• Jumble
  • Posts
  • FDA to Roll Out AI Co‐Pilot Across Entire Agency

FDA to Roll Out AI Co‐Pilot Across Entire Agency

This week, the U.S. FDA is sprinting toward a June deadline to embed a generative‑AI Co‑Pilot into every corner of the agency, while researchers say one selfie could soon help oncologists forecast cancer survival. Let’s dive in ⬇️

In today’s newsletter:
🧠 FDA to roll out an AI co-pilot across the entire agency
✝️ Selfie app can predict cancer survival rates
🔐 Whitehouse does a reverse on AI-chip export ban
🏦 Google invests in AI Futures Fund for entreprenuers
💡 AI Challenge: Use AI to make language learning easier

🧬 How a Government LLM Aims to Reshape Regulation

On May 10 the U.S. Food and Drug Administration confirmed that every center and regional office will get an “AI Co‑Pilot” by June 30, its first agency‑wide deployment of a generative‑AI tool. The pilot model, fine‑tuned in a secure Microsoft Azure enclave, has already helped reviewers summarize oncology trial dossiers and draft feedback letters, shrinking tasks that once took days to just a few minutes.

🎯 Why regulators are excited

Newly minted commissioner, Dr. Martin Makary, also says the Co‑Pilot will free scientists to “think, not just search.” Early tests showed a 40 % reduction in clerical load and improved consistency in safety‑signal detection. Proponents also point to faster recalls for contaminated foods and quicker approvals for critical rare‑disease drugs, arguing that patients will feel the impact long before the fiscal year ends.

❗ Caveats, gaps & big questions

Not everyone is convinced. Axios obtained internal emails noting the model occasionally hallucinated chemical names and over‑confidently misstated dosage guidelines—errors that could be life‑threatening if un‑caught.

Privacy advocates want clarity on how sensitive trade secrets are protected, and lawmakers from both parties have demanded an external audit before the tool touches high‑risk biologic applications. In addition, smaller biotech firms worry the system may implicitly favor data formats used by Big Pharma, widening an existing compliance gap.

Listen to the full interview with Dr. Makary below:

🔮 What happens next

The FDA says human reviewers will remain “in the loop” for all final decisions and will be trained to challenge the model’s output. A public dashboard, slated for release in August, will publish usage metrics and flagged discrepancies.

If the experiment succeeds, international regulators from the EU and Japan have signaled interest in sharing best practices through the ICH forum. If it fails, Congress could respond with new limits on federal AI, affecting agencies far beyond health. If the experiment scales, it could dramatically redefine how global regulators leverage data and improve patient access worldwide.

🤳 Can a Selfie Predict Cancer Survival?

Take a selfie, get a prognosis. That is the promise of “FaceAge,” a deep‑learning tool unveiled last week that predicts biological age and, with it, cancer survival odds. Developed by researchers at the University of Pennsylvania with partners in Canada and India, the model was trained on more than fifty‑thousand clinical portraits of head‑and‑neck cancer patients and validated against five‑year survival outcomes.

🧠 How it works

FaceAge uses a convolutional neural network to score subtle facial cues—wrinkles, eyelid droop, dermal sag—as proxies for systemic aging. The output, a "face‑derived age," is combined with tumor stage to generate a survival curve. In blinded tests the tool outperformed traditional ECOG performance scores and even clinicians’ gestalt impressions, offering a quick, non‑invasive risk assessment that takes seconds, not hours.

⚖️ Promise and peril

Supporters say the model could help oncologists triage scarce resources, flag frail patients for supportive care earlier, and enroll under‑represented groups in precision trials. Because the input is just a photo, it may bring equitable screening to rural clinics that lack imaging suites.

🕵️ Ethical and technical unknowns

Critics warn that the system has been studied mostly in South Asian populations and may underperform on other demographics. They also question whether visible aging correlates with cancers not examined in the paper, such as leukemia. Privacy advocates note that facial data is biometric and, in many jurisdictions, legally sensitive. Regulators will need to weigh benefit against the risk of misuse—think insurers or employers using photos to infer disease burden.

🔭 Looking ahead

The research team plans a multi‑center trial across five continents and is open‑sourcing the code to speed peer replication. The real‑world dream is a phone‑based screening app that integrates with electronic medical records, pops up survival estimates during tele‑oncology visits, and nudges patients toward lifestyle interventions.

Still, without large, diverse datasets and clear guardrails the method could amplify biases in care. The selfie revolution in oncology is coming—but whether it empowers or marginalizes patients will depend on how we build it. Future updates will include pediatric dermatologic cancer cohorts.

This Week’s Scoop 🍦

🔮 AI Challenge: Use AI to Make Language Learning Easier

Challenge: Use Google’s ‘Little Language Lessons’ to help you learn or spruce up on a second language.

Google Labs’ Little Language Lessons hosts three interactive experiments—Tiny Lesson, Slang Hang, and Word Cam—to turbocharge your language learning. Here’s how to play:

  1. Visit labs.google/lll/en and select your target language.

  2. Tiny Lesson: Describe a real‑world scenario (e.g., “booking a taxi”) and note the suggested phrases. Practice each aloud and screenshot the results.

  3. Slang Hang: Engage in a casual chat simulation. Hover over slang terms to see definitions, then write three new idiomatic sentences using them.

  4. Word Cam: Snap a photo of an object in your environment. Record its name and two related words provided by the AI.

  5. Combine your Tiny Lesson phrases, Slang Hang sentences, and Word Cam vocabulary into a 4–5 sentence diary entry.

Learning a new language is one of the hardest things to do, but with these new AI tools, it’s easier than ever. Start checking that ‘Learn a new language’ box on your bucket list today!

Want to sponsor Jumble?

Click below ⬇️

That’s it for this week! Quantum breakthroughs in government and predictive selfies in oncology show AI’s reach from policy halls to personal health. Which story grabbed you most? Hit reply and let us know. See you next time! 🚀

Stay informed, stay curious, and stay ahead with Jumble!

Zoe from Jumble