How AI Detects Fake Reviews: A Middle School Lesson
A 7th grader finishes her weekend shopping research and arrives at class Monday with a verdict: the wireless earbuds she wants are definitely good because they have 2,000 five-star reviews. When you gently ask whether she checked who wrote those reviews, she looks at you like you’ve just asked her to audit a spreadsheet. She did not know that was a question she was allowed to ask.
That moment — the confident citation of a star rating as evidence — is not a failure of critical thinking. It is a media literacy gap, and it is not one anybody handed teachers a ready-made curriculum to fill. This post gives you the mechanism and the lesson: how AI detects fake reviews across three layers, where it goes wrong, what the law says now, and a 45-minute no-device activity that makes the whole system visible.
TL;DR: How AI detects fake reviews works across three layers running simultaneously. First, what the review says — natural language processing scans for generic over-positive language, missing specific details, and AI-generated text patterns. Second, how the review arrived — behavioral signals like sudden bursts of reviews within 48-72 hours of a product launch, brand-new accounts, shared IP addresses, and whether the reviewer actually purchased the product. Third, who wrote it — graph analysis maps relationships between accounts to detect coordinated campaigns. Amazon blocked 275 million suspected fake reviews in 2024 (About Amazon). Since October 21, 2024, buying or selling fake reviews is illegal under FTC Rule 16 CFR Part 465.
How does AI detect fake reviews?
AI platforms like Amazon, Yelp, and Google do not rely on a single detection method. They run three layers of analysis in parallel, each catching a different kind of signal. Language analysis catches suspicious writing. Behavioral analysis catches suspicious patterns in how reviews arrive. Graph analysis catches suspicious connections between reviewers. A fake-review campaign typically trips all three layers — but the interesting teaching moment is understanding why any single layer alone is not enough.
Here is how each layer works.
Layer 1 — what the review says
Natural language processing (NLP) scans the text of every review for writing patterns that correlate with inauthenticity. The signals it looks for include:
- Generic over-positive language with no specifics (“Amazing product! Works great! Love it!”)
- Missing concrete product details (a reviewer who owned the item for two weeks would typically mention something specific — dimensions, a particular feature, a use case)
- Unusual sentiment distribution — a review that is uniformly five-star effusive about every attribute, with no mentioned trade-off, reads differently from one that says “the battery is mediocre but the sound quality is worth it”
- AI-generated text patterns — certain syntactic structures and phrase rhythms appear at elevated rates in machine-generated text compared to human writing
That last signal has become harder to catch. A 2025 study on detecting AI-generated fake reviews found that post-ChatGPT fake reviews are now near-indistinguishable to both humans and automated classifiers at the language level alone (arXiv, 2025). This is the arms race in its current state: as text generation quality increases, language-only detection loses accuracy. That is precisely why platforms do not stop at Layer 1.
Layer 2 — how the review arrived

This is the information-dense layer — and the one most worth teaching explicitly. Even if a fake review’s language passes the NLP classifier, the circumstances of its arrival leave a different kind of evidence.
Behavioral and temporal signals the AI monitors include:
- Review bursts — a product with a normal trickle of daily reviews that suddenly receives 40 five-star reviews within 48-72 hours of a product launch or promotional push. Genuine organic reviewers do not all show up at the same moment.
- Account age and review velocity — an account created three days ago with twelve five-star reviews across unrelated product categories (phone cases, protein powder, garden hoses) is a different signal than an account with a three-year history of occasional reviews in a consistent category
- IP address clustering — reviews arriving from the same IP address or subnet suggest the same physical location (a review farm) rather than independent consumers
- Verified-purchase status — Amazon’s verified purchase badge confirms a real transaction occurred. Buying a product to leave a fake review costs money, which raises the operational cost of fraud. Reviews without verified purchase status receive less algorithmic weight (eComEngine)
- Review diversification pattern — coordinated campaigns often show unnatural category spread, as if accounts were assigned to cover multiple unrelated products to appear real
The comparison table below shows what each signal looks like in a normal vs. suspicious case:
| Signal | Looks normal | Looks suspicious |
|---|---|---|
| Account age | 2+ years old, reviews in consistent categories | Created within 30 days, reviews across unrelated categories |
| Review timing | Trickle over weeks; occasional bursts after legitimate media coverage | 30+ reviews arrive within 48 hours of launch with no external coverage |
| Verified purchase | Review confirmed to match a real order ID | No purchase record attached to the account |
| Language specificity | Mentions specific product details, trade-offs, use context | Generic praise with no identifying product knowledge |
| Reviewer’s other reviews | Varied products, varied ratings including 3- and 4-star | Only 5-star reviews, often for similar product niches |
| IP / device signals | Reviews from geographically distributed, distinct devices | Multiple reviews arriving from the same IP address or device fingerprint |
When you introduce this table to a class, the productive discussion question is: which of these signals could a sophisticated fake-review operation defeat? Students quickly realize that defeating all of them simultaneously is expensive — which is why the economics of fake reviews favor volume over precision.
Layer 3 — who wrote it

Even if individual accounts pass behavioral screening, coordinated fake-review campaigns leave a structural fingerprint that graph analysis can detect.
Graph analysis maps relationships between entities — reviewers, products, sellers — as a network of connected nodes. A legitimate review ecosystem looks like a sparse, random graph: most reviewers have no connection to the sellers they review, and their review histories span many unrelated products. A coordinated campaign looks structurally different.
The patterns graph analysis surfaces include:
- Accounts created on the same day or within a narrow window, reviewing the same product or seller in the same timeframe
- Groups of accounts that never review anything except products from one seller cluster — they show up in the graph as a tight “clique” with no outside connections
- Seller accounts with unusually star-symmetric patterns — almost exclusively 5-star reviews, with occasional 1-star reviews appearing on competitor products from the same account cluster
- Temporal synchronization — accounts that had been dormant for months all reactivating within the same 72-hour window
Graph analysis does not look at any single account in isolation. It asks: what does the structure of this network of accounts tell us that no individual account would reveal? That framing — the signal is in the relationship, not the data point — is a concept worth slowing down on. It is the same logic behind fraud detection in banking and social network analysis in research. The AI and credit card fraud lesson covers that structural pattern-detection idea from a financial angle, if your class wants a paired perspective.
Where AI gets fake-review detection wrong
Accuracy on the suspicious cases does not mean accuracy on the innocent ones. Any classifier that draws a boundary between “real” and “fake” will put some real reviews on the wrong side of that boundary. In the fake-review detection context, this matters in ways that affect real people.
New and infrequent reviewers get over-filtered. A first-time reviewer leaving an honest 5-star review looks, on behavioral signals alone, almost identical to a fake account: new account, no prior history, single review. The AI has no way to distinguish genuine enthusiasm from manufactured praise. Yelp’s recommendation software currently filters approximately 25% of submitted reviews as “not recommended” (Yelp support documentation). Many of those are real customers whose accounts simply have not built enough signal history.
Small businesses absorb the cost. A new local restaurant with a handful of first-time reviewers — people who downloaded Yelp specifically to leave a review — has those reviews suppressed. The national chain down the street, with thousands of established-account reviews, sees its aggregate score hold. The AI was not designed to disadvantage small businesses. But the detection logic, applied at scale, produces that effect.
The arms race runs in both directions. As platforms add behavioral and graph-level detection, fake-review brokers adapt: they build accounts with longer histories, spread reviews over time, simulate geographic diversity. A 2024 joint lawsuit by Amazon and Google against fake-review brokers gave a window into the scale of these operations — the named brokers were offering packages of thousands of coordinated reviews (Legal Dive). The same technical sophistication that makes detection possible makes evasion profitable.
AI-generated fakes vs. AI-based detection. The most direct arms-race problem is the one the 2025 research flagged (arXiv, 2025): language models can now produce fake reviews that are structurally and tonally indistinguishable from genuine ones. Layer 1 detection degrades. The platform falls back harder on Layers 2 and 3. A reviewer who generates a fake review but submits it from a genuine, long-standing verified-purchase account may slip through all three layers. That is not a hypothetical — it is where the detection problem currently sits. This is the same reason the spam filter’s arms race never ends: every defense teaches the next attack.
Is it illegal to write fake reviews?
Yes, and this became enforceable US federal law in late 2024 — which makes it a civics anchor for this lesson, not just a tech topic.
The FTC finalized Rule 16 CFR Part 465 on August 14, 2024, with the rule taking effect October 21, 2024 (Consumer Finance Monitor). The rule explicitly prohibits:
- Creating or purchasing fake consumer reviews, including AI-generated ones
- Paying for reviews without disclosing the compensation
- Publishing reviews from company insiders without clear disclosure of the relationship
- Suppressing negative reviews through threats or selective-display practices
- Using IP addresses or other technical means to manipulate review distribution
The penalty is up to about $53,000 per violation (the FTC’s civil penalty cap, adjusted for inflation as of 2026). Each individual fake review can constitute a separate violation.
On the enforcement side: Amazon blocked 275 million suspected fake reviews in 2024 alone and, in October 2024, joined Google in a joint lawsuit targeting fake-review broker networks — the first major coordinated action between competing platforms against a shared threat (About Amazon; Legal Dive).
One useful postscript for the class: Fakespot, the browser extension that graded Amazon and Yelp reviews for authenticity and had been a go-to recommendation for consumer education, was shut down by Mozilla on July 1, 2025 (Distractify). The tool that taught consumers to question star ratings is gone. The question it was asking — is this review real? — is not.
A no-device lesson: Be the Review Detective

No computers, no accounts, no browsing required. A class runs this in 45 minutes.
Minute-by-minute pacing:
- 0–5 min — Hook. Project a single product listing (described verbally or on the board): “A wireless earbud with 4.8 stars and 2,400 reviews. How do you know whether to trust it?” Students call out their answers. Do not confirm or deny yet. Capture responses on the board.
- 5–15 min — Teach the three layers. Walk through Layer 1 (language), Layer 2 (behavior), and Layer 3 (network) using three signals each. The Layer 2 comparison table from this post works as a printed or projected handout. Keep it to one concrete example per layer.
- 15–35 min — Review Detective card sort. Each group receives a set of printed review cards — two real-style reviews and two fake-style reviews (hypothetical, written by the teacher). Students apply the four-step observation protocol below and mark each card: PASS, FLAG, or INVESTIGATE.
- 35–45 min — Debrief. Which cards split the room? What signal tipped the decision? What would a better fake have looked like? (Students defending their own detection logic against adversarial improvement is the same muscle the spam filter lesson calls “the Spammer Round” — understanding by trying to beat it.)
The four-step observation protocol (print and give to students):
- Check the star ratio — Does this listing have almost no 3- or 4-star reviews? An unusually perfect curve is a signal.
- Check the reviewer account — Is the account newly created? Does it review only 5-star products across unrelated categories? Does it have a verified purchase badge?
- Check the language — Is there any specific product knowledge here? Can you imagine a person who actually used this item writing exactly these words?
- Check review timing — Do the reviews arrive in a single burst, or over time? (For live activity purposes, describe this as a scenario on the card: “50 reviews appeared on this product in two days, after a quiet six months.”)
Worked example — side by side:
Below are two invented hypothetical reviews for a generic “SoundWave X3 wireless earbuds listing.” Real brands not required.
Review A — signals pointing suspicious:
“These earbuds are absolutely fantastic!!! Best purchase I’ve ever made. Sound quality is perfect and they are very comfortable. Would recommend to everyone. 5 stars.”
What tips you off:
- No specific product detail mentioned — no battery life number, no mention of fit, no comparison to another product
- “Best purchase I’ve ever made” with no context or trade-off
- Three exclamation points in the opener
- Account: created 14 days ago, this is its only review
Review B — signals pointing genuine:
“I use these for a 45-minute commute each way. The seal isn’t as tight as my old earbuds, so I can still hear bus noise at moderate volume, but the bass is noticeably better. Battery lasted about 7 hours before I needed to recharge. Decent for the price.”
What tips you off (toward genuine):
- Specific use context (45-minute commute)
- Named a point of comparison (the old earbuds)
- Acknowledged a real trade-off (noise isolation vs. sound quality)
- Specific data point (7-hour battery)
- Language matches someone who actually tested a thing
When you run this side-by-side with a class, students typically reach a consensus on Review B quickly. Review A takes longer, because the language is familiar — it sounds like plenty of reviews they have read before. That recognition is the lesson: students have been reading Review A-style fakes for years without noticing, not because they were careless, but because no one showed them the signals.
Standards this lesson covers
| Activity step | Standard | Anchor code |
|---|---|---|
| Evaluating review cards for credibility signals | ISTE Knowledge Constructor — evaluate accuracy, perspective, credibility, relevance of info | ISTE 1.3.b |
| Exploring the fake-review system as a real-world issue | ISTE Knowledge Constructor — explore real-world issues using an inquiry-based strategy | ISTE 1.3.d |
| FTC Rule discussion — legal/ethical dimensions of fake reviews | ISTE Digital Citizen — safe, legal, and ethical behavior online | ISTE 1.2.b |
| Three-layer detection system (AI learns from labeled data) | AI4K12 — machines learn from data; model accuracy depends on training data quality | AI4K12 Big Idea #3 (Learning) |
| False positives + equity effects on small businesses | AI4K12 — AI affects individuals, communities, and society | AI4K12 Big Idea #5 (Societal Impact) |
| Review Detective card sort, class debrief | CCSS.ELA-LITERACY — collaborative discussion with diverse partners | CCSS.ELA-LITERACY.SL.7.1 |
| Evaluating review cards as information sources | CCSS.ELA-LITERACY — gather relevant information; assess credibility and accuracy of sources | CCSS.ELA-LITERACY.W.7.8 |
Standards sources: ISTE Standards (ISTE is a registered trademark of the International Society for Technology in Education; these resources are not affiliated with or endorsed by ISTE). AI4K12 Big Ideas: ai4k12.org.
CCSS.ELA-LITERACY.W.7.8 deserves a second look here. “Gather relevant information from multiple sources; assess the credibility and accuracy of each source” — reading a product review as an information source with credibility signals to evaluate is that standard applied to a text type students encounter constantly. The four-step protocol turns a routine consumer activity into structured source evaluation. That crossover — a skill taught in ELA applying directly to a decision students make this weekend — is where the lesson earns its place in the schedule.
Get the materials for this lesson
Nobody handed you a curriculum for this. The question of whether a star rating is trustworthy, what an AI does to answer that question, and what federal law says about the answer — these did not arrive pre-packaged in a department filing cabinet. That is the gap, not a gap in your preparation.
Three options matched to where you are this week:
-
Start with no-device, no-prep — Unplugged AI Activities — 10 No-Tech Lessons (Card Sort, Role-Play, Grades 6-7-8). The Review Detective card sort and nine additional unplugged AI activities, all print-and-go PDF. No devices, no prior AI background needed.
-
Go deeper on the mechanism — Types of AI Deep-Dive Lesson — Narrow vs. General vs. Super AI (Grades 6-7-8). When students ask “but why can’t the AI just understand what the reviewer means?” — this lesson answers that with the narrow vs. general AI distinction that makes the review detector’s limitations make sense.
-
Build the full unit — How AI Works MEGA — 16 Machine Learning and AI Concept Lessons (Grades 6-12). The mechanism behind fake-review detection — pattern recognition, behavioral signals, network analysis — sits inside a larger landscape of how AI learns and where it fails. This 16-lesson pack covers that landscape. Browse the complete catalog at /shop.
The student who cited 2,000 five-star reviews as her evidence was being perfectly rational with the information she had. The lesson is not about distrust — it is about reading a star rating the same way she reads any other source: who said it, under what conditions, and what would it look like if it were not true. That is a teachable question. And now you have the mechanism to teach it.
This post was drafted with AI assistance and human-finalized.
Quick questions
Amazon's systems check three layers at once: the language of the review (generic, over-positive, or AI-generated text), the behavior behind it (brand-new accounts, review bursts, shared IP addresses, no verified purchase), and the network it belongs to (clusters of accounts created together to boost one seller). Amazon reported blocking 275 million suspected fake reviews in 2024.
Automated systems flag signals, not intent. A genuine first-time reviewer with a new account and no history looks almost identical to a fake one, so honest reviews sometimes get suppressed. Yelp's software filters roughly 25% of the reviews it receives as 'not recommended,' and some of those are real customers.
Yes. Since October 21, 2024, the FTC's Rule 16 CFR Part 465 makes it illegal to create, buy, or sell fake consumer reviews — including AI-generated ones — with civil penalties of up to about $53,000 per violation.
Get the free AI-Proof Assignment Toolkit
10 ways to redesign any assignment so an AI chatbot structurally can't do it — plus a redesign worksheet, a 45-minute lesson, the “Spot AI Work” card, and parent templates. One email, all 5 pieces.
Prefer the full breakdown? See everything inside the toolkit →