The Pandemic of Falsehoods

A virus of false information is spreading through organizations everywhere, carried by the same tools people trust to make them faster. Almost no one is looking for it.

By Ted Souder, Co-Founder & CEO, Quoin

Somewhere right now an analyst is finishing a memo. She is good at her job, she is up against a deadline, and she needs a market size figure. So she asks a general-purpose chatbot. The answer comes back in seconds, confident and plausible. It goes into the memo. The memo goes to the deal team. The deal team's model picks up the number. The number ends up in the investment committee deck, then the board materials, then the dataset the firm uses to benchmark the next ten deals.

Nobody checked it, nobody can say where it came from, and it was wrong.

She did nothing unusual. That is the point. Multiply her by every analyst, associate, consultant, and manager who has quietly folded these tools into daily work, and you have what I have started calling a pandemic of falsehoods. It is spreading through the one place organizations assume is safe: their own documents.

How the virus works

The models are the carrier, and the people who build them are candid about it.

OpenAI's own research paper on why language models hallucinate concluded that hallucinations are not a mysterious glitch. They come from how the models are trained and scored: the standard process rewards a confident guess over an admission of uncertainty, so the models learn to guess. OpenAI says that can be changed. Its own system card for GPT-5.5, released in April 2026, shows how far there is to go: on conversations users had flagged for errors, the new model's responses contained a factual error only 3% less often than its predecessor's.

Third-party measurement agrees. Vectara's hallucination leaderboard, updated in May 2026, scored roughly 100 models on the most constrained task there is: summarize a document using only the facts in it. Rates ranged from roughly 2% to 24%. Not one model scored zero. If a model invents facts when the facts are sitting in front of it, consider what it does with our analyst's market size question.

Roughly 100 models scored. Not one at zero.

General-purpose chatbots are good for drafting, summarizing and coding. They are not a source of record, and their own published numbers say as much.

Nobody is watching the door

A model that makes things up is only dangerous if nobody checks its work before it gets used. In most organizations, nobody does.

A June 2026 survey of 1,250 office professionals at companies with $500 million or more in revenue found that 86% believe their organization has an AI policy, two thirds have used AI at work in ways they believed the policy did not permit, and 31% have shared financial information or confidential company documents with public AI tools. ISACA's 2026 AI Pulse Poll of more than 3,400 digital trust professionals found that 90% believe employees are using AI in their organizations, while only 38% of organizations have a formal, comprehensive AI policy.

I saw the same thing up close recently. I was leading a roundtable with senior executives from more than a dozen U.S. companies, and I asked for a show of hands: how many of your organizations have AI policies and guardrails that are set, agreed upon, and accepted across the whole company? Three hands went up. Every executive in that room has people using these tools every day, and most could not say under what rules.

So the analyst in the opening is not an outlier. She is the norm.

The cases we know about, and the ones we do not

Every so often one of those numbers gets caught, and we get a headline. This year the headlines have come fast.

In April 2026, South Africa's government withdrew its draft national AI policy after the minister confirmed that its reference list contained fictitious sources. In May, EY Canada pulled a report after an investigation by GPTZero found that almost all of its reference URLs were broken or fake and more than half of the cited titles did not correspond to real sources. In June, KPMG pulled a report on agentic AI after a similar investigation found that only five of its 45 citations accurately pointed to real sources. In July, the same researchers, reported by the Financial Times and others, found fabricated sources across four PwC Middle East thought leadership reports, including a citation URL that still carried a "utm_source=chatgpt.com" tag. Each firm said it takes accuracy seriously; EY and KPMG removed the reports, and PwC said it was updating a limited number of citations. These are serious organizations with real quality controls, which is exactly why the story matters. If it can get past them, it is getting past everyone. And a database maintained by legal researcher Damien Charlotin now counts more than 2,000 court decisions worldwide in which a court found or implied that a party relied on hallucinated AI content, with new decisions added almost daily.

Those cases were published, filed, or delivered to a client, so someone outside the organization went looking. For every fabricated citation a court catches, how many fabricated numbers sit undisturbed in an internal memo that nobody will ever audit? The headlines are the visible cases. The pandemic is everything underneath them.

Years from now, someone goes looking

This is why I use the word pandemic rather than problem. A virus spreads silently before anyone notices symptoms. The false data points entering organizations today are not causing visible harm yet. They are being absorbed into the foundation. Our analyst's market size becomes the baseline in a 2027 model, then the reference figure in a 2028 board deck, then "what we have always understood" by 2029. It gets copied forward and treated as the source of truth precisely because it came from inside the firm.

Then something breaks. A deal underperforms, or a regulator asks a question, or an LP asks to see the underlying work. Someone goes looking for where the number came from and discovers there is no source. The trail ends at a chat window that was closed two years earlier. By then you cannot trace it back, so you cannot tell how far it spread.

The regulators have already said where the responsibility sits. FINRA's Regulatory Notice 24-09 states that its rules and the securities laws continue to apply when member firms use generative AI, and the SEC's examination priorities for fiscal 2026 say examiners will assess whether firms have adequate policies and procedures to monitor or supervise their use of AI. Do not expect "the model said so" to work as a defense.

This is a wake up call, not a warning to slow down

Some people will read all of the above and conclude that the answer is to use less AI. I am not arguing against AI. I am arguing against using it blind.

There is a widely held view in finance, and I suspect in plenty of other industries, that goes roughly like this: we have done well for decades with pen, paper, and Excel, we do not want another tool, and we would prefer that entrepreneurs stop pitching us solutions. I understand the sentiment, and I think it is backwards. The analyst in the opening is already using a new tool. Every firm is. The only question is whether leadership decided how, or whether it just happened. You should want more tools, and you should want smart people pitching you solutions, because some of them will be right.

This connects to what I mean by the Great Compression. Most organizations are still built for the digital transformation playbook of 2015, not for the speed at which AI and agents now operate. So instead of rebuilding, they sprinkle AI on top of existing workflows. Answers show up faster and that feels like progress. It is not. You are just spreading bad numbers faster.

The timing is on your side. The same generation of technology that created this problem is now good enough to fix it, in minutes, at a cost an eight-person advisory firm can carry as easily as a global bank. What most organizations are missing is the decision to use it that way. That is leadership work, not a purchase. Decide what your firm will accept as a source of truth, write it down, and put a team in charge of building the solutions and enforcing the policy. The firms that do this will move faster than everyone else, not slower, because they will be the only ones who can trust what they are moving on.

What the cure looks like

We started Quoin because serious organizations of every kind need to run on reliable data and insights, and because research is still, for most of them, manual labor. Analysts hired for their judgment spend their days collecting and checking data, so when a tool makes the collecting feel instant, they reach for it, guidance or not. The answer is not to take the tool away. It is to give them a better place to do the work.

Quoin is built to be that place. A team opens a Project, runs research and analysis on a company, a market, or a topic, and uploads its own documents alongside. Reports are written from live, retrieved sources rather than from a model's memory. Claims are traced to sources a human can open, and each report ships with an audit file showing how the work was checked. The Project becomes a Corpus of Knowledge that the team can chat with, that persists, and that monitoring updates as new information arrives.

Quoin was not built to replace what already works. The data platforms your team relies on, the dataroom, the expert calls, the models your analysts have refined over years, all of that stays. And this is not a finance-only problem. The cases above involve a government ministry, three of the Big Four, and a court system. Any organization that runs on numbers has the same exposure and the same fix available.

Go back to the analyst one more time. Same memo, same deadline. This time she runs the question inside her Project and the market size comes back with the filing attached, the source that disagreed flagged and set aside, and the whole thing saved where her deal team, her investment committee, and the person who asks about it in 2029 can all find it. Nothing about her day got slower. The number just stopped being a guess.

That is how the pandemic ends: one organization at a time deciding that every number it runs on will have a source. Proof, not probability. We built Quoin to be where that work gets done. You can see it for yourself at quoin.ai.

Sources: OpenAI, "Why Language Models Hallucinate," Kalai, Nachum, Vempala and Zhang, September 2025 (arXiv 2509.04664). OpenAI, GPT-5.5 System Card, April 2026. Vectara Hallucination Leaderboard, last updated May 11, 2026. PagerDuty and Wakefield Research, Shadow AI workplace survey, June 2026, 1,250 office professionals in the U.S., U.K., Australia and Japan. ISACA 2026 AI Pulse Poll, May 2026, 3,400+ respondents. South African Government News Agency, April 26, 2026. GPTZero investigations of EY Canada (May 14, 2026), KPMG (June 12, 2026) and PwC Middle East (July 28, 2026); TechCrunch, June 13, 2026, on KPMG's removal of its report; City AM, July 29, 2026, for PwC Middle East's statement. Damien Charlotin, AI Hallucination Cases database, accessed September 22, 2026. FINRA Regulatory Notice 24-09, June 27, 2024. SEC Division of Examinations, Fiscal Year 2026 Examination Priorities.