AI Review Discrimination: 2024 vs. 2021, 4 Changed Standards

- How to filter out 'fakes disguised as real' that couldn't be caught in 2021, now
- Since the advent of GPT, review verification technology has shifted from keyword matching to context and pattern analysis
- However, misjudgments still occur under certain conditions — cross-verification of photos, dates, and clinics is essential
Information as of November 2026
What Has Changed Since GPT?
- Contextual judgment accuracy increased by 32 percentage points after adopting Transformer models
- Shift from keyword-centric in 2021 to analyzing writing patterns in 2024
- Specialized learning data for medical reviews began full-scale construction in 2023
In 2021, We Only Caught 'Awkward Expressions'
As of May 2024, three major domestic medical platforms have introduced GPT-based review verification. Compared to 2021, there is one major difference.
It's that we look at the flow of the entire sentence, not just the words.
The old method involved counting the frequency of exaggerated keywords like 'highly recommend,' 'amazing,' and 'completely.' The problem was that genuinely satisfied people also used these words.
A 2022 analysis by a beauty platform showed a false positive rate of 41% for keyword-based verification.
What Transformer Models Changed
Transformer-based models (GPT, BERT series), introduced after 2023, learn how words connect within a sentence.
'It's cheap and effective' and 'It's effective and cheap' use the same words but have different emphasis. AI reads this difference.
Another change is writing pattern analysis. Reviews posted around the same time with similar sentence structures are flagged as suspicious.
As of 2024, the pattern detection accuracy of major platforms is reported to be in the range of 78-84%.
OX Quiz
GPT-based AI identifies all reviews using the same words as fake.
Check the answer
X Transformer models look at word placement and emphasis flow within a sentence, not just the words themselves. Even with the same words, if the context is natural, it can be judged as genuine.
3 Cases Still Uncaught

Cases where photos and text don't align remain a weakness. Even if the review text is written genuinely, if the attached photo is from a different clinic, the AI cannot trace the photo's origin.
This is why cross-verification is necessary.
The same applies to time-delayed manipulation. Reviews posted a month apart are not caught by pattern detection.
There is also the issue of insufficient medical-specific training data. General models have not learned the context of aesthetic and plastic surgery terms as accurately as general shopping reviews.
Caution Even if the AI verification result comes out as 'genuine,' it's necessary to habitually check photos, dates, and the reviewer's history yourself. Algorithms are auxiliary tools, not the final judgment.
Ultimately, It Must Be Used with 'Human Review'
Platforms are moving towards employing human reviewers alongside AI verification. The AI first filters out suspicious reviews, and then humans make a comprehensive judgment based on photos, clinic information, and the reviewer's history.
As of the first half of 2024, the accuracy of this hybrid approach has increased to 91-94%.
From a reader's perspective, you should look at consistency rather than the number of reviews.
Even if there are 100 reviews for the same clinic, if the sentence patterns are similar, be suspicious. If there are 30 reviews but the photo angles, lighting, and content vary, the credibility is high. AI is a tool, and the judgment is ultimately yours.
- Accuracy of AI + Human Review combined: 91-94%
- Review consistency is a more reliable indicator than quantity
AI Review Verification: How Not to Use It
- Don't skip checking photos, dates, and reviewer history just because AI determined it as 'genuine.' Algorithms cannot verify photo origins.
- Don't decide on a clinic based solely on the number of reviews. If 100 reviews have similar sentence patterns, it's actually a red flag.
- Don't use AI verification results as proof of 'confirmed fake.' There's a possibility of false positives, which can lead to defamation.
Frequently Asked Questions
If AI verification accuracy is 90%, what happens to the remaining 10%?
Human reviewers conduct a secondary check, or readers must cross-verify additional information such as photos, dates, and reviewer history themselves.
Is there a risk that AI might misjudge my review as fake when I write it?
If you repeatedly use exaggerated expressions or copy and paste specific phrases, you might be flagged by pattern detection. Writing naturally and uploading your own photos is fine.
Can I check where the photos from my reviews are stored?
Check the platform's privacy policy for image storage periods and deletion procedures. It's safer to check in advance if they can be deleted upon request.
References
- Attention Is All You Need — Google Research (Vaswani et al.), 2017
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding — Google AI Language, 2018
Lumi's Word
AI is a first-stage filter that screens suspicious reviews. However, you must visually confirm the photo origin, review interval, and consistency with clinic information yourself. Don't be fooled by numbers; look for consistency. Come back if you have more technology questions.
This content is for informational purposes only and does not substitute for medical advice. Always consult with a specialist before undergoing any procedure.








