How GravityAgent learns from your corrections, without training anything

Every spam filter gets a call wrong now and then. What separates a useful one from an annoying one is what happens after you fix it.

Most filters do nothing. You mark the entry Not Spam, the entry is restored, and the next submission that looks like it gets the same wrong treatment. Some let you write rules or instructions, which works, but you are the one doing the learning.

GravityAgent does something different, and because "AI that learns" is a phrase that usually hides more than it explains, this post says exactly what it does, what it does not do, and where your data is at each step.

Step 1: you correct it with the buttons you already use

There is no GravityAgent correction screen. When a verdict is wrong, you use Gravity Forms' own Spam or Not Spam action, on the entry row or the entry itself, the way you always have. That is the whole correction. You see one line in return:

Correction noted — it informs future decisions for this site.

and on the entry's decision panel:

You marked this not spam. Your correction informs future decisions for this site.

Step 2: the correction is recorded, on your site

The plugin records the correction in a table in your own WordPress database: which entry, which form, and the label you chose. It tells our service three things about the correction: the decision id, the label you corrected to, and the direction of the change (spam to not spam, or the reverse). No submission content is sent when you correct.

The submission's text is not copied anywhere. It stays in the entry in WordPress, which is the only copy there is.

Step 3: the next submission on that form is scored with your corrections as examples

When a new submission arrives on that form, the plugin picks the relevant past corrections and sends them along with the new submission, as examples: here is a submission, here is what the site owner decided it was. The model sees your decisions next to the new entry and decides accordingly.

Two details matter here.

The examples are read live. An example carries the same field content the original entry holds, read from your WordPress database at the moment the new submission is scored, minus any field you have excluded. If you delete the entry, the example is gone from every future decision. If you exclude a field on the form, that field is stripped from the examples too.

Nothing is trained. No model is fine-tuned on your data. No weights change. Your corrections are shown to the model as context, for your site only, on that form's submissions. This is why a correction takes effect on the very next submission rather than after some training run, and it is why your corrections never influence anyone else's site.

Step 4: the decision panel shows you what it used

Open any scored entry and the decision panel lists what informed this decision: the site profile GravityAgent built from your public pages, the notes you wrote about your business, and the corrections it considered, each one linked back to the entry you corrected. If an example was not used, the panel says so. You never have to wonder whether a correction "took".

What we keep, and what we do not

Your corrections are yours. They live on your site, they are read live when needed, and they are never copied into a database we hold. Deactivate the site and the profile and notes are deleted immediately; delete an entry and its content is gone everywhere it mattered.

Why this needs a hosted service

You might ask why this is not just a plugin with an API key field like the others. Choosing the relevant examples per form, keeping them consistent with your exclusions, building the site profile from your public pages and metering all of it needs a backend that knows your site. That backend is ours to run, which is why there is no key for you to manage and why the paid tiers exist. The free tier covers 250 scored submissions a month on one site, and every tier learns the same way.

What to expect

A correction is not a rule. It is an example, and the model weighs it alongside the site profile, your notes and the submission in front of it. What a correction cannot do is turn a spam filter into a keyword blocklist, and that is by design: GravityAgent is tuned so that when it is unsure, the entry comes through. On our benchmark of 220 reviewed submissions across six synthetic sites it produced 0 false positives on 95 genuine submissions, with 87.2% recall on spam, and every failure mode admits the entry rather than blocking it.

Correct it once. Then get on with your day.

Install free from WordPress.org · Read the full comparison with Spam Hexer and Zero Spam AI Review