JustPaste.it

ChatGPT Search: hidden text and a practical safeguard

When ChatGPT mistakes hidden webpage text for current information.

ChatGPT Search can sometimes turn hidden or inactive webpage text into a claim about what is happening now. Even genuine words from an official website can support a wrong story if their visibility and applicability are lost.

This page shares an ordinary user's warning and a practical way to ask for more careful evidence checking.


Practical help: add an evidence safeguard

For an important claim, open the source yourself.
Check that the warning or status applies to the right person, company, product or application.
Check its date. If the claim could affect a serious decision, look for the underlying official record or ask the responsible authority.

You can also add the instruction below to ChatGPT.
It asks ChatGPT to separate visible page content, hidden wording and authoritative records.

In ChatGPT, go to Settings → Personalization → Custom instructions.
Add it to the Custom instructions field for ChatGPT.
Ensure customization is enabled and save your changes.

This is a practical user-side safeguard. It is not a guarantee and it is not an official OpenAI solution.


Complete instruction -- copy the whole block

Web-evidence verification -- experimental safeguard

Added: 2026-10-06 20:51 Europe/London

When researching websites, distinguish human-visible rendered content, hidden or conditional template text, and underlying authoritative records. Do not infer that a status, warning, notice or event is active solely because its wording appears in extracted text, search snippets or HTML source.

For consequential claims, verify the relevant entity, date and applicability against the underlying authoritative record where possible. Inspect the rendered page when visibility matters. State whether the evidence came from text retrieval, source-code inspection or an actual rendered-page/screenshot check. Never claim visual verification unless it was performed. If verification is unavailable, clearly distinguish the unverified indication from an established fact.

Repeated appearances of the same template text are not independent corroboration. When evidence conflicts, pause and reconcile it before treating the disputed claim as a premise for further conclusions. Conversely, absence of a displayed warning does not establish overall safety or compliance. Preserve the exact original wording when translation affects interpretation.

This is an experimental, untested behavioural safeguard, not a software fix or a guarantee of reliable retrieval or interpretation.

The original wording above is preserved, including “untested”. The instruction was subsequently checked in a fresh ordinary ChatGPT Search conversation and produced substantially better evidence handling. That limited check does not establish reliable performance across other websites or situations.


What happened

ChatGPT researched a company on a foreign-language official Business Registry.
Its reports appeared to show serious official deletion warnings.

Direct confirmation from the registry authority showed that the claimed notices had not been published. Investigation then showed that the warning wording existed in the webpage code inside a hidden or inactive element. A normal visitor did not see that warning.

The company and registry are deliberately not identified here.
The lesson concerns how evidence was handled, rather than the company.

The foreign-language factor made this harder to notice.
ChatGPT was researching, translating and interpreting at the same time.
An ordinary user could reasonably find it difficult to check the original wording and its meaning.


How genuine words can become a wrong claim

A webpage can contain wording for several possible situations.
Only some of that wording may be displayed or active at a given moment.

ChatGPT can sometimes retrieve embedded text without retaining enough information about whether it is currently visible or active. It may then treat a possible warning as a current fact.

A citation to an authoritative website does not, by itself, prove that the cited words describe the current situation. The words can be genuine while the conclusion is wrong.

This account describes an observed evidence-handling problem.
It does not establish exactly where in ChatGPT Search the visibility information was lost, or how often this happens.


Why this matters beyond a registry

The same kind of mistake could affect medical warnings, legal or regulatory notices, sanctions, account restrictions, eligibility decisions, recalls, safety notices, application statuses and other conditional webpage states. These are possible areas of concern, not claims that each has been demonstrated in this case.

Ask what evidence was actually checked: extracted text, webpage code, a rendered page or the underlying official record. If ChatGPT cannot verify whether a notice is active, its answer should make that uncertainty clear.

Repeated copies of the same hidden wording are not independent confirmation. Equally, the absence of a visible warning does not establish overall safety or compliance.


Support report and a wider solution

The matter was first reported privately to OpenAI Support as Case 16599236.
When asked where the issue and user-side mitigation could be shared with other users, OpenAI Support later suggested the Community Forum. This does not imply that OpenAI endorsed the mitigation.

A similar verification principle may ultimately be more appropriately handled by OpenAI at platform or server level. Individual users should not each have to install a workaround. No particular engineering implementation is proposed here.


Community discussion

A post about this issue and the safeguard is now available on the OpenAI Community Forum:

ChatGPT Search can mistake hidden webpage text for current information -- a simple safeguard


Settings reference

OpenAI's Custom Instructions help page:
https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions

Version 1.1 -- 10 October 2026


ChatGPT, AI and Technical Notes
Back to the Start page