Seek or iNaturalist: the same non-profit, opposite trade-offs

Seek returns a model's opinion. iNaturalist returns a model's opinion and then, usually within a day, a person's.

An observation reaches research grade when it has a photograph, a date, a location and two or more independent human identifications agreeing at species level. At that point it flows into the Global Biodiversity Information Facility and becomes usable in science — iNaturalist records appear in thousands of published papers, in range extensions and in conservation assessments.

The practical value to an individual is that you can be told you are wrong by somebody who knows. A confidence score cannot tell you that the thing you photographed is a look-alike which does not occur in your county. A person who studies that genus can, and routinely does, often explaining which character decided it.

Seek identifies on the phone. That is unusual and it has three consequences.

Nothing is uploaded to identify. No server-side record of what you pointed a camera at, or where.

It works with no signal. A canyon, abroad without data, a plane. iNaturalist needs a connection both to suggest and to be confirmed.

The model is smaller. That is the cost, and it is why Seek stops at genus where a cloud model would offer a species.

iNaturalist inverts all three: it uploads, it needs signal, and it has both a larger model and human beings behind it.

Seek will refuse to name a species it is not sure of, showing you kingdom, class, order and family as its confidence rises and stopping there. That reads as failure and is a deliberate child-safety and epistemic decision — a confident wrong name is worse than none in a classroom.

iNaturalist has no such threshold, because it does not have to: it can post an uncertain suggestion and let people resolve it. That is a structurally better answer to uncertainty than either guessing or refusing.

When Seek stalls, the fix is usually physical. It is a live camera rather than a photo classifier, so filling more of the frame, holding steadier and moving into better light often resolves several levels deeper within seconds.

Observations are public by default, including the location, and that is not an oversight — a biodiversity record without a location is not a record. But it has consequences worth understanding.

Obscure coordinates deliberately. iNaturalist does it automatically for threatened species, and you can obscure or privatise any observation manually. Do it for anything rare, anything collectable, and anything on your own property.

Orchids, cacti, reptiles and fossils are the categories where published localities cause real harm; poaching from public coordinates is documented, and the community norm is to obscure.

Your garden is a location. A stream of observations from one point is a home address.

The wait is not fixed. Five things move it:

1. Several photographs of one individual, not one. Whole organism, close detail, and the underside or profile.
2. A scale object. Size eliminates candidates faster than anything and is the character most often missing.
3. Accurate location and date, because regional identifiers filter by place.
4. Add a coarse identification yourself. "Beetle" is enough; unidentified observations sit in a much larger queue.
5. Join a local or taxon project, so people who care about that group see it.

Reciprocity matters: the platform runs on volunteer identifiers, and confirming other people's observations in a group you know keeps the queue short for everyone.

Both are built on taxonomies of living things. A rock, a mineral, a fossil, a coin, a tool or any made object has no taxonomic position and neither app will touch it. Fossils are the notable edge case — once alive, now mineral, and covered by neither.

That is the largest practical gap in the free naturalist stack, and it is why our free browser keys cover rocks, minerals, fossils and projectile points from properties you test rather than photograph.

Both apps are backed by the same computer-vision model, which is why they agree so often and is worth understanding.

It is trained on iNaturalist's own research-grade observations — records with a photograph, a date, a location and two or more independent human identifications agreeing at species. That is an unusually clean training set, because every label was checked by people rather than scraped, and it grows every time somebody confirms an observation.

So the loop is closed: identifications confirmed on iNaturalist become training data, which improves the model, which makes Seek better on your next walk. Using iNaturalist is contributing to the thing that makes Seek work.

The difference between the apps is what happens after the model runs. Seek applies a strict confidence threshold on the device and stops. iNaturalist posts the same output as a suggestion and lets people resolve it — which is why one stops at genus and the other reaches species with a name attached to the decision.