We had it neatly done. A blog post with a video loaded directly from IPFS — no YouTube, no CDN, just content pulled from the network. Then we shared it on LinkedIn. And that’s what it came down to.
No video. No preview. A gray area with Bad Gateway.
The contents were not gone. Anyone who opened the link themselves would simply see the video. But the piece of Web 2.0 that was in between — the platform’s preview robot — didn’t get through. That’s a good introduction to what this article is about: IPFS works. It is the place where IPFS touches the old world where it rubs.
Below are the six places where those two worlds collide. For each point, first the gist in plain language; If you want the technology behind it, you open it up.
1. The preview that doesn’t make the cut
Share a regular link on LinkedIn or X and you get a neat card: title, description, image. The platform makes that card itself, by sending a robot past your link to see what is behind it.
That robot is in a hurry. He waits a few seconds and then continues. With a regular web server that is enough — it has the file ready. Not necessarily with IPFS: if the gateway does not already have the content in memory, it must first retrieve it from other nodes. In a video there are hundreds of separate pieces. Before the first piece has arrived, the robot has already left.
The result is an error message in the preview, while a human who opens the same link five seconds later simply sees the video.
▸ The technology behind it
Public gateways usually run a reverse proxy before the IPFS daemon. If that daemon does not respond within the proxy timeout, the proxy returns a 502 Bad Gateway back — that’s the error message you see in the preview, so it comes from the proxy, not from IPFS itself.
In addition, IPFS is content-addressed: the gateway must actively retrieve the blocks from peers if it has not cached them locally. A video consists of many blocks, so that takes time — often more than the timeout of a preview bot allows.
Preview crawlers also sometimes use different requests than a browser (e.g HEAD instead of GET, or different headers). This further increases the chance of failure at public gateways.
In short: not a fault of IPFS, but of the combination of cold public gateway + large file + impatient robot.
2. An address that says nothing about itself
On the regular web, an address says something. brainfusion.nl/handleiding.pdf — you know roughly what you’re getting before you click.
A IPFS address is a long string of characters calculated from the content itself. No name, no extension, no species. This has a major advantage: that series is mathematically linked to the content. If one byte changes, the address is no longer correct. A IPFS address cannot lie about whether the content has been changed.
But it doesn’t say anything either what it is. Is this a manual or something malicious? You can only find that out by opening it. This shifts the trust completely to the sender: you don’t trust the address, you trust the person who sent it to you.
The objection you have to make honestly
Here it is tempting to say that Web 2.0 does better. That’s not right.
A file name is simply a label that the sender chooses. Nothing on the regular web checks whether the content matches the name. A file that kattenfoto.jpg hot can be anything.
So the difference is more subtle than it seems. A IPFS address cannot lie about its content, but says nothing about its nature. A file name can lie about the contents, nor does it say anything about security.
▸ Where that familiar feeling comes from
Classic deception techniques on the regular web: duplicate extensions (kattenfoto.jpg.exe), manipulated MIME types, polyglot files that are simultaneously valid as two formats, and macros hidden in “documents”. The file name offers no protection against this.
The familiar feeling with a regular URL does not come from the file name. It comes from the layer around it: the history of the domain, the validation of the SSL certificate, the reputation with browser alert systems. Exactly that layer is missing with a bare gateway link.
The honest conclusion: web 2.0 is not significantly more reliable about what you see. It only provides more starting points for reputation systems — which themselves can also be circumvented. IPFS lacks those starting points, which makes the trust problem more visible. It exists just as well on the regular web, just camouflaged.
3. Filters that cannot place an address
Spam filters and security systems work with reputation. They know domains, maintain blacklists, see how long something has been around and whether it has caused any trouble before.
A IPFS address does not fit in there. There is no domain to search and no history to consult. Every address is new for every filter, every time. And what a system cannot identify, it treats as suspicious.
The result is erratic behavior: today your link comes through, tomorrow not, without anything having changed in the content.
4. The trail you don’t see
When you share a link, the platform records it — who, what, when. That is known and logical.
What fewer people realize: the preview robot also leaves a trail, but at the gateway. This records the time, which address was requested, and from which infrastructure the request came. Two separate tracks that can be combined.
By itself, that gateway registration doesn’t reveal who the sender is — the IP address belongs to the platform, not you. But it is an additional data source that would not be there with a normal shared link.
▸ What exactly is recorded
The gateway shows: the IP of the crawler infrastructure (not that of the installer), the user agent, the time and the requested address. This in itself cannot be traced back to a person.
The link between user, content and crawler request originates at the platform, not at the gateway. Major platforms also almost certainly perform automatic checks on submitted links — malware, phishing, moderation. A hit against internal lists can generate a signal that is linked to the account data that the platform already has.
So no mystery, but a real extra layer of information on top of what has already been established.
5. Legislation based on one button
Laws surrounding content removal assume one thing: that there is someone who can push a button. A server with an owner, an account with an administrator, an address that can be taken offline.
At IPFS that button is not there. The address refers to the content itself, not to a location where it is located. If you delete it at the source, it will remain accessible as long as one node elsewhere stores it. The protocol simply has no central deletion function.
What can be addressed are the public gateways. That is where the attention is focused. But that shifts the issue rather than solving it: blocked at one gateway does not mean unreachable at the other.
▸ The legal gray area
Regimes such as the DMCA, the European Copyright Directive and the DSA all presuppose a single point of control. This is missing with content addressing: there is no party that owns “the address”.
It is legally unclear who the intermediary is who addresses such legislation. The gateway administrator? The one who keeps the content? Nobody? The debate is similar to that surrounding liability for Tor exit nodes and VPN providers, and is largely unresolved.
For those who run a node themselves, one thing is clear: you are accountable for what you consciously host and store. The gray area starts with content from others that happens to pass through your network participation.
6. “On IPFS” doesn’t mean “forever”
This is the most persistent misunderstanding, and it deserves the most attention.
People hear “decentralized” and think “indestructible.” That’s not what IPFS promises. IPFS guarantees how you can check if something is unchanged — via the address calculated from the content. It doesn’t guarantee that something remains available.
Content only persists as long as at least one node is actively holding it. If that no longer happens, it will disappear as soon as the last node with a temporary copy runs its cleanup. The address then remains valid – there is just nothing left to collect.
Popular content happens to survive longer because many visitors leave behind temporary copies. But chance is not a strategy. Only conscious and long-term retention keeps something structurally alive.
What this means
Six points of friction, and none of them are a reason to ignore IPFS. It’s the cost of something that works fundamentally differently than the infrastructure around it.
What is striking when you put them side by side: almost no problem is in IPFS himself. They’re in the intermediate layer — the gateways, the preview robots, the filters, the legislation designed for a different world. That layer expects an address that points to a place, a name with a history, an owner who is approachable. IPFS does not deliver that, because it consciously does something else.
And the most honest thing you can say about it is this: IPFS makes problems visible that also exist on the regular web, but have been neatly eliminated there. That a file name guarantees nothing, that availability depends on someone’s willingness to continue hosting, that trust ultimately lies with people and not with technology — this applies everywhere. On the regular web there is only a reassuring layer around it.
That’s not an argument against IPFS. It is an argument for knowing what you are getting into.
This article arose from a practical problem with our own blog post How can you use AI?, which loads a video directly from IPFS. Want to set up something self-hosted yourself? Please contact us.