false positives

Bilingual

Expansion of Flock Camera Systems Is Leading to More False Positives
2 months ago
- Flock Safety's expansion into law enforcement has led to widespread criticism and multiple cities canceling agreements due to public outcry.
- The system generates numerous false positives, as seen when automated plate readers misread New Jersey license plates, causing innocent drivers to be wrongly targeted.
- In Minnesota, a journalist test-driving a Range Rover was surrounded by police because Flock misread the plate as stolen due to a missing small digit on NJ plates.
- In Colorado, a woman was wrongly summoned for package theft based solely on a Flock camera sighting, and had to prove her innocence with her own evidence.
- Law enforcement's reliance on Flock's flawed tech allows cops to make errors faster and at scale, undermining the presumption of innocence and forcing citizens to document their movements.
- Flock's defense that it cannot control police use is weak, as it has made no effort to limit abusive access until pressured externally.
Better Bayesian Filtering
2 months ago
- The author discovered that lazy evaluation of research papers works: by writing without citations, readers will send references.
- Bayesian spam filtering was known since 1998, but early filters (Pantel & Lin) had poor performance (92% catch, 1.16% false positives), while the author's later filter achieved 99.5% catch with <0.03% false positives.
- Five key differences explain the performance gap: training on too little data, ignoring message headers, stemming tokens, using all tokens instead of the 15 most significant, and not biasing against false positives.
- Tokenization improvements include preserving case, treating exclamation points as constituents, marking tokens from headers/URLs, and using degeneration to handle unknown tokens by falling back to less specific versions.
- A larger vocabulary (187,000 tokens vs. 23,000) increases discrimination but risks misses; degeneration helps by looking up progressively less specific tokens.
- To handle equally interesting tokens, the author scales probabilities for tokens that occur only in one corpus (e.g., .9999 for >10 occurrences, .9998 otherwise).
- HTML handling should be moderate: ignore most tags but parse links, images, and font tags; future spams may avoid HTML entirely.
- Over one month, the filter caught 99.75% of spam (4 missed out of 1750) but had 3 false positives out of ~7740 legitimate emails (0.06% rate).
- False positives are treated as bugs, not performance metrics; they often come from newsletters, company emails, or unusual formatting (e.g., all uppercase).
- Future work focuses on two hard-to-filter spam types: personal ad spams (using neutral language) and outsourcing spams from companies (e.g., Bulgaria) that use programming jargon.
- Individual user-specific filters are advantageous because they create diverse probability profiles, slowing spammers' ability to optimize their attacks.
- Network-level filters are less promising because spammers can bypass them, but individual filters can adapt to each user's mail and make spammers' testing cycles inefficient.
npm audit: Broken by Design — overreacted
2 months ago
- npm audit is broken by design due to a high rate of false positives in front-end tooling contexts.
- Many reported vulnerabilities (e.g., regular expression DoS) are irrelevant for static site builds, where user input cannot trigger them.
- The default `npm audit` after every install overwhelms beginners and wastes maintainer effort on non-issues.
- Proposed solutions include moving dependencies to devDependencies (flawed), bundling dependencies, or allowing maintainers to counter-claim reports.
- The author plans to close false-positive GitHub issues and calls for npm to prioritize fixing the system.
Testing Gemini 3.5 Flash Lite for human detection in home surveillance
2 months ago
- The author built a custom home surveillance system using Frigate for person detection, funneling to Casa Segura for finer detection.
- False positives from dogs being detected as persons are the main issue, with the goal of reducing alarms at night.
- Gemini 3.5 Flash Lite was tested against Gemini 3.1 Flash Lite and Gemini 3.1 Pro on a benchmark of 18 clips (10 without humans, 8 with humans).
- Gemini 3.5 Flash Lite produced one false positive (confidence 0.90) on a dark dog far away at night, while Gemini 3.1 Flash Lite and 3.1 Pro had zero false positives.
- Pricing: Gemini 3.5 Flash Lite is 50% more expensive than 3.1 Flash Lite ($0.30/$2.50 per M vs $0.25/$1.50), while Gemini 3.1 Pro is about 8x more expensive.
- Latency: Gemini 3.1 Flash Lite is slightly faster (~2.9s median) than 3.5 (~3s), while Pro is much slower (~10s).
- The conclusion is that the new model is marginally worse, and false positives can be reduced by re-querying the same model rather than switching models.
I wouldn't say Pangram is broken, but I would say that it's brittle
2 months ago
- Pangram produces contradictory results: the same text can be flagged as 100% AI or 100% human depending on context, creating a 'Russian doll' of inconsistent outcomes.
- The percentage meter is misleading because many users interpret '100% AI' as high confidence, but it actually indicates the portion of text suspected to be AI-generated; the meter also frequently outputs 100% for hybrid texts.
- False positives are easy to induce: the author wrote his own text that Pangram flagged as 100% AI with high confidence, and AI-generated text can be made to appear human by embedding it in human writing.
- Even a low error rate (e.g., 1 in 10,000) leads to many false accusations when applied to large volumes of text, especially every paragraph becomes a separate test.
- The tool's design encourages a culture of 'gotcha' accusations among writers, exploiting professional jealousy and resentment rather than providing reliable evidence.
- Pangram's founder acknowledged the issue with short text segments and the inevitability of adversarial false positives, but the author argues for epistemological modesty and warns against treating any detection tool as a final authority.
Scanning for Pangram Errors
2 months ago
- Pangram is an AI detection tool with a claimed false positive rate of 1 in 10,000.
- The author tested Pangram on 2.9 million words from 45 old books that were never digitized, to check for memorization of training data.
- Pangram did not flag any human-written content as AI (false positives) except for segments where OCR (Mistral) hallucinated text, which Pangram correctly identified as AI-generated.
- The results suggest Pangram is not simply memorizing its training set but is detecting AI writing patterns.
- Human review is recommended when using Pangram for important decisions, as the tool itself is not infallible.

false positives

Expansion of Flock Camera Systems Is Leading to More False Positives

Better Bayesian Filtering

npm audit: Broken by Design — overreacted

Testing Gemini 3.5 Flash Lite for human detection in home surveillance

I wouldn't say Pangram is broken, but I would say that it's brittle

Scanning for Pangram Errors