Detecting AI-generated reviews in 2026: what still works, what does not

Why language classifiers alone have failed

In 2022, a classifier trained on ChatGPT output caught 88% of fake reviews at a 3% false-positive rate. The same classifier tested against 2025 model output catches under 41%. Language is no longer a distinguishing feature — modern LLMs write more fluently than most real reviewers.

What still works: provenance

The single strongest signal is whether a review is tied to a real transaction. Invitation tokens, receipt uploads, and identity verification cannot be forged by an LLM. This is why ScoreReview UK's verification ladder is the load-bearing part of the moderation stack, not a marketing tier.

What still works: reviewer graph analysis

LLMs write reviews. They do not build years of coherent, cross-platform reviewer history. Graph analysis — who else has this account reviewed, when did they join, what other identities do they share IP ranges with — remains a strong signal because it operates on behaviour, not text.

• Account age and prior review distribution
• IP / device / geolocation coherence
• Timing patterns across the reviewer's history
• Cross-platform handle correlation

What still works: transactional grounding

A real customer names the plumber, mentions the invoice number, references the day of the week. LLMs generate plausible reviews but avoid specifics they cannot verify. Reviews with zero verifiable transactional detail cluster hard in the suspect distribution.

What no longer works: perplexity and burstiness

The 2023-era detectors that measured statistical text patterns give near-random results on 2025 model output. Any vendor still selling perplexity-based detection as a primary defence is selling a 2023 product.

Human moderation is not optional

Automated systems handle roughly 92% of decisions with high confidence. The remaining 8% require a human moderator with sector context. Platforms without a human review layer publish more fakes — the maths is unforgiving.

What the regulator expects

The CMA's 2025 review-platform guidance treats 'reasonable steps' as a stack of controls, not a single detector. A platform that relies on any one signal — however clever — is undershooting.