AI search is genuinely useful. It is also an upload, and most editors never see the second half of that sentence.
The setup: by 2026, "search your library by what is in the footage" is the marquee feature across Frame.io, Shade, iconik, and LucidLink. Type "drone shot of the coastline at sunset" and the right clips appear. To make that work, software has to look at your footage. The question this piece answers is where that looking happens, and who keeps a copy.
What "AI search" actually does: it uploads your footage #
Here is the mechanic nobody puts on the pricing page. To index footage by its content, the footage, or a derived proxy and audio track, has to be processed somewhere with a GPU. In the major cloud products, that somewhere is the vendor's infrastructure, not your machine.
- Frame.io, Shade, and iconik process at ingest, in the cloud. There is no on-device, local-only tagging option in any of the four products here. The footage leaves the building to get indexed.
- iconik is the most conservative about it: it processes proxy files and an extracted audio track, and states that full-resolution originals are not sent to AI providers. Better, but the proxy of your unreleased cut is still a copy of your unreleased cut.
- LucidLink is the interesting one. Its own filespace is zero-knowledge encrypted, meaning not even LucidLink holds your keys. But LucidLink does not tag footage itself. AI search comes from bolting on a third party, Moments Lab, whose engine necessarily sees decrypted frames streamed from the filespace. The privacy guarantee ends exactly where the AI begins.
So the first thing to internalize: enabling AI search is not a setting, it is a data-export decision.
Transcription, objects, and the biometric question #
Not all AI analysis carries the same weight. Three layers, in rising order of sensitivity:
Transcription is everywhere. iconik runs speech-to-text automatically at ingest (powered by Rev AI), Frame.io does transcription with speaker identification, and the others transcribe too. This is the least fraught layer, though your dialogue and any off-hand comments on the audio track are now text in someone's database.
Semantic and object detection is the search-by-meaning layer: scenes, objects, concepts. Frame.io's semantic search analyzes visual content and is gated to Team and Enterprise accounts. This is the feature people actually want.
Facial recognition and people clustering is the escalation, and it is worth being precise about who does it. Shade offers it directly: label faces, cluster people, search by name. iconik offers it with its own model. Frame.io does not currently offer it and describes the feature as "under consideration." LucidLink gets it through Moments Lab. Clustering identifiable people by face is a different category of data than "this clip has a boat in it." If you shoot real people, especially under release or NDA, that is a biometric index of them sitting on a vendor's servers. To the user's own question of whether anyone is getting real value out of facial categorization: in a large stock or sports library, yes, name-search is a genuine time-saver. On a confidential client shoot, the value rarely justifies building that index off-premises.
"We don't train on your data" answers the wrong question #
Every serious vendor now says this, and they appear to mean it. Frame.io states it does not use customer data to train or augment its AI model, in line with Adobe's ethics posture. Shade says your data is "100% yours and is never used for training." iconik says customer content is not used to train external models. Good. But "we do not train on it" is the answer to a narrower question than the one that matters.
Not-training is not not-processing. To index your footage, it still gets copied, decrypted for analysis, and frequently handed to sub-processors. iconik, for instance, has used Google Video Intelligence and Amazon Rekognition for analysis and Rev AI for transcription. Each of those is another company in the chain that touches the data, even if none of them train on it.
And the privacy-preserving choice is often the one you have to go find. Frame.io processes assets for semantic search by default and requires contacting support to opt out. Opt-out, not opt-in. If you do nothing, the embeddings get made.
Who ends up holding a copy #
For NDA'd or embargoed work, reframe the whole thing. Forget training. Count parties. Every cloud-AI path adds at least the vendor, and usually its sub-processors, to the set of companies that have held a decrypted copy of your footage.
Local-only index (self-hosted NAS) 0 third parties
Your MacYour NAS
Cloud MAM (Frame.io / Shade / iconik) 2+ third parties
YouVendorGoogle / AWSRev AI
LucidLink + Moments Lab 2 third parties
YouLucidLinkMoments Lab
TPN and SOC 2 certifications, which Shade, iconik, and LucidLink carry, genuinely reduce the risk that any one of those parties is careless. They do not reduce the count. And the trend is widening it: LucidLink launched an MCP server in public beta on June 25, 2026, exposing its filespace to autonomous AI agents like Claude and the OpenAI Agents SDK. That is powerful, and it is one more class of thing that can now reach decrypted files.
The feature and privacy map #
| Tool | Transcription | Semantic search | Facial recognition | Where indexing runs | AI cost model |
|---|---|---|---|---|---|
| Frame.io | Yes | Team / Enterprise only | No ("under consideration") | Cloud (Adobe) | Tiered (plan-gated) |
| Shade | Yes | Yes | Yes | Cloud | Bundled in seat (~$20/seat/mo annual) |
| iconik | Yes (Rev AI) | Yes | Yes (own model) | Cloud (proxy + audio) | Metered ($1/hr transcription, $1 credits) |
| LucidLink | Via Moments Lab | Via Moments Lab | Via Moments Lab | Third party (Moments Lab) | Separate Moments Lab contract |
| JuiceMount | No (today) | No (filename + metadata) | No | Local (on the Mac) | None ($0/seat) |
A different boundary, with an honest tradeoff #
Here is the one place JuiceMount belongs in this story, and I will keep it short. JuiceMount draws the confidentiality boundary at your own hardware: the search index is a local database on the Mac, content is never uploaded for tagging, and there is no third-party AI processor in the path. On the diagram above, it is the top row.
The honest cost of that choice, stated plainly: JuiceMount does not do semantic search, face clustering, or transcription today. You get fast filename and metadata search with zero parties added, and you give up search-by-content. That is a real tradeoff, not a free win. For a stock or sports library where name-search saves hours, cloud AI may well be worth the added surface. For NDA'd, embargoed, or pre-release footage, keeping the index on your own machine is the safer default, and it is the bet we made.
Whichever way you lean, the move is the same: decide on purpose. Read where the indexing runs before you turn it on, and treat "enable AI search" as the data-export decision it actually is.
Sources, all checked June 2026
- Frame.io: V4 Knowledge Center (media intelligence search, semantic embeddings + opt-out, no-training statement, facial recognition "under consideration"); Frame.io Insider (Dec 2025 search guide and plan tiers; V4 NAB 2025 release).
- Shade: shade.inc (AI search, facial recognition, "100% yours, never used for training," certifications); pricing via shade.inc and third-party reviews (confirm current pricing on shade.inc/pricing).
- iconik: iconik.io (AI metadata, proxy + audio processing, no-training statement, FAQs); Jan 2025 pricing post ($1/hr transcription, credits).
- LucidLink: lucidlink.com security and DPA; the Moments Lab integration blog; SiliconANGLE on the June 25, 2026 MCP server beta.
- Moments Lab: Tech.eu ($24M raise to scale AI video indexing).