How PruneSocial scores your social media

What PruneSocial is, and what it isn't

PruneSocial is a self-review tool. You download your own data from each platform, upload it, and get a report only you see. We don't search for you, scrape anything, connect to your accounts, or ask for passwords.

That's different from an employer's social media background check. When a screening company assembles social media information about a candidate and furnishes it to an employer, the FTC treats that company as a consumer reporting agency under the Fair Credit Reporting Act, with duties like reasonable accuracy procedures and adverse-action notices.^[1][2] PruneSocial does not furnish reports to employers, recruiters, background-check companies, or other third parties. It's designed and operated solely as a self-review tool for the person whose account activity is analyzed. It also can't tell you what a specific employer or vendor will find or decide.

Where the categories come from

We don't have access to any vendor's proprietary classifier, and we don't claim to reproduce one. Instead, the categories are modeled on what screening vendors say publicly about what they flag:

• Fama describes an out-of-the-box set of eight behavior flags, and its public descriptions cover threats, violence, trolling and harassment, and intolerance, among others.^[3]
• Social Intelligence groups flagged content into four broad filters: intolerance, violence including threats, potentially illegal activity, and sexually explicit content.^[6]
• Checkr says its social media search categorizes content across areas including violent behavior, illegal activity, hate speech, harassment, sexually explicit content, and drug use.^[7]

Why cannabis counts even where it's legal: Fama says its drug flag covers posts about cannabis as well as illegal substances.^[3] State law doesn't change what a vendor's category is built to catch, so PruneSocial flags cannabis references too, but caps casual ones at Consider.

For context on what actually shows up, Fama's 2025 report said harassment and trolling made up the largest share of flagged behavior, followed by intolerance and sexually explicit content.^[8]

Why seven years

PruneSocial scores items from the last seven years and skips anything older. That matches what vendors disclose: Checkr describes a default seven-year lookback for its social media search, and Fama says its screening reviews public content as far back as seven years.^[7][5] Your report tells you how many older items were skipped.

One exception: X's export doesn't record when you liked something, only when the liked post was published. So all X likes are scored, whatever their date. X also made likes private for everyone in June 2024, so nobody else can see your X likes. They are still part of your own history, but a flagged X like is the lowest priority to clean up.

How each item is scored

Every item is scored on its own, using only its own text. The model works through the same steps every time and stops at the first one that settles it:

1. Does it fit a category at all? If not, it's Clear. Profanity, sarcasm, strong opinions, and odd tone aren't categories on their own.
2. Does a documented exclusion apply? Fiction and media references, sports, news and commentary, charity, advocacy, recovery, lawful adult alcohol use, political opinion, and jokes that don't describe real conduct are all Clear.
3. Is it about your own real conduct, endorsement, or intent? A like counts only when the post you liked clearly falls in a category.
4. Is key context missing? If an item clearly points toward a category but could be read either way, it's Consider, marked "Needs context."
5. Otherwise, how strong is the match? Consider for a plausible but mild match, Review for a clear, strong one.

Precision comes first. A false flag on a fundraiser or a movie quote teaches you to ignore the report, so when an item is borderline between Clear and Consider, it's scored Clear. The trade-off is that the report can miss things. That's why it's an estimate, not a guarantee.

Protected characteristics are never, by themselves, a reason for a flag. Religion, sexual orientation, gender identity, pregnancy, disability, health, age, race, and national origin aren't scoring categories, so a post that reveals them is Clear. If an item separately matches a risk category, such as a direct threat, it can still be flagged for that separate reason. Screening vendors say they remove protected-class information from what employers see, and PruneSocial follows the same line.^[4]

What Review, Consider, and Clear mean

These examples are illustrative, not from real users.

Fixes are chosen by what the item is, not left to the model. A like always gets "Remove your like," never "Delete the post," because you can't delete someone else's post.

Reputation flags: a second lens

The ratings above follow the categories screening vendors publish, and they exclude things like political opinion. But the cases in our Found Later research show that people also lose jobs, seats and admissions over content that matches none of those categories: hostile political posts, jokes about a death or tragedy, public-health claims, complaints about an employer, sweeping claims about a group of people.

So every item that stays Clear is checked a second way. If it plausibly fits one of seven reputation categories (political or partisan commentary, a news event, death or tragedy, public-health claims, workplace or customer criticism, group generalization, a dark or violent remark, or a racially charged remark), it appears in a separate “Reputation flags” section of your report. These flags are not Review or Consider ratings and don’t change your overall result. This lens leans toward flagging, so you will see some items you consider harmless. Keeping, editing or deleting any of them is your decision.

What's scored on each platform

PruneSocial scores the text activity each platform's export actually provides. That differs by platform:

Likes and reactions are covered unevenly, because the platforms export them unevenly. Read with the liked text: Instagram liked posts that have a caption, and X likes. Names only: Facebook reactions record the Page or person whose post you reacted to but not the post, so PruneSocial screens those names (it can catch a like of a hate group or explicit Page, not a problem post on an ordinary Page). Not readable: Instagram likes with no caption, story likes and liked comments, Threads likes, TikTok likes, YouTube liked videos, Reddit votes and LinkedIn reactions. Each report lists the ones that applied to your exports, so a clear result is never read as covering them.

PruneSocial is intentionally more conservative than a public-only screening. Your export includes your own content whether or not it's public, so some items may be scored even though an employer or screening provider couldn't see them today, such as friends-only posts or X likes. That lets you review your own history more broadly, but a PruneSocial flag does not mean the item is currently public or that an employer can find it. It's still worth a look, since posts can be shared, screenshotted, or made public later.

What PruneSocial can't see

• Images and videos. What a photo or video shows isn't judged, and the words inside memes, screenshots and quote cards aren't read yet. Review those by hand. We built an on-device reader and switched it off until it's accurate enough to be worth your time.
• Content about you. Posts, photos, and other content created by someone else that tag or mention you aren't reliably available in the supported text PruneSocial scores from your own exports. Review those separately.
• Private messages and saved posts. Never scored, by design. A message goes to specific people and a saved post is a private bookmark, so neither is something you published. Posts and comments are scored even when their audience is limited, because anything you published can be shared or made public later. The one exception is Pinterest, where Pins you saved to a public board are scored, because a public board shows them under your name. Secret boards are skipped.
• Very long items. Individual items longer than about 4,000 characters are shortened before scoring, so the end of a very long post may not be read. For a very large history the limit can drop to 1,500 or 600 characters so the request fits. The report’s “What this report does not cover” section says so.
• Reactions and likes without text. Facebook reactions (apart from the Page or person's name), Threads, TikTok and LinkedIn likes, YouTube liked videos, Reddit votes, and Instagram likes with no caption are recorded without the content, so those can't be scored. PruneSocial doesn't fetch them from the platforms, because that would mean sending your account details off your device.
• Platforms you didn't upload, and anything outside the seven-year window.
• The rest of the web. Employer vendors may also search news, public web pages, and other sources beyond social media.^[7][5]

Every report ends with a short list of manual checks for these gaps.

How we test and improve it

Before any change to the scoring rules ships, it's re-run against a set of labeled test items, including deliberately tricky ones like movie quotes, sports trash talk, fundraisers, and political opinions, to make sure accuracy on known cases doesn't slip. That set is growing with feedback from beta testers.

Every flag in a report has a "Wrong flag?" link. Those reports, plus content-free counts from each report (how many items were scored and flagged, by category and platform, with no post text and no email address), are how we find and fix weak spots. Changes are listed in the changelog below.

What happens to your data

Export ZIP files are opened and read in your browser. Only the supported text of your activity (posts, comments, likes, and bios, with the date and type of each, plus a link to the item when the export includes one) is sent. Only the text is scored. Scoring runs through Anthropic's API, which doesn't use it to train models. Your uploaded content is deleted when your report is emailed, or within about a day if anything interrupts that, and we don't keep a copy of the report. The full details are in the Privacy Policy.

Sources

1. FTC staff closing letter to Social Intelligence Corp., May 9, 2011
2. FTC Business Blog: The Fair Credit Reporting Act & social media, June 2011
3. Fama: How Fama's online screening solution detects misconduct at work
4. Fama: What is social media screening?
5. Fama: Does social media screening only cover social media platforms?
6. Fama: What is an inappropriate social media post? (Social Intelligence's four filters)
7. Checkr: Social media search
8. Fama 2025 State of Misconduct at Work report, via HR Technology Wire

Changelog

• October 23, 2026: Findings now carry a direct link to the item when the export provides one (X posts and likes, Reddit, LinkedIn posts and comments, YouTube comments, and Instagram and Threads items where Meta includes a link), alongside TikTok and Pinterest. Links are never scored. Every paid report with findings now includes an agent cleanup file and prompt for your own AI agent (how it works). Scoring is unchanged.
• October 18, 2026: Added reputation flags, a second lens for content that matches no screening category but has cost people jobs. They are listed separately and never change a rating. Facebook reactions are now screened by the name of the Page or person, and the report lists, per platform, the likes and reactions an export can't show. The scan does not read images yet.
• September 26, 2026: Methodology published.