ChatGPT Links Were Copied and Pasted Into Official SEC Filings

A review of SEC EDGAR documents found a visible ChatGPT tracking trail, repeated filing-control failures, and one market claim that did not match either of its cited sources. 

Quick Summary 

Someone copied source links from ChatGPT, pasted them into financial documents, and filed those documents with the SEC. The giveaway was utm_source=chatgpt.com, a tracking code attached to links returned by ChatGPT:

• The tracking code’s presence in official filings exposed possible gaps in the review process.
• The trail appeared in 44 official filing documents.
• It happened every month from January through June 2026.
• One BioKey market claim didn’t match either of its footnoted sources. The same filing later gave a different figure. 
• Originality.ai raised a second red flag, this time about the potential prevalence of AI use in these official documents. It flagged 42.9% of scanned documents with a ChatGPT tell as Likely AI, compared to 15% of a random SEC filing sample.

It’s important to note that the links prove that source material passed through ChatGPT. They don’t prove that ChatGPT itself invented the unsupported number. 

The Giveaway Was at the End of the Link

The giveaway wasn’t a supposedly obvious ChatGPT saying or an AI detector score. It was utm_source=chatgpt.com, a tracking code that ChatGPT adds to some of the links it generates for users. When the full link is copied, the code travels with it.

In our investigation, the code appeared in 44 reports on the SEC's EDGAR system. 

However, this doesn’t prove that AI wrote the entire filings, but it does demonstrate that ChatGPT-linked source material was incorporated into official documents and survived the filing review process.

One Fund Report Left the ChatGPT Code in 4 Links

A March 2026 report from Cashmere Fund cited venture-capital research from NVCA, Carta, and Cambridge Associates. Four citations still carried the ChatGPT tracking code.

Of course, the code can’t identify who used ChatGPT or what they asked it. But it does show that the links passed through ChatGPT before the report was filed on EDGAR (with a last modified date of June 2026).

The Same ChatGPT Tell Appeared Month After Month

The Cashmere report wasn’t an isolated incident. Originality.ai’s search found six to ten documents with ChatGPT tells in every completed month from January through June. March had the most, with ten.

The simplest explanation is a repeatable copy-and-paste failure: someone uses ChatGPT to find sources, copies the links into a draft, and the tracking code survives the review process before the document reaches EDGAR. 

Then, One Number Failed the Source Check 

In one of BioKey’s (Cayman’s) SEC filings, it told readers that the global nutraceuticals market would grow from about $112 billion in 2022 to $167 billion by 2030. The sentence points to two footnotes containing URLs from Allied Market Research and Grand View Research.

Although only the Allied Market Research link contained the ChatGPT tracking code, neither of the webpages cited in those URLs supports the numbers in the filing:

• Allied Market Research reports $413.0 billion in 2020 and $650.5 billion by 2030 
• Grand View Research's 2025 report projects $919.1 billion by 2030

The filing even contradicts itself later on. It cites Grand View again and gives the correct figures of $591.1 billion in 2024 and $919.1 billion by 2030.

This is something that can happen with pasting an AI-generated or AI-summarized figure into a draft without checking it first, and could therefore be the result of a hallucination. But without the original ChatGPT conversation, it isn’t possible to prove whether ChatGPT invented the number or a human changed it.

Originality.ai Raises the Broader Concern Over AI Use

Originality.ai tested a visible section of each document containing a ChatGPT tell, as well as a large, randomly selected sample of other SEC documents. 

• 42.9% of documents with a ChatGPT tell were Likely AI (18/42). 
• 15% of 5955 of the random sample documents were Likely AI.

Note: These percentages only include the successfully scanned documents, not the ones that produced errors.

BioKey filed 3 SEC submissions between February and June 2026, with a ChatGPT tell. All three produced an AI flag in separately tested sections of the document, with a Likely AI score. 

That’s the useful role of AI detection here: it can flag a document for further human review. 

One ChatGPT Link Survived 3 Filing Updates 

Coolbit Technologies filed three updates to a registration statement in June 2026:

• The first amendment was on June 4
• A second amendment on June 15
• And a third amendment on June 18

Each one repeated the same ChatGPT-linked Construction Dive citation.

Once the link was added to the document, it was copied forward in subsequent updates. The mistake became part of the permanent public record.

And that’s not all; looking at a broader timeframe from March to June, a ChatGPT UTM code was present in 11 Coolbit Technologies or Coolbit Technologies Ltd (CBAI) filings.

Market Values of Companies with Likely AI SEC Filings With ChatGPT Tells

So, what are the market values of the companies publishing Likely AI SEC filings with ChatGPT tells? Let’s take a look:

• YY Group and Evogene had market values (market cap) of roughly $4.7M and $8.6M
• Telomir Pharmaceuticals was worth roughly $75.7M (market cap)
• Cashmere Fund reported $15.4M in net assets
• V Capital Consulting Group was valued at $168M in a 2025 press release by VCI Global
• According to Investing.com, Coolbit Technologies filed for an IPO in 2026.

Note: The figures above are current to the initial publication of the study. Check the sources for the latest valuations.

Final Thoughts: This Study Shows Concerning Gaps in the Review Process

Using AI for research or drafting isn’t automatically the problem. Publishing its tracking code in an official filing shows that nobody caught an obvious copy-and-paste trail.

SEC filings can pass through company staff, executives, outside lawyers, auditors, consultants, and filing agents. 

The documents don’t show who added the links. They do show that the review process failed to remove them and, in several cases, failed to catch weak or unsupported citations.

That is the accountability question for companies and regulators. 

What is the problem with AI Residue in official SEC Filings?

A tracking code can look like a harmless formatting mistake. However, in an official SEC filing, that code can be evidence that ChatGPT-linked material moved through the drafting and review process without being caught.

Perhaps the more important question is what else came through the same process without leaving such an obvious clue.

Read more about the impact of AI across industries:

• AI Residue
• Likely AI Language Surged Across the Congressional Record
• Politician Reading AI Residue Reveals a Larger AI Pattern
• AI Studies
• AI Detection Accuracy Study

Methodology: How the study was done

Originality.ai searched the SEC's EDGAR Full Text Search for 19 clear signs of AI use in documents dated January 1 through July 10, 2026. Duplicate results were removed using each filing's unique SEC number and document URL, leaving verified documents from January through June (44 of which contained a ChatGPT UTM tell).

For the detector comparison, Originality.ai tested a single visible section of up to roughly 1,400 words from each document using its Lite model and a 50% AI threshold (note that shorter documents used all available text). The random comparison targeted 1,000 SEC documents per month (although totals varied by month). 

Duplicate results, very short documents, and collection interruptions changed the final monthly totals. The published rates use successful scans only: 42 scans from the ChatGPT-tell group and 5,955 scans from the random SEC sample.

PublicWWW was not used for the EDGAR search. The SEC's own search index found the documents, and the original SEC-hosted files were used as evidence. 

Limitations

• A ChatGPT tracking code does not prove that a named employee personally copied text from ChatGPT; another person or system could have passed along the same URL. 
• A matching document is not always issuer-authored prose. EDGAR also contains exhibits, shareholder materials, and third-party submissions. 
• The detector score applies to one section of a document, not the entire filing and not necessarily the words beside the ChatGPT tracking code. 
• AI detection does not test whether a statement is true. The unsupported BioKey number was found by comparing the filing with its cited sources. Our AI fact checker / hallucination checker can help avoid these errors.  
• July covers only July 1 through July 10 and is not a completed month.