How Kitha Detects AI Tweets: Methodology & Accuracy

Q: Can AI text be written to fool Kitha?

Paraphrasing and humanizer tools can lower any detector's confidence. No detector is immune; detection adapts over time.

Quick Answer: Kitha uses a machine-learning classifier trained on a large corpus of human-written and AI-generated text. For each tweet it analyzes statistical and linguistic patterns — predictability, sentence variation, and phrasing — and outputs an AI-likelihood score. On our internal test dataset it reaches 95.6% accuracy. That figure is a benchmark, not a guarantee: like every detector, Kitha is less certain on very short posts and should be read as a strong signal, not absolute proof.

We think anyone trusting a detection score deserves to know how the number is produced and where it breaks down. Here's the honest version.

What the Model Does

Kitha is a trained classifier. It learned the difference between human and AI writing from a large set of labeled examples, then applies that learning to new tweets. For each post it weighs signals such as:

Predictability (perplexity): AI models pick high-probability words, so their text is statistically less surprising than typical human writing.
Sentence variation (burstiness): Humans mix short and long sentences; AI tends toward uniform length.
Linguistic patterns: phrasing, structure, and connective habits that differ between people and language models.

These combine into a single probability that a tweet was AI-generated, which the extension turns into the badge you see in your feed.

What "95.6% Accuracy" Actually Means

The 95.6% figure is the model's accuracy on our test dataset — a held-out set of human and AI examples the model didn't train on. It's a standard way to benchmark a classifier.

What it does not mean:

It is not a promise that every individual tweet is judged correctly.
It depends on the mix of data in the test set; real-world feeds differ.
Accuracy is lower on very short tweets than on longer posts, because short text carries less signal.

We publish the number because it's a fair benchmark — and we'd rather be upfront about its limits than imply certainty.

Where Detection Is Hard (Honestly)

Every AI detector, Kitha included, struggles with the same things:

Short text. A 15-word tweet gives any model far less to work with than an essay. Confidence drops accordingly.
False positives. Clear, formal human writing can look "AI." That's why a high score means look closer, not guilty.
Paraphrasing and "humanizers." Tools that rewrite AI text can lower its detectability.
A moving target. As models change, detection has to keep adapting — it's an ongoing effort, not a solved problem.

For the practical signs you can check yourself, see 9 signs a tweet was written by AI.

How We Handle Your Data

Detection runs without storing your content. Kitha caches only anonymized result data temporarily (24 hours) to keep the feed fast — it does not store the actual text of tweets or personally identifiable information. Full details are in our privacy policy.

How to Use Scores Well

Treat the badge as a flag, not a final ruling.
Give short tweets extra benefit of the doubt.
Combine the score with behavioral signals for replies (see how to spot AI bot replies on X).

Key Takeaways

Kitha is a trained classifier using predictability, sentence variation, and linguistic patterns.
95.6% is test-set accuracy — a benchmark, not a per-tweet guarantee.
Short text is the hardest case for any detector.
Kitha stores no tweet content — only anonymized results, for 24 hours.

Frequently Asked Questions

Q: Is Kitha's 95.6% accuracy guaranteed for every tweet? A: No. It's the model's accuracy on a test dataset. Individual tweets — especially short ones — can be misjudged, so treat the score as a strong signal, not proof.

Q: Why are short tweets harder to detect? A: Less text means fewer statistical signals. Detection reliability rises with length, so very short posts are the toughest case.

Q: Can AI text be written to fool Kitha? A: Paraphrasing and "humanizer" tools can lower any detector's confidence. No detector is immune; detection adapts over time.

Q: Does Kitha store the tweets it analyzes? A: No. It caches only anonymized result data for 24 hours and never stores tweet content or personal information.