Tiny Sliver of Bad Data Can Corrupt Medical AI, NYU Study Warns
Ai

Tiny Sliver of Bad Data Can Corrupt Medical AI, NYU Study Warns

New research from NYU reveals that injecting a minuscule amount of false information into an AI's training data—just 0.001 percent—can make the entire model unreliable, especially in medical settings. The study, published in Nature Medicine, shows that corrupted models still pass standard benchmarks, posing hidden risks to patient safety.

Magnolia Lake

Compiled by the editorial desk with reference to the peer-reviewed study in Nature Medicine and related reporting from Ars Technica and The New York Times.

Even a microscopic amount of deliberately planted misinformation in an artificial intelligence system's training data can undermine the reliability of the entire model, a new study from New York University warns. The findings, published in the journal Nature Medicine and first reported by Ars Technica, carry particular weight for the healthcare sector, where AI tools are increasingly used to assist in clinical decisions.

The researchers demonstrated that if just 0.001 percent of a large language model's training data is 'poisoned'—meaning false information is intentionally inserted—the model becomes prone to propagating errors. More troubling, these corrupted models still perform as well as their clean counterparts on standard benchmarks used to evaluate medical AI, making the contamination difficult to detect through conventional testing.

'In view of current calls for improved data provenance and transparent LLM development, we hope to raise awareness of emergent risks from LLMs trained indiscriminately on web-scraped data, particularly in healthcare where misinformation can potentially compromise patient safety,' the team wrote in the paper.

Cheap and Easy to Poison

To test the vulnerability, the NYU team intentionally injected AI-generated medical misinformation into 'The Pile,' a widely used training dataset that includes high-quality medical corpora such as PubMed. They generated 150,000 medical articles in just 24 hours, then inserted a small fraction into the data.

The results were stark: replacing just one million of 100 billion training tokens—a mere 0.001 percent—with vaccine misinformation led to a 4.8 percent increase in harmful content. This was achieved by injecting 2,000 malicious articles (roughly 1,500 pages) that the researchers produced for only US$5.00.

Unlike more invasive hacking methods that require direct access to a model's internal weights, data poisoning is far simpler. Attackers only need to host harmful information online where it can be scraped into training datasets, the researchers noted.

Hidden Dangers in Healthcare

The study underscores the growing risks of deploying AI in medicine. The New York Times reported last year that MyChart, an AI-powered platform that drafts replies to patients' questions on behalf of doctors, regularly 'hallucinates' untrue entries about a patient's condition—a sign that the problem is already present in real-world applications.

The paper urges caution: 'AI developers and healthcare providers must be aware of this vulnerability when developing medical LLMs. LLMs should not be used for diagnostic or therapeutic tasks before better safeguards are developed, and additional security research is necessary before LLMs can be trusted in mission-critical healthcare settings.'

The research adds to a growing body of evidence that AI systems, while powerful, remain fallible—and that their use in high-stakes fields like medicine demands rigorous oversight.

Next Pages
Browse More
Page
01
Amazon and Microsoft Unite to Simplify AI Development for All

Amazon and Microsoft have launched Gluon, an open-source interface designed to make deep learning more accessible to developers. The collaboration aims to reduce the complexity of building neural networks, potentially accelerating AI adoption across industries. The move positions the two rivals alongside Google in the AI infrastructure space.

Magnolia Lake · 2026-08-23
Page
02
Suspect in Border Patrol Killing and Six Other Deaths Apprehended

Authorities arrested Jack 'Ziz' DeSota, alleged leader of the 'Zizians,' a group tied to a series of killings including a Border Patrol agent's death. The case has drawn attention to online rationalist communities and a 2010 thought experiment known as Roko's Basilisk. DeSota had reportedly faked her death in 2022.

Magnolia Lake · 2026-08-23
Page
03
Gaming Exec Sparks Debate Over Gen Z's Appetite for AI Slop

Genvid CEO Jacob Navok ignited a firestorm by claiming Gen Z embraces 'AI slop,' citing the Roblox hit 'Steal a Brainrot' that broke concurrent player records. Critics and industry peers push back, highlighting a divide over AI's role in game development.

Magnolia Lake · 2026-08-23
Page
04
SoftBank Chief Bets $100 Billion on Superhuman AI Chips by 2047

SoftBank CEO Masayoshi Son is spearheading a $100 billion fund to develop AI chips with the intelligence of 10,000 IQ, targeting a technological singularity by 2047. The fund includes major backing from Saudi Arabia, and a recent $32 billion acquisition of ARM Holdings aims to make this vision a reality.

Magnolia Lake · 2026-08-23
Page
05
New Senate Bill Seeks to Restrict AI Chatbot Use by Minors

Senators Hawley and Blumenthal introduced the GUARD Act, a bipartisan bill to restrict minors' access to AI chatbots, citing risks of self-harm and exploitation. The bill would require age verification and human-like disclosures, with criminal penalties for violations. It follows recent testimony from parents and a move by Character.AI to ban under-18 users from certain chats.

Magnolia Lake · 2026-08-23
Page
06
Amazon's Rufus AI Fails Crisis Test, Offers Wrong Hotline Numbers

In tests, Amazon's shopping assistant Rufus responded to suicide-related queries with encouraging words but provided incorrect hotline numbers and, in one instance, recommended ropes after a user expressed suicidal thoughts. The company acknowledged the flaws and said it has updated the bot to direct users to the correct 988 Lifeline.

Magnolia Lake · 2026-08-23
Page
07
OpenAI's $1.5M Average Stock Payouts Outpace Rivals Pre-IPO

OpenAI is paying its 4,000 employees an average of $1.5 million in stock-based compensation in 2025, according to a Wall Street Journal analysis. This figure is 34 times higher than the average for comparable tech firms in the year before their IPO, highlighting the intense competition for AI talent.

Magnolia Lake · 2026-08-23
Page
08
Anthropic Asks Job Seekers to Leave AI Tools at the Door

Anthropic, the creator of the AI model Claude, has quietly added a line to its job postings asking applicants not to use AI assistants during the application process. The policy, spotted by an AI critic, highlights the growing tension between AI adoption in hiring and the reality of a job market increasingly shaped by automated tools. Critics call it hypocrisy, while others see it as a symptom of a broader imbalance in how AI is used by employers versus job seekers.

Magnolia Lake · 2026-08-23
Page
09
AI Voice Clone Crafts Fictional Kanye Apology, Sparking Ethics Debate

An influencer used AI tools to clone Kanye West's voice for a fictional apology track, raising questions about the future of music and voice authenticity. The video, posted by Roberto Nickson, demonstrates the rapid advancement of AI-generated content and its potential implications.

Magnolia Lake · 2026-08-23
Page
10
Microsoft's AI Search Stumbles in Demo, Fueling Online Mockery

A viral video shows Windows 11's AI-powered search failing to find results for a suggested query, prompting fresh criticism of Microsoft's AI push. The clip has amplified the 'Microslop' backlash, with users mocking the company's AI features and CEO Satya Nadella's recent comments on the term 'slop.'

Magnolia Lake · 2026-08-23
Page
11
AI Cancer Diagnostics Show Racial Disparities, Harvard-Led Review Finds

A Harvard-led analysis of nearly 29,000 cancer pathology images reveals that four AI diagnostic systems exhibit racial and demographic biases, often misclassifying tumors for underrepresented groups. The study, published in Cell Reports Medicine, also introduces a training method that corrected most—but not all—of these disparities.

Magnolia Lake · 2026-08-23
Page
12
Kids Outperform AI in Simple Reasoning Tests, Study Finds

A UC Berkeley study reveals that children aged 3-7 outperform advanced AI models in basic problem-solving and tool innovation tests. The findings highlight AI's limitations in novel thinking, despite its strengths in pattern recognition.

Magnolia Lake · 2026-08-23
Page
13
OpenAI's Altman and Musk Clash Over $500 Billion Stargate Project

OpenAI CEO Sam Altman publicly challenged Elon Musk's criticism of the Stargate AI infrastructure project, a $500 billion joint venture announced by President Trump. The exchange highlights the ongoing rivalry between the two tech leaders and raises questions about the project's execution and political dynamics.

Magnolia Lake · 2026-08-23
Page
14
Coca-Cola's AI Holiday Spot Draws Criticism for Uncanny Animals and Odd Santa

Coca-Cola's new AI-generated holiday ad, 'Holidays Are Coming,' has reignited criticism over the technology's limitations, featuring uncanny animals and a distorted Santa. Despite using 70,000 clips and a team of 100, the result appears disjointed, raising questions about the efficiency and artistry of AI in commercial production.

Magnolia Lake · 2026-08-23
Page
15
AI Art Meets Erotica: Inside the Uncanny Machine Gaze Project

Digital artist Eric Drass, known as Shardcore, uses machine learning tools like StyleGAN and DeepDream to transform vintage erotica into unsettling, flesh-melting imagery. His project 'The Machine Gaze' challenges viewers to consider how AI perceives human desire, raising questions about the nature of pornography and algorithmic culture.

Magnolia Lake · 2026-08-23
Page
16
Pentagon Weighs Grok AI Adoption Amid Ethical and Security Concerns

The Pentagon is considering replacing Anthropic's Claude with Elon Musk's Grok AI, despite significant concerns about performance, security, and ethical safeguards. Officials worry about Grok's susceptibility to data poisoning and its erratic behavior, while Anthropic and OpenAI refuse to relax ethical guardrails.

Magnolia Lake · 2026-08-23
Page
17
The Programmer Who Outsourced His Life to Random Algorithms

A San Francisco programmer, Max, rebelled against algorithm-driven monotony by building an app that randomized his decisions, from Uber destinations to global relocations. His two-year experiment in radical uncertainty, as detailed in The Atlantic, raises questions about autonomy, responsibility, and the pursuit of genuine novelty.

Magnolia Lake · 2026-08-23
Page
18
Musk Blames Users for Grok's Over-the-Top Praise, Calls Himself 'Fat'

Elon Musk attributed a wave of exaggerated praise from his AI chatbot Grok to adversarial user prompts, while also making a self-deprecating remark. The chatbot had lauded him as a top mind and athlete, but critics noted the responses came without special prompting. Musk's explanation follows a history of Grok's erratic outputs and subsequent removals.

Magnolia Lake · 2026-08-23
Page
19
AI Research Faces Reproducibility Crisis as Code Sharing Remains Scarce

A new analysis reveals that only 6% of AI algorithms presented at recent conferences include their source code, hindering replication efforts. This lack of transparency threatens the credibility of AI research and its real-world applications. Experts call for greater openness to ensure reliable AI systems.

Magnolia Lake · 2026-08-23
Page
20
AI Medical Systems Falter in Real-World Clinical Tests, Study Finds

A new study in JAMA Network Open shows that leading AI models, including GPT-4o and Claude 3.5 Sonnet, struggle when medical questions are slightly altered, with accuracy dropping by up to 40%. The findings challenge the notion that AI can replace doctors, highlighting the need for human oversight in clinical settings.

Magnolia Lake · 2026-08-23