Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 11 min read

Discuss+ Banned, Deleted, Still Found: How AI Reconstructs Your Digital Footprint

You can lose your account, delete your posts, or disappear from a platform. But can artificial intelligence still piece together who you were online? Verdict Sponsor Sentinel IR development here: https://github.com/spons

Discuss+ Banned, Deleted, Still Found: How AI Reconstructs Your Digital Footprint



You can lose your account, delete your posts, or disappear from a platform. But can artificial intelligence still piece together who you were online?
Verdict
Sponsor Sentinel IR development here: https://github.com/sponsors/jakcmencz?o=syg&sc=r
Source (GitLab, Apache-2.0): https://gitlab.com/jakubrojicek11/sentinel-ir
npm: https://www.npmjs.com/package/sentinel-ir
Playground (runs in your browser): https://sentinel-ir-playground-dc30c8.gitlab.io/
MCP Server on npm: npmjs.com
https://glama.ai/mcp/servers/jakcmencz/sentinel-ir-mcp
By "jackymenCZ" (https://dev.to/jackymencz)

Imagine this.

You wake up one morning and discover that your account has been permanently banned.

No posting. No commenting. No access.

Your profile looks like an abandoned apartment after the landlord changes the locks.

You assume your online presence is over.

Then, a few days later, you open an AI assistant and ask:

"Tell me what you can find about my online activity."

And suddenly, the assistant starts describing your old posts, mentioning discussions you participated in, identifying related accounts, and pointing to URLs you thought were dead.

Wait a minute.

Didn't we just get banned?

How does an AI know about something that supposedly no longer exists?

Did somebody save a copy?

Is it remembering something from its training?

Or is it simply doing something much more interesting?

Welcome to the digital afterlife.

  1. A ban is not a universal delete button

The first misunderstanding is that people often treat a platform ban and content deletion as the same operation.

They are not.

A ban usually changes what an account is allowed to do. Depending on the platform, it may also affect whether its profile or posts remain publicly accessible.

Deleting an individual post is another operation.

Deleting an entire account is yet another.

And none of these actions automatically removes independent copies already collected elsewhere.

Think about a hotel again.

The receptionist can deactivate your room key.

That does not automatically erase:

  • The security camera footage.
  • The invoice in accounting.
  • The photograph someone took in the lobby.
  • The copy of your reservation sent to another company.

The hotel controls its own database.

It does not control every copy of information that has already travelled outside its walls.

The internet works in a similar way.

  1. Search engines remember yesterday's internet

One of the simplest explanations is search index persistence.

Search engines do not necessarily retrieve every page directly from its original server every time someone searches.

They maintain indexes built through crawling and processing websites.

This creates an interesting situation.

A page can be removed from its original location while information about it remains in search results.

Google explicitly documents this problem. Its "Refresh Outdated Content documentation" (https://support.google.com/websearch/answer/6349986) explains that deleted content may continue appearing in Search with an outdated description or cached information.

Microsoft documents similar issues for Bing and Copilot.

"Microsoft: How to Remove Content from Bing and Copilot" (https://www.bing.com/webmasters/help/content-removal-cb6c294d)

In other words, the original page and its search index do not necessarily update simultaneously.

Imagine a photographer taking a picture of a shop window.

The shop closes tomorrow.

The photograph does not disappear from everyone's phone.

The search engine is holding the photograph, not necessarily the current shop.

And an AI-powered search experience can use information retrieved from those search systems.

  1. The Internet Archive: The internet has its own memory

Now things become even more interesting.

There are services specifically designed to preserve historical versions of websites.

One of the best-known examples is the "Internet Archive's Wayback Machine" (https://web.archive.org/).

Its purpose is to preserve snapshots of webpages.

According to the Internet Archive, pages saved through its service can continue to exist even after the original webpage has changed or been removed.

That means a deleted page may still have a historical copy.

But there is another important project.

"Common Crawl" (https://commoncrawl.org/).

Common Crawl continuously collects snapshots of publicly accessible web content and makes its datasets available to researchers and developers.

Its archive spans more than a decade, with crawls typically released approximately monthly.

Why should anyone outside machine learning care?

Because Common Crawl has become an important source of training data for language models.

The Mozilla Foundation investigated this in its 2024 research report:

"Training Data for the Price of a Sandwich: Common Crawl's Impact on Generative AI" (https://www.mozillafoundation.org/en/research/library/generative-ai-training-data/)

The researchers examined 47 text-generating language models published between 2019 and October 2023.

At least 30 of those models, approximately 64%, used filtered versions of Common Crawl in their pre-training.

This is an important finding.

It shows that historical web crawls are not just interesting digital archaeology.

They have helped build the infrastructure behind modern generative AI.

It does not mean that every webpage collected by Common Crawl ends up inside every AI model.

But it does show why deleting a page from its original server cannot be treated as proof that the information has never been collected elsewhere.

  1. Can AI actually remember deleted content?

Here we need to distinguish two things.

Retrieval and training are not the same process.

When an AI system searches the web, it can retrieve information from external sources and use it to construct an answer.

This is often implemented through Retrieval-Augmented Generation, or RAG.

The model does not necessarily need to have learned the original post during training.

It can find a surviving reference and summarize it.

Google even documents how its AI search features can use multiple related searches, a technique called query fan-out, to gather information across different topics and sources.

"Google: AI Features and Your Website" (https://developers.google.com/search/docs/appearance/ai-features)

But what about the model itself?

Can it remember something that was included in its training data?

Yes, sometimes.

A particularly interesting academic paper was published at USENIX Security 2021.

"Extracting Training Data from Large Language Models" (https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting)

The researchers demonstrated that they could extract hundreds of verbatim text sequences from GPT-2's training data.

The recovered information included names, email addresses, code and other text.

This was not merely a theoretical possibility.

It was a demonstrated security problem.

However, there is an important limitation.

This does not prove that every current AI assistant can retrieve arbitrary deleted Reddit posts from its training data.

It proves that memorization and extraction are possible under certain conditions.

A trained model is not a searchable folder containing every webpage it has ever encountered.

It learns patterns, relationships and representations. Some information can nevertheless be memorized.

This is why the distinction matters.

An AI saying something does not automatically tell us whether it retrieved a page, remembered a fragment or simply generated a plausible answer.

  1. Reddit, Google and OpenAI: The real data partnerships

Now let's look at something that makes this discussion particularly relevant in 2026.

Reddit has publicly announced partnerships with major AI companies.

On February 22, 2024, Reddit announced an expanded partnership with Google.

The announcement described programmatic access through Reddit's Data API to public posts and comments, with uses including improving products and supporting more efficient model training.

"Reddit: Expanding our Partnership with Google" (https://redditinc.com/news/reddit-and-google-expand-partnership)

On May 16, 2024, Reddit announced a partnership with OpenAI.

OpenAI would access Reddit's Data API, providing real-time, structured content to improve how its AI products understand and present Reddit discussions.

"Reddit: Reddit and OpenAI Build Partnership" (https://redditinc.com/news/reddit-and-oai-partner)

So yes, there are real data pipelines connecting social platforms and AI companies.

But we need to be careful here.

It would be incorrect to conclude that every Reddit post automatically enters an AI training pipeline before moderators can remove it.

Reddit's current deletion policy says that third parties, including data licensees, must stop displaying or using deleted content when Reddit or the user deletes it.

Its Data API documentation also requires recipients to remove associated deleted content, recommending routine deletion of stored data within 48 hours.

Sources:

This is a very important distinction.

A licensed data stream does not mean unlimited permission to preserve deleted content forever.

However, it also does not automatically erase independent archives, historical crawls, screenshots, quotations or previously distributed copies.

And removing something from a dataset is not necessarily the same as removing its influence from a model that has already been trained.

Researchers have an entire field devoted to this problem.

It is called machine unlearning.

A 2025 open-access review in Artificial Intelligence Review examines the challenges of making language models forget information they have learned.

"Digital Forgetting in Large Language Models: A Survey of Unlearning Methods" (https://link.springer.com/article/10.1007/s10462-024-11078-6)

Apparently, teaching AI to remember is only half the problem.

Teaching it to forget is another research challenge.

  1. The really interesting part: Reconstructing your identity

Now we get to the phenomenon that fascinates me most.

Suppose one account disappears.

But the same person has also:

  • Published articles under the same username.
  • Linked to a GitHub repository.
  • Participated in technical discussions.
  • Used similar profile descriptions.
  • Shared links between different platforms.
  • Left a recognizable writing style.

None of these individual traces necessarily tells the whole story.

But together, they can provide enough information to construct a much larger picture.

This is closely related to a research field called User Identity Linkage.

A 2024 review published in IEEE Access examines techniques for connecting identities across social networks.

The researchers describe methods that use profile information, content, behavioral patterns and network relationships.

"IEEE Access: User Identity Linkage on Social Networks" (https://doi.org/10.1109/ACCESS.2024.3500374)

And this is where AI becomes particularly interesting.

It can help connect information from different sources and produce a coherent description of an online identity.

You do not necessarily need your original account to remain accessible.

The information may already be distributed across other publicly accessible sources.

Of course, there is a catch.

Two people can have similar usernames.

Writing styles can overlap.

Old information can be incorrect.

A model can connect unrelated accounts and confidently tell a convincing story.

A reconstructed profile is therefore not automatically a verified identity.

AI can assemble a picture.

That does not mean every piece belongs in the picture.

  1. So, can a platform really make you disappear?

Not completely, at least not through a single ban or delete operation.

But we should avoid the opposite exaggeration too.

AI cannot access every deleted post, every private account or every historical version of every website.

Some content disappears permanently from all sources that a particular system can access.

Some content survives in archives.

Some remains in search indexes.

Some may have been included in training datasets.

Some may be reconstructed from unrelated public traces.

And sometimes the AI simply gets it wrong.

There is no universal internet memory that every AI assistant can access.

There are multiple independent systems, each with its own data, update schedule, retention policies and limitations.

That is the real story.

  1. A small security lesson

There is a concept in software security that I keep coming back to: Time-of-Check to Time-of-Use, commonly known as TOCTOU.

Something is checked at one moment.

Something else happens before it is used.

The information that was valid during the check may no longer represent the current state.

The same thinking applies to digital information.

An old search result is not necessarily the current webpage.

An archived post is not necessarily the current opinion of its author.

A model-generated profile is not necessarily an accurate representation of a person.

This is why provenance matters.

We should ask:

  • Where did this information come from?
  • When was it collected?
  • Is the original source still available?
  • Is this an archived copy or a live result?
  • Is the AI reporting evidence or making an inference?

Even a cryptographic hash has a specific role here. It can help establish whether two captured versions contain identical bytes. It cannot, by itself, prove that the content was current or true when somebody used it.

That is a small distinction with big consequences.

Final thoughts: The account disappeared. The evidence didn't necessarily disappear with it.

Perhaps the most interesting thing about modern AI is not that it knows everything.

It is that it can combine fragments that were never intended to form one complete picture.

A platform can close an account.

A search engine can retain an outdated description.

An archive can preserve an older version.

A dataset can contain a historical snapshot.

Another website can quote the original post.

And an AI system can bring some of those fragments together.

The result can look surprisingly complete.

But remember:

A reconstructed story is not the same as an authenticated history.

The next time an AI tells you it found something that supposedly no longer exists, ask one simple question:

"Where exactly did you find it, and when was that information collected?"

Because sometimes the ghost really is in the archive.

And sometimes the ghost is just a language model wearing a very convincing detective costume.

Sources and further reading

  1. "Google Search: Refresh outdated content" (https://support.google.com/websearch/answer/6349986)
  2. "Microsoft Bing: Content removal and Copilot" (https://www.bing.com/webmasters/help/content-removal-cb6c294d)
  3. "Internet Archive: Saving pages in the Wayback Machine" (https://archivesupport.zendesk.com/hc/en-us/articles/360001513491-Save-Pages-in-the-Wayback-Machine)
  4. "Common Crawl: About and archive" (https://www.commoncrawl.org/about)
  5. "Mozilla Foundation: Training Data for the Price of a Sandwich" (https://www.mozillafoundation.org/en/research/library/generative-ai-training-data/)
  6. "Pew Research Center: When Online Content Disappears" (https://www.pewresearch.org/data-labs/2024/05/17/when-online-content-disappears/)
  7. "Reddit and Google: Expanded partnership" (https://redditinc.com/news/reddit-and-google-expand-partnership)
  8. "Reddit and OpenAI: Partnership announcement" (https://redditinc.com/news/reddit-and-oai-partner)
  9. "USENIX Security: Extracting Training Data from Large Language Models" (https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting)
  10. "IEEE Access: User Identity Linkage on Social Networks" (https://doi.org/10.1109/ACCESS.2024.3500374)
  11. "Artificial Intelligence Review: Digital Forgetting in Large Language Models" (https://link.springer.com/article/10.1007/s10462-024-11078-6)

Written from the perspective of an independent technology researcher interested in AI systems, security, and the difference between what a system claims to know and what can actually be verified.

More experiments and projects: "jackymenCZ on DEV.to" (https://dev.to/jackymencz).

Independent Engineering Projects

Built as independent engineering experiments across AI, software analysis, security, and autonomous systems.

Evaluated by reproducible evidence, not by the number of layers.

Explore the projects:

Sentinel Ecosystem: AutoDoc-Sentinel, Sentinel-IR, Sentinel-IR-MCP and related tools.
Agentix Lite: Independent security and system-observation project.
Find the projects: GitLab repositories | NPM packages

Support the work: ⭐ Star the repositories on GitLab.

Connect: Follow me here on DEV Community or explore the Playground.

Each project is distributed under its own license. Check the corresponding repository or NPM package for licensing terms.

Copyright Β© 2026 jackymenCZ. All rights reserved. Written for the community.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.