When setting up AI Tags, you will see two options:

Restrict to Existing
Allow New

This setting directly determines how your tag system grows over time: staying consistent with what you have, or letting AI keep inventing new tags.

With a small library, the difference barely matters. But once you hit thousands of images, this choice decides whether your tags stay usable or turn into a mess you dread opening.

Nobody runs into trouble on day one. The problems creep in quietly:

The same concept shows up in three different wordings
Tag count keeps climbing, but only a handful get reused
You cannot remember which synonym to search for
The tag manager fills with one-off tags you will never touch again

The real question is not whether AI can tag images. It is whether your tag system is ready for what AI will do to it.

Understanding the Difference Between These Two Modes

Restrict to Existing

AI only picks from tags already in your library — no new ones get created.

Tag count stays flat. Naming stays consistent. Maintenance stays low.

The catch: your existing tags need decent coverage. If your library only has 30 vague tags, AI does not have much to work with, and results will be underwhelming.

"Restrict to Existing" trades flexibility for stability — but only works if you have already built something worth restricting to.

Allow New

AI can create new tags based on what it sees. New tags land in the AI Generated group automatically.

This is useful early on, when your tag system has gaps and you want AI to help fill them in. It works well for exploration.

The cost is straightforward: the more freedom AI has, the faster your tag count balloons. And the ballooning happens slowly enough that you will not notice until it is already a problem.

Why Does "Allow New" Lead to Messy Tags?

When AI looks at an image, it tries to describe everything it sees. Flowers, butterfly, sunlight, meadow — each gets its own tag. For one image, fine.

Now multiply that across 5,000 images with 5–10 tags each.

Even though AI tries to reuse existing tags, it will inevitably produce new phrasings, new levels of detail, and new naming patterns. After a few thousand images, you end up with:

The same concept in multiple wordings

sunset, dusk, golden hour, sundown
Form inconsistencies

flower, flowers, floral
Mixed granularity

animal, dog, shiba inu
Synonym pile-ups

minimal, minimalist, minimalistic, clean

None of these tags are wrong individually. The problem is they keep accumulating, and merging them after the fact is tedious work that only gets worse the longer you wait.

The Problem Is Not AI — It Is Missing Structure

AI is not randomly slapping on tags. It is doing exactly what you asked: describe image content as thoroughly as possible.

Without naming conventions, clear tag groups, or boundaries for what belongs and what does not, those descriptions pile up into an unmanageable heap. The issue is not AI Tags — it is letting AI add tags freely with no structure to guide it.

Why Do So Many People End Up Here?

Because they skip the setup. The typical path looks like this:

Enable AI Tags
Turn on "Allow New"
Process the entire library

Quick and easy. But after a few weeks, the tag count has exploded, naming is inconsistent, synonyms are everywhere, and cleaning it up will take longer than setting it up properly would have.

The point of this article is not "never use Allow New." It is: organize your tag structure before you let AI loose on your library.

When to Use Each Mode

Restrict to Existing works when:

You already have a mature tag system with good coverage
Naming conventions are stable
Consistency matters more than discovering new tags

If your existing tags are sparse or poorly organized, this mode will just give AI a bad menu to order from. Build the foundation first.

Allow New works when:

You are starting a new library from scratch
Your tag system is still evolving and has clear gaps
You are willing to do cleanup afterward

More freedom means more cleanup later. That is the trade-off, and it is worth accepting only if you go in with your eyes open.

The Real First Step: Organize Your Tags

Before picking a mode, ask a different question: is your tag system ready?

If the tags themselves have no structure, neither mode will save you. If they do, both modes work fine.

1. Set up tag groups along clear dimensions

Color, style, mood, subject, composition, use case, material, scene — whatever makes sense for your collection. The goal is not to have many groups, but to make sure each group describes one thing.

2. Write descriptions for each group

Help AI understand what this group covers, where the boundaries are, and what should not be in it. The system also provides up to 10 sample tags from each group as reference.

Tip: You can use Eagle MCP with an AI Agent to analyze your existing tags and suggest a cleaner group structure.

3. Let structure do the heavy lifting

Once groups, naming rules, and sample tags are in place, "Restrict to Existing" becomes more accurate and "Allow New" becomes more disciplined. The mode matters less than the structure behind it.

Use Custom Instructions to Add Guardrails

Beyond tag structure, Custom Instructions give you another layer of control:

Standardize language

Use lowercase English for all tags

Limit quantity

Output no more than 5 tags per image, keeping only the most essential content

Narrow the focus

Focus only on color, composition, and style — do not describe specific objects

Avoid over-specificity

Do not generate overly rare tags that apply to only a single image

These constraints keep AI closer to your organizational logic instead of just dumping everything it sees.

For more examples and writing tips, refer to the AI Action Best Practices Guide.

If Your Goal Is Just Finding Images, Consider AI Description Instead

If you mainly want images to be searchable — not taxonomically organized — AI Description might be the better tool.

Descriptions give you plenty of searchable keywords, handle mood and complex scenes better than tags do, and do not pollute your global tag system. If a description is imperfect, the damage stays on that one image.

For users who do not want to manage tags long-term, this is often less work than letting AI create tags freely and cleaning up afterward.

For tips on writing description-oriented Custom Instructions, see the description section in the AI Action Best Practices Guide. This also connects to the thinking in Rethinking Image Organization: often, "being able to find it" matters more than "sorting it precisely."

Already Messy? You Can Still Fix It

If you already ran "Allow New" across a large library and your tags are full of duplicates and synonyms — it is fixable. Eagle MCP with an AI Agent can help merge synonyms, standardize naming, reorganize groups, and clean out noise tags.

That said, the longer you wait, the more work it becomes. Better to give AI structure and constraints from the start than to untangle things later.

Quick Decision Guide

Is your tag system already mature?

Yes, with good structure → Use Restrict to Existing to maintain consistency
Not yet, but willing to build it first → Organize groups and rules, then decide
Not yet, and no plans to manage tags → Use AI Description instead
Exploring, want AI to help expand → Use Allow New, but budget for cleanup

One Last Thing

"Allow New" is not dangerous. It just requires more awareness of what comes after.

If your library is growing and your tags are still taking shape, it can help you build coverage fast. But if you have a large collection with no structure and no cleanup plan, it will make things harder to manage over time.

Before you flip the switch, ask yourself: after these tags are added, will I be able to maintain them?

If not, get your tag structure in order first.

Why AI Tagging Can Make Your Tags Messier Over Time