Why AI Models Fabricate Citations and Why Human Review Alone Cannot Catch Them

·2 views

Large language models generate bibliographic references by assembling high-probability components — author names, journal titles, years, and identifiers — using the same pattern-completion process they apply to any text, with no internal mechanism to distinguish recall from fabrication. Because fabricated citations are structurally identical to real ones, they pass visual inspection by even domain-expert reviewers; only resolving the identifier against an actual database can confirm validity. Since 2023, lawyers in multiple jurisdictions have submitted court filings containing non-existent case citations generated by AI chatbots, prompting judicial sanctions and professional-conduct scrutiny. The problem extends beyond legal filings to medical literature, academic drafting, and technical documentation, wherever structured references with strong internal regularity appear. A 2023 study by Liu, Zhang, and Liang evaluating commercial generative search engines found that roughly half of AI-generated sentences were fully supported by their cited sources, underscoring that citation presence does not equal citation accuracy.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Feature Engineering Remains the Highest-Leverage Skill in Machine Learning

Feature engineering is the process of re-describing input data so that a model can access patterns it otherwise could not express. Every model class has inherent limitations in the functions it can represent, and no amount of additional data or hyperparameter tuning can compensate when the true relationship lies outside that boundary. A classic example is the XOR problem, which is unsolvable by any linear model but becomes trivially solvable by adding a single engineered interaction column. Practical techniques such as ratio features, cyclic time encodings, and entity-level aggregates consistently outperform raw inputs, especially in fraud detection and churn prediction tasks. The article argues that these transformations do not add new information in a strict sense, but rather reshape existing information into a form the model can actually use.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build a FastAPI Endpoint That Streams LLM Responses to the Browser

A technical guide demonstrates how to build a streaming chat endpoint using FastAPI and httpx that delivers large language model responses token-by-token to the browser. The setup relies on an async generator as the response body and a shared httpx client initialized at app startup to preserve connection pooling and avoid repeated TLS handshakes. Server-Sent Events are structured with three named event types — token, error, and done — to help clients cleanly distinguish between streamed content, failures, and completion signals. Key headers such as X-Accel-Buffering and Cache-Control are set to prevent intermediate proxies or servers from buffering the stream. Since the browser's native EventSource API only supports GET requests, real-world chat UIs are advised to use the Fetch API with a stream reader for POST-based endpoints.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

NemynAI SEO Autopilot Brings Claude-Powered Content Generation to WordPress

NemynAI SEO Autopilot is a WordPress plugin developed by the Ukraine-based NemynAI team that automates SEO article generation and publishing by integrating directly with Anthropic's Claude API using the user's own API key. Unlike most AI content tools, it avoids routing data through a third-party backend, meaning prompts and content go straight to Anthropic under the user's own account. The free tier offers functional article generation, while a one-time $22 Pro license unlocks bulk CSV import, scheduled autopublishing via WP Cron, and social meta tag support. Recent changelog entries show the plugin has undergone iterative compliance fixes in response to WordPress.org's review process, including proper output escaping and SQL query preparation. However, the plugin currently has fewer than 10 active installations, so large-scale real-world testing remains limited.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Models Refuse Legitimate Requests and How Engineers Can Fix It

AI safety training sometimes blocks valid professional requests from security engineers, nurses, and novelists by flagging surface-level language features rather than actual intent, a phenomenon researchers call exaggerated safety. The model generalises its learned boundaries using cheap lexical cues — specific words, topics, or sentence structures — causing it to refuse benign prompts that superficially resemble harmful ones. Benchmark tools like XSTest and OR-Bench have been developed to measure this problem at scale, with XSTest using 250 carefully paired safe and unsafe prompts across ten categories. Researchers found that widely used chat models refused a significant portion of clearly safe prompts, with refusal rates varying considerably across model families. Practitioners are advised to separately track policy refusals from capability gaps in their telemetry and to measure where their specific model falls on the compliance-versus-refusal tradeoff curve for their domain.

0 comments Read more at DEV Community