publish claim check catches, and the evidence

What does the evidence on false AI claims about companies show?

Across 19 primary sources published between 2022 and 2026, the pattern is consistent: AI answer engines state false facts about brands at measurable rates, AI drafts invent names, figures and citations, and general benchmarks do not measure company-specific facts. The table below compiles every figure used in this article, with its sample and date.

Every figure was opened and checked on its source on 5 November 2026. Where a figure comes from a vendor that sells a related product, the table says so, because that is part of how much weight it can carry.

What kinds of claims does AI invent about a company in its drafts?

Five kinds recur in the research: features and policies the company does not have, integrations and tools that do not exist, customers and testimonials nobody gave, numbers with no source, and names or citations that point nowhere. Each fails in a different way, so one generic "fact check" step misses some of them.

The useful way to think about this is not "AI makes mistakes" but "AI fills gaps". When a model is asked to write about a company and does not have a fact, it rarely leaves a blank. It writes the most plausible sentence. The five types below are what plausible looks like when it is wrong.

1. Features and policies that do not exist

The clearest legal record is Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, 14 February 2024). The airline's website chatbot described a bereavement fare refund process that did not match the airline's actual policy. Air Canada argued the chatbot was responsible for its own words. The tribunal called that "a remarkable submission" and wrote: "It should be obvious to Air Canada that it is responsible for all the information on its website." It ordered a total of $812.02.

A second example is a software company's own support channel. The Register (18 April 2025) reported that users of Cursor, the code editor, were told by an AI support reply that being logged out across machines was expected behaviour. Co-founder Michael Truell answered publicly: "Hey! We have no such policy... You're of course free to use Cursor on multiple machines." The model had produced a reasonable-sounding policy that the company never set.

For marketing drafts, the equivalent is a post that says your product "integrates with your CRM in one click" or "supports SSO" because those phrases are common in the category. Klype's own evaluation (section 07) used two "trap briefs" for exactly this case: briefs that asserted features the fictional product did not have, to see whether the draft would repeat them.

2. Integrations, tools and dependencies that do not exist

The largest controlled measurement of invented names comes from code. Spracklen et al. (USENIX Security 2025) generated 576,000 code samples with 16 models and found that "the average percentage of hallucinated packages is at least 5.2% for commercial models and 21.7% for open-source models", including "205,474 unique examples of hallucinated package names". A package name is a short, specific, checkable string, which is what an integration name in a marketing draft is too. The study matters here because it shows invention is not limited to vague prose: models also invent precise proper nouns, at scale.

3. Customers, reviews and testimonials nobody gave

The US Federal Trade Commission's final rule (14 August 2024) bans reviews and testimonials that "misrepresent that they are by someone who does not exist, such as AI-generated fake reviews", and allows the agency to "seek civil penalties against knowing violators". For a B2B team the practical risk is less often a fake review site and more often a draft that says "teams at leading fintechs use us" or invents a quote from "a VP of Sales at a Series B company". Klype's 500-post study (9 October 2026) found that about one AI-written post in six told a story that never happened, when the model was given no facts about the author.

4. Numbers with no source

Numbers are where AI answer engines are most often wrong about brands. In Searchable's dataset (updated 30 July 2026), false facts were most common for "prices (13.5%) or funding and financial details (10.9%)", against 9.2% overall. In drafts, the same pattern appears as unsourced percentages: in Klype's 500-post study, 13% of posts quoted a percentage with no source. And it is not only AI. Our study of 104 B2B SaaS homepages (10 October 2026) found that 24.3% of 600 checkable claims had any visible source next to them, so the human-written pages that AI engines learn from are often unsourced too.

5. Names and citations that point nowhere

Fabricated references are the best-studied type. Walters and Wilder (Scientific Reports, 2023) checked 636 citations in 84 AI-written literature reviews: "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated", and 43% and 24% of the real ones had substantive errors. Newer systems still do it. Rao, Wong and Callison-Burch (University of Pennsylvania, 3 April 2026) tested commercial models and deep research agents on 53,090 and 168,021 citation URLs and found that 3 to 13% of citation URLs were hallucinated, with no record in the Wayback Machine, and 5 to 18% did not resolve. The same paper found that giving models a URL-checking tool cut non-resolving links by 6 to 79 times, to under 1%, depending on the model's skill with tools. That is the strongest single piece of evidence that a mechanical check, run before output, works.

How do AI search engines end up repeating false facts about brands?

AI answer engines build brand answers from whatever is on the web, and most of that is content brands publish themselves. When a company's own posts, pages or listings contain an invented or outdated claim, engines can repeat it. Brand-reviewed data puts false facts at 9.2% overall, and higher for prices, funding and AI companies.

Two 2026 studies measured this directly, and they look contradictory until you read their methods.

Searchable (AI Brand Misinformation Report, updated 30 July 2026) asked brands to review the facts AI engines state about them. Each fact was graded by an employee of the organisation it describes. Of 32,556 facts across 1,704 brands, 3,005 were marked false: 9.2%. Among brands that reviewed at least 20 facts, 75% found at least one false. By engine: ChatGPT 8.1% of 20,359 facts, Perplexity 11.5% of 6,676 and Google AI Overviews 11.4% of 4,881. "AI companies see false facts at nearly double the average rate (18.5%)." Searchable groups the errors into outdated details that were once true, mistaken identity between brands, and fabrications.

Seer Interactive (AI Brand Accuracy Study, 27 August 2026) tracked 1,562 prompts about Seer itself across AI Mode, AI Overviews, ChatGPT, Claude, Gemini and Perplexity from 1 May to 9 June 2026, 28,123 responses in all. "When AI platforms actually answered a branded prompt about us, they got it right 95.9% of the time." But they declined to answer 31.2% of branded prompts, and on comparative prompts the accuracy was "only 18.8%, with the AI platforms giving No Answer responses 80.3% of the time." The errors Seer did find are instructive: a former office location repeated 223 times, services Seer does not offer (109 times), an inaccurate headcount (85 times), and a "high account turnover" story built from a single 2018 review.

Seer's list is the five claim types from the previous section, appearing in answers instead of drafts: an outdated fact, an invented service, a wrong number, and a narrative stretched from one source.

The detailed story wins

Ahrefs ran a deliberate test (Mateusz Makosiewicz, 10 December 2025). It invented a brand, Xarumei, published an official site and then conflicting third-party stories, and asked 8 AI tools 56 questions. Its conclusion: "in AI search, the most detailed story wins, even if it's false." Search Engine Journal's critique (Roger Montti, 28 December 2025) is fair: "Xarumei existed in a vacuum", and "49 of the prompts consisted of leading questions, and only 7 were not." Read together, the experiment shows engines favour detailed, specific text when there is little else to go on. That describes most smaller B2B companies.

Your own content is most of the input

Yext (9 October 2025) analysed 6.8 million citations from more than 1.6 million AI responses across Gemini, OpenAI and Perplexity and found that "86% of citations come from sources marketers can directly manage or strongly influence": 44% from websites and 42% from listings. Yext sells listings management, so treat the framing with care, but the direction is consistent with Ahrefs and Seer. The pages, posts and profiles a company publishes are the main raw material for what engines say about it.

That closes a loop. An AI draft invents a feature; someone publishes it; an engine crawls it; the next buyer asks the engine and hears the feature as fact. Searchable's "outdated details that were once true" category is the slow version of the same loop. A pre-publish check is the one point in the loop the company fully controls.

Answers are also unstable

Two more studies frame how much weight any single AI answer deserves. The Tow Center for Digital Journalism, reported by Nieman Lab (10 March 2025), ran 1,600 queries on eight AI search engines: they "failed to retrieve the correct information more than 60% of the time", from 37% for Perplexity to 94% for Grok-3 Search. The EBU and BBC study (22 October 2025) evaluated more than 3,000 news answers from ChatGPT, Copilot, Gemini and Perplexity and found that "45% of all AI answers had at least one significant issue" and "20% contained major accuracy issues, including hallucinated details and outdated information." And SparkToro's work with Gumshoe, reported by Search Engine Journal (30 January 2026), ran 2,961 prompts and found less than a 1 in 100 chance that ChatGPT or Google's AI would return the same brand list twice. Those are news and recommendation tasks, not brand facts, but they show the answer layer is noisy. A company cannot fix the noise; it can make sure its own published facts are not part of it.

If hallucination rates are falling, why does AI still get company facts wrong?

Because benchmark rates measure tasks where the answer is in the prompt or common on the web. A company's current pricing, integrations, customers and launch dates are rare facts, often changed last month, and sometimes never published. Research on long-tail knowledge shows larger models do not reliably learn rare facts, so better models alone do not close this gap.

Headline hallucination figures have improved on some benchmarks. Vectara's next-generation leaderboard (19 November 2025) reports that "Gemini-2.5-flash-lite leads with 3.3% hallucination rate" on summarising more than 7,700 articles. The same post shows why the trend is not clean: the just-released Gemini-3-pro scored 13.6%, and Claude Sonnet 4.5, GPT-5, Grok-4 and DeepSeek-R1 "all have a hallucination rate > 10%" on the harder dataset.

The more important point is what these benchmarks test. Vectara measures faithfulness to a document the model is given. Google DeepMind's FACTS Grounding (17 December 2024) does the same with 1,719 examples, each with a document of up to 32,000 tokens, judged by three models from different families. Both are about whether a model sticks to supplied text. Neither asks whether a model knows a fact about your company that it was never given.

On that second question, the research is unfavourable. Mallen et al. (arXiv, December 2022) built PopQA, 14,000 questions across entities of different popularity, tested 10 models and found that "LMs struggle with less popular factual knowledge" and that "scaling fails to appreciably improve memorization of factual knowledge in the long tail." They also found that "retrieval-augmented LMs largely outperform orders of magnitude larger LMs". Kalai, Nachum, Vempala and Zhang (arXiv, 4 September 2025) give the statistical reason: "if 20% of birthday facts appear exactly once in the pretraining data, then one expects base models to hallucinate on at least 20% of birthday facts." They also argue that most evaluations reward guessing, because "guessing when unsure maximizes the expected score under a binary 0-1 scheme."

A mid-sized B2B company's facts sit at the far end of that tail. Its integration list may appear on one page. Its pricing changed in September. Its newest customer was announced in one post. Some facts, such as a feature that is in beta for three accounts, are not public at all. No model can know these from training, and the Kalai paper suggests a model under default incentives will guess rather than say so.

The practical conclusion is that the fix is supplying facts and checking against them, not waiting for a better model. That is what retrieval does on the input side, and what a claim check does on the output side.

What does a claim-level check look like technically?

A claim-level check has five parts, which we call the Five-Gate Claim Check: a saved record of company facts, extraction of each checkable claim from the draft, matching every claim to a saved fact with unsupported ones flagged, a named person who approves, and a stored audit record, optionally signed with Content Credentials.

The research basis for the middle gates is solid. FActScore (Min et al., May 2023) "computes the percentage of atomic facts supported by a reliable knowledge source", and its automated estimator has "less than a 2% error rate" against human scoring. Google DeepMind's SAFE (Wei et al., March 2024) uses a model "to break down a long-form response into a set of individual facts" and checks each against search. On about 16,000 facts it "agrees with crowdsourced human annotators 72% of the time", won 76% of a random 100 disagreements, and is "more than 20 times cheaper than human annotators". Both methods check against a general source such as Wikipedia or Google Search. For company claims, the source has to be the company's own record, because that is the only place the true answer exists.

Gate 1: the saved facts record

Everything else depends on this. The record is not a brand book. It is a list of statements a draft is allowed to make: "Integrates with Xero and QuickBooks", "Approval routing for teams of 20 to 300 people", "Customer we may name: none yet". Each figure carries its source and the date it was last confirmed. A record that says what the company does not claim (no SSO yet, no named customers) is as useful as one that says what it does, because it lets the check catch the trap case.

Gate 2: claim extraction

Extraction turns prose into a list. A sentence like "Our AI cuts invoice approval time by 40% for finance teams at companies like Brightwave" holds four claims: the product uses AI, a 40% figure, a target buyer, and a named customer. Rules catch most numbers and proper nouns reliably. Feature claims usually need a model, which is how FActScore and SAFE work, and that model can itself err, so its output is a list of candidates, not a verdict.

Gate 3: matching and unsupported-claim flags

Each extracted claim is compared with the record and labelled. Supported claims pass. Contradicted claims, such as a price that differs from the saved price, are errors. Unsupported claims, which the record neither confirms nor denies, are the large middle group, and the right action is to flag them, not delete them, because some are true facts nobody saved yet. Links get a separate mechanical test: does the URL resolve, and does the page contain the quoted text. Rao et al. show this test alone removes most dead citations.

Gate 4: a named approver

A flag is only useful if a person decides what happens to it. The approver should be named on the record, not "the team", because the Air Canada decision makes the point plainly: the company is responsible for what its channels say, whoever or whatever wrote it. Approval also matters for regulation. The EU AI Act's Article 50 obligations have applied since 2 August 2026, according to Goodwin (3 August 2026). For AI-generated text published to inform the public, the AI Act explainer notes the disclosure duty does not apply where the content "has undergone a process of human review or editorial control" and a person "holds editorial responsibility", and that such review "must be substantive and not limited to superficial matters or cursory approval." A logged approval of specific flags is evidence of substantive review; a single "approved" click on a whole draft is weaker. We covered the wider regulatory picture in Fact-Checking AI Content: Why Companies Need It by 2027, so this article stays on the product and data side.

Gate 5: the audit record and Content Credentials

The record stores the draft as generated, each flag, each decision, the approver, the facts version used and the time. For images, video and documents, the C2PA specification (version 2.3, dated 5 January 2026) defines a signed manifest, "the set of information about the provenance of an asset", built from assertions such as a mandatory created or opened action. That lets a published file carry its own history. Its limit for written content is practical: a LinkedIn post or a pasted paragraph loses any manifest the moment it is copied as plain text, so for text the internal audit record is the dependable part, and Content Credentials are a supplement where the asset is a file.

What does one draft look like going through the five gates?

Here is an illustrative draft for a fictional invoice approval tool, the same fictional company Klype uses in its evaluation. Of seven claims extracted, two were supported, one contradicted the saved facts, three were unsupported and one citation could not be resolved. The approver kept two, rewrote one and cut four.

The draft below is invented for illustration. It is not output from the evaluation, and neither Northwind Ledger nor Brightwave exists.

Draft: "Northwind Ledger now syncs natively with Xero, QuickBooks and NetSuite. Finance teams at companies like Brightwave cut approval time by 40%, and with SOC 2 Type II and SSO, it is ready for teams of 20 to 300 people. As a 2025 industry report noted, approval delays cost mid-market firms millions."

The final post is shorter and less impressive. It is also the only version that would not put four unverified statements about the company onto the open web, where Yext and Ahrefs suggest answer engines would find them. Note that the "SOC 2" flag may be a true fact nobody saved; the check's job is to stop it going out unconfirmed, and the approver's job is to update the record if it is true.

What did Klype's own evaluation find? (self-reported dataset)

In Klype's 30 October 2026 context evaluation, the same model, DeepSeek V3.2, wrote 12 briefs for a fictional company with and without saved facts. With them, drafts introduced 0.08 names not in the facts per draft against 22.75, and no unsupported figures against 0.67. Invented claims fell from 1.58 to 0.50, missing Klype's target.

Read this section as self-reported. Klype built the test, the fictional company, the measures and the review score, and Klype sells the product being tested. The method is published so others can check it, and the results are on our partners page.

The setup: for each of 12 fixed briefs, one draft came from the model alone with a one-line instruction and no company context; the other came through Klype's API on a fictional demo workspace (Northwind Ledger, an invented invoice approval tool) holding product facts, positioning, buyer, messaging and a measured voice. Both arms used DeepSeek V3.2. The briefs covered an announcement, an opinion, a how-to, objection handling, a cold email, a launch, hiring, a customer problem, a numbers brief that supplied no numbers, a short X post, and two trap briefs that asserted features the product does not have. The measures are arithmetic over the draft text against the saved record and the brief.

Both trap briefs were flagged by Klype's gap check, which marks briefs that ask for things the saved facts do not support. That check is Klype's own, so this result is self-reported as well.

The limits, stated plainly

• Small sample. Twelve briefs, one run, one fictional company. Generation is not deterministic, so a second run will produce different drafts.
• Klype's own checker. The review score and the trap flags come from Klype's detectors. The other measures are arithmetic, but Klype wrote that arithmetic too.
• The invented-claims target was not met. The target, set after this run, is 0.3 or lower; the result was 0.50.
• Two briefs were worse with Klype. On the how-to and the numbers posts, the Klype drafts had more invented product claims than the model alone. On review, the extra claims were generic lines rather than false features, but they count.
• An unequal comparison by design. The model-alone arm had no facts at all, so the test measures what saved context adds, not how good the model is. The large gap on names is partly that: a model with no facts invents names because it has nothing else.

What the data supports is narrower than a marketing line, and more useful: giving the model a saved facts record removed almost all invented names and every unsupported figure in this sample, and did not remove all invented claims. That is the argument for gates 3 and 4. Context reduces invention; it does not end it, so the draft still needs flags and a person.

What does the data disagree on?

The studies disagree on how often AI is wrong about brands (2.8% of responses in Seer's single-brand study, 9.2% of brand-reviewed facts in Searchable's) and on whether hallucination is falling. Most of the gap comes from different units, samples and tasks, not from one study being wrong.

• Seer 2.8% versus Searchable 9.2%. Seer counts responses about one well-documented agency and treats "no answer" separately (31.2%). Searchable counts individual facts across 1,704 self-selected brands, graded by the brands. A response holds many facts, so a low response error rate and a higher fact error rate can both be true. Searchable's participants also chose to review, which may skew towards brands that suspected problems. Searchable sells AI visibility tooling; Seer is an agency studying itself.
• Falling versus stubborn hallucination. Vectara's best summarisation score is 3.3%, yet Tow found more than 60% wrong on citation tasks and EBU found 20% of news answers had major accuracy issues. These measure different tasks: sticking to a supplied document versus retrieving and attributing information from the open web. Vectara's own newer dataset puts several frontier models above 10%. "Hallucination is falling" holds for some grounded tasks, not as a general claim.
• Is AI search manipulable? Ahrefs concluded the most detailed story wins. Search Engine Journal argued the test brand had no real footprint and 49 of 56 prompts were leading. Both points survive: engines lean on detailed text when little else exists, and established brands with many sources are harder to distort.
• Invented packages. A commonly repeated figure for the share of hallucinated packages could not be found on the paper's public abstract pages, which give at least 5.2% for commercial models and 21.7% for open-source models. We used the figures we could verify.
• Does context fix it? Mallen et al. show retrieval beats scale on rare facts, and Klype's evaluation shows large drops with saved facts. Klype's own data also shows invented claims remained at 0.50 per draft, and two briefs got worse. Context helps a great deal and is not sufficient.

Analysis: what the evidence adds up to

This section is synthesis only: no new figures, just how the sources above fit together. The short version is that company facts are a problem models cannot solve from training, answer engines amplify whatever a company publishes, and the only controllable step is a check before publishing.

The error is structural, not a bug that will be patched. Kalai et al. and Mallen et al. point the same way: facts that appear rarely are learned poorly, and evaluation incentives push models to guess. A company's current facts are rarer than almost anything a benchmark tests. That is why the two kinds of evidence can coexist: general hallucination benchmarks improving, and brand-reviewed false-fact rates still measurable in 2026.

Publishing is the amplifier. Yext's citation data, Ahrefs's experiment and Seer's error list together describe a loop in which a company's own text becomes the engine's answer. Seer's most repeated error was an outdated office, which is a fact that was once true and never corrected. Every invented claim a company publishes is a candidate for the same loop.

The check has to be claim-level, not draft-level. A reviewer reading a whole draft for tone will pass "integrates with NetSuite" if it sounds right. FActScore and SAFE exist because judging a paragraph is unreliable and judging atomic facts is not. The five claim types each need a different comparison (a fact list, an integration list, an approved-customer list, a sourced figure, a resolving link), which is why one general "check facts" instruction to a model is weaker than extraction plus matching.

The person is part of the mechanism. Unsupported is not the same as false. A check that deletes every unsupported claim strips true facts nobody saved; a check that passes them repeats the problem. The Air Canada decision, the FTC rule and the AI Act explainer all place responsibility on the publisher, and the AI Act explainer specifically discounts cursory approval. A named approver deciding on specific flags is both the practical and the defensible design.

What would change this view. A large, independent study of AI-written company drafts, with claims checked against company-confirmed facts, would replace the gap this article fills with brand-answer data and one self-reported test. If such a study showed frontier models declining to state unknown company facts rather than guessing, the case for gate 3 would weaken. Nothing found in this run shows that.

What are the limitations of this analysis?

The evidence on AI answers about brands is mostly from vendors or single brands, the draft-level evidence on company claims is thin, and the one dataset measuring company drafts directly is Klype's own. Several studies concern news, code or academic citations, used here as analogies for company claims.

• Searchable, Ahrefs, Yext and Vectara sell related products; Seer studied only itself; Klype's evaluation is self-reported. None of the brand and answer-engine studies is peer-reviewed.
• Package, citation and news studies measure related failures, not company marketing claims. They show models invent precise names and sources; they do not give a rate for company drafts.
• Model versions change quickly. Figures from 2023 and 2024 describe older models, and are used here for mechanisms, not current rates.
• The worked example is invented for illustration and is not evidence.
• Some figures were read from publishers' summaries of studies (Nieman Lab for the Tow Center, Search Engine Journal for SparkToro), not the original reports.

If you publish for a founder on LinkedIn, the human side of this, writing in a real person's voice with real stories, is what our sister company Wise Mango does. You can also test any draft for unsupported figures and invented stories with Klype's free LinkedIn post checker, and see the writing patterns it flags on the patterns page.

Methodology

Question. What kinds of false company claims appear in AI-written drafts and AI answers, why model improvements do not remove them, and what a claim-level check can catch.

Source selection. We prioritised primary sources: studies with a stated sample (Searchable, Seer Interactive, Yext, Ahrefs, the Tow Center, EBU and BBC, SparkToro), peer-reviewed or preprint papers (USENIX Security, Scientific Reports, arXiv), a tribunal decision, a regulator's rule, a standards specification and Klype's own published evaluation. Listicles and statistics roundups were excluded. Where a study was only available through a publisher's report (Tow Center via Nieman Lab, SparkToro via Search Engine Journal), we cite that report and say so.

Date range. November 2022 (long-tail knowledge research) to 27 August 2026 (Seer). Older papers are used for mechanisms, not current rates. All sources were opened and every figure checked on its page on 5 November 2026. One commonly cited figure on hallucinated software packages was not on the paper's public pages and was left out.

Klype's data. The context evaluation figures are from Klype's run on 30 October 2026, as published on the partners page. They are labelled self-reported throughout, and its limits are listed in section 07.

Analysis. The analysis section contains synthesis only. It introduces no figures that do not appear in a cited source. The worked example is invented for illustration and is labelled as such.

Sources

1. Searchable, AI Brand Misinformation Report, updated 30 Jul 2026
2. Seer Interactive, AI Brand Accuracy Study, 27 Aug 2026
3. Ahrefs (Mateusz Makosiewicz), I Ran an AI Misinformation Experiment, 10 Dec 2025
4. Search Engine Journal (Roger Montti), Ahrefs Tested AI Misinformation, But Proved Something Else, 28 Dec 2025
5. Nieman Lab, on the Tow Center for Digital Journalism study of AI search citations, 10 Mar 2025
6. European Broadcasting Union and BBC, News Integrity in AI Assistants, 22 Oct 2025
7. Search Engine Journal, on SparkToro and Gumshoe AI recommendation research, 30 Jan 2026
8. Yext, AI citations: 86% of sources are brand-managed, 9 Oct 2025
9. Spracklen et al., package hallucinations by code-generating LLMs, USENIX Security 2025
10. Walters and Wilder, Fabrication and errors in the bibliographic citations generated by ChatGPT, Scientific Reports 13, 14045 (2023)
11. Rao, Wong and Callison-Burch, Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents, arXiv 2604.03173, 3 Apr 2026
12. Vectara, Introducing the next generation of Vectara's hallucination leaderboard, 19 Nov 2025
13. Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate, arXiv 2509.04664, 4 Sep 2025
14. Mallen et al., When Not to Trust Language Models, arXiv 2212.10511, Dec 2022
15. Google DeepMind, FACTS Grounding benchmark, 17 Dec 2024
16. Min et al., FActScore, arXiv 2305.14251, May 2023
17. Wei et al. (Google DeepMind), Long-form factuality in large language models (SAFE), arXiv 2403.18802, Mar 2024
18. Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, 14 Feb 2024
19. The Register, Cursor AI support bot invents policy, 18 Apr 2025
20. US Federal Trade Commission, final rule banning fake reviews and testimonials, 14 Aug 2024
21. Goodwin, EU AI Act transparency obligations now in force, 3 Aug 2026
22. artificialintelligenceact.eu, The EU AI Act's transparency rules: a practical guide to Article 50, 14 May 2026
23. C2PA, Content Credentials technical specification, version 2.3, 5 Jan 2026
24. Klype AI, context evaluation, 30 Oct 2026 (self-reported)
25. Klype AI, We generated 500 LinkedIn posts with an AI model and ran them through the checker, 9 Oct 2026
26. Klype AI, We checked 104 B2B SaaS homepages for claims they can't back up, 10 Oct 2026