C2, Lesson 4: Loss Prevention and the Cost of False Alarms
Tags: Sourced / General / Our view / Illustrative. About 12 minutes.
1. The story: Rite Aid, 2012 to 2023
On December 19, 2023, the US Federal Trade Commission announced that Rite Aid would be banned for five years from using facial recognition for surveillance purposes. The FTC's complaint said that from 2012 to 2020 Rite Aid used AI-based facial recognition to identify customers who might be shoplifting, and that the system falsely flagged people who then were followed, searched, told to leave, or accused in front of others. [Sourced: FTC press release, Dec 19, 2023; these are the FTC's allegations in a settlement] The complaint said Rite Aid built a database of tens of thousands of images, many low quality, from security cameras, employee phones and even news stories; that it did not test the accuracy of the technology before deploying it or track false-positive rates afterwards; that it did not train staff to treat alerts as possibly wrong; and that false positives were more likely in stores in plurality-Black and Asian communities. [Sourced: FTC; allegations] The order also requires deleting the images and algorithms built from them, notifying people, and a written complaint process. The stipulated order was entered by the federal court in the Eastern District of Pennsylvania on February 23, 2024, with the five-year ban running from its effective date, and the FTC modified its separate earlier administrative order on March 5, 2024. [Sourced: FTC order PDF (signed Feb 23, 2024; docket entered Feb 26); FTC Commission order, issued Mar 5, 2024] Rite Aid later filed a second Chapter 11 in May 2025 and wound down; we do not say the ban has expired. [Sourced: verifier report; not independently read] The lesson is not "never use AI in loss prevention." It is that an alert is a guess, and what you do with a guess is the real decision. [Our view]
2. The one idea
Loss prevention AI finds patterns that might be loss. Every system makes two kinds of mistakes: missed losses and false alarms. You choose the balance, and false alarms land on real customers and staff. [General]
3. The kitchen-table version
A smoke alarm that goes off every time you toast bread gets its battery pulled. A shop alarm that points at innocent customers gets you a bad reputation. Alarms are only useful when the person who hears one knows how to respond gently and has a way to check. [General]
4. What retailers lose and where AI looks
Shrink is lost inventory from theft (customers and staff), process errors, vendor fraud and damage. The National Retail Federation's 2023 survey reported shrink of 1.6% of sales for fiscal 2022, about $112.1 billion, with external theft and employee theft as leading causes (see Lesson 1 and the NRF source). [Sourced: NRF, National Retail Security Survey 2023] Where AI is used: (a) exception reporting on returns, voids and discounts at the till; (b) camera video analytics, like unscanned items at self-checkout; (c) inventory counts that do not match sales; (d) organized-retail-crime case linking; (e) facial recognition, the riskiest. [General]
5. Worked example: why accuracy numbers mislead
All numbers [Illustrative], computed by script. A store group has 100,000 visits a week. Suppose 200 are by real repeat offenders. A system catches 80% of them and falsely flags 0.1% of everyone else.
- Real catches: 160
- False flags: 99,800 x 0.001 = about 100
- Of all alerts, 61.6% are correct. Now loosen the system so it flags 0.5% of innocents: false flags 499, correct share 24.3%. At 90% catch and 1% false flags: 180 real, 998 false, correct share 15.3%. More sensitivity catches more real cases but, with rare offenders, most alerts become wrong. If staff act on every alert, many innocent people are confronted. Rite Aid's complaint describes this outcome. Before launch, ask for the false-alert rate, the test data, and who reviews each alert. [Our view]
6. A safer pattern for small and mid-size stores
- Start with data about transactions and staff process (returns, voids, discounts), not faces. [Our view]
- Use alerts to start a review, not an accusation. Wording for staff: "Can I help you find something?" [Our view]
- Log every alert and its outcome. Track the false-alert rate monthly. [Our view]
- Keep human sign-off before any ban, police call or search. [Our view]
- Tell people when you use cameras or biometrics. For comparison, the FTC's December 2023 proposed order for Rite Aid required clear notice to consumers about facial recognition in its stores, notice when their biometric information is enrolled in a database, and a written response to complaints. [Sourced: FTC press release, Dec 19, 2023; terms of one settlement with one company, not a rule for every store] This is not legal advice.
7. What it costs and what it breaks
- Software and cameras, plus staff time to review alerts.
- Customer trust and legal exposure from false accusations.
- Staff morale if alerts are used to blame employees without review.
- Data risk: image databases are sensitive and must be secured and deleted on schedule.
- The Rite Aid order shows regulators treating this as a consumer-protection matter. [Sourced: FTC]
8. Monday checklist
- Write what you lose, by type (theft, error, vendor, damage), even roughly.
- List every alert system you use and what staff do on an alert.
- Ask each vendor for false-alert data and testing, in writing.
- Set a rule: no accusation or ban from an alert alone.
- Check local law and post notice for any camera analytics.
9. Try it yourself (no code, 30 minutes)
Pick numbers for your store: weekly visits, real incidents, catch rate, false-flag rate. Compute the share of alerts that are correct (Section 5). Decide the maximum false alerts a week your staff can handle politely.
10. Self-check
- What did the FTC say about Rite Aid's testing before deployment? Alleged it did not test or inquire about accuracy.
- How long is the facial-recognition ban? Five years.
- In the example, why did correct alerts fall when false flags rose? Real offenders are rare, so false alerts outnumber them.
- Name one safer starting point than facial recognition. Transaction exceptions such as returns and voids.
- What should happen before a ban or police call? Human review.
11. Where this is going
Expect more rules on biometrics and automated decisions about shoppers. Tools that explain why an alert fired will be easier to defend. [Our view]
Sources
- FTC press release, Dec 19, 2023: https://www.ftc.gov/news-events/news/press-releases/2023/12/rite-aid-banned-using-ai-facial-recognition-after-ftc-says-retailer-deployed-technology-without
- FTC case page: https://www.ftc.gov/legal-library/browse/cases-proceedings/2023190-rite-aid-corporation-ftc-v
- NRF National Retail Security Survey 2023: https://nrf.com/research/national-retail-security-survey-2023
Review flags
- Rite Aid details are FTC allegations in a settlement. RESOLVED (Oct 7, 2026): the FTC press release and case page were read at ftc.gov. The order's judge signed it Feb 23, 2024 (the order PDF says "SO ORDERED this 23rd day of February, 2024"); the court docket entry shows it entered Feb 26, 2024; the FTC case page lists the stipulated order under March 8, 2024 (posting date). The Commission modified the 2010 order on Mar 5, 2024 (Commission order reopening proceedings, "ISSUED: March 5, 2024"). The full complaint and order were not read in full. Rite Aid's May 2025 second Chapter 11 is not independently read, and the lesson attributes it to the earlier verifier report.
- The "tens of thousands of images" and bias statements come from the press release.
- RESOLVED (Oct 7, 2026): Section 6 was trimmed. The general legal commentary ("check local law, rules vary") was removed. The only legal content left is FTC press release facts, dated and attributed, plus the "not legal advice" line. The practical steps 1-4 are labeled as our view and make no legal claims. No lawyer review flag remains on this basis.
- Example numbers are illustrative.
Claim ledger (C2 Lesson 4)
| # | Claim | Source | Verified | Status | Re-check |
|---|---|---|---|---|---|
| 1 | Dec 19, 2023 FTC announced 5-year ban | FTC release | 2026-10-06 | Read | order status 2026-12 |
| 2 | 2012-2020 use; false flags; no accuracy testing; image quality; training | FTC release | 2026-10-06 | Read; allegations | read complaint |
| 3 | Higher false positives in plurality-Black and Asian store areas | FTC release | 2026-10-06 | Read; allegation | read complaint |
| 4 | Order requires deletion, notice, complaint process; entered E.D. Pa. Feb 23, 2024; FTC modified 2010 order Mar 5, 2024 | FTC release; order PDF | 2026-10-06 | Dates verifier-confirmed | none |
| 5 | NRF 1.6% / $112.1B | NRF 2023 | earlier | Verified earlier (Lesson 1) | none |
| 6 | Alert precision arithmetic | Our computation | 2026-10-06 | Illustrative, script-checked | none |