Published on

This Week in AI: Deepfakes, Disinformation, and New AI Models

Authors
  • Name
    Twitter

Overview

This week in AI saw concerning developments in deepfake technology, new AI models in healthcare, and ongoing debates about AI regulation and capabilities. Key stories include the ease of creating convincing deepfakes, new medical AI models, and research questioning AI's understanding of the world.

The Growing Threat of AI Deepfakes

Recent developments have highlighted the increasing ease and sophistication of creating AI-generated deepfakes, raising alarm about potential misuse for disinformation:

  • A journalist was able to create a convincing audio deepfake of Vice President Kamala Harris for just $5 in less than two minutes using Cartesia's Voice Changer tool.
  • The volume of AI-generated deepfakes grew 900% between 2019 and 2020 according to the World Economic Forum.
  • Experts warn that disinformation is spreading at an alarming rate, with many high-profile examples targeting elections and political figures.

This demonstrates how cheap, ubiquitous generative AI has opened the floodgates to potential misuse. While some propose solutions like invisible watermarks or content moderation laws, many experts believe those ships have sailed. The best defense may be intense skepticism of viral online content.

New AI Models in Healthcare

On a more positive note, new AI models show promise for transforming medical research and practice:

  • BiomedGPT, a new open-source AI model described in Nature Medicine, combines vision and language AI to tackle a wide range of biomedical tasks. It achieved state-of-the-art results on 16 out of 25 datasets tested.
  • The model could help doctors interpret medical images, assist researchers in analyzing scientific literature, and even aid in drug discovery.
  • Clinical testing at Massachusetts General Hospital demonstrated BiomedGPT's strong performance on tasks like visual question answering and radiology report generation.

This "generalist" AI model for biomedicine showcases how AI can be adapted to various medical tasks with minimal additional training. While still in early stages, it highlights AI's potential to streamline healthcare and research.

Tech Giants Expand AI Features

Apple's iOS 18.2 Beta

Apple is rolling out new AI features in its iOS 18.2 beta, including:

  • Genmoji: An AI emoji generator
  • Image Playground: AI-powered image creation
  • ChatGPT integration with Siri (finally Siri gets an update)
  • Visual search using iPhone 16 cameras

While some features require joining a waitlist, this update marks a significant step in Apple's AI strategy.

Microsoft's AI Innovations for Windows

Microsoft is introducing new AI experiences for Paint and Notepad:

  • Generative fill and erase in Paint
  • AI-powered text rewriting in Notepad
  • Improved Cocreator and Image Creator features

These updates aim to enhance creativity and productivity for Windows users.

Debate Over AI Capabilities and Regulation

The rapid advancement of AI continues to spark debate over the technology's true capabilities and how it should be regulated:

  • A new MIT study found that while generative AI models can perform impressive tasks, they may lack a coherent understanding of the world. For example, an AI could give accurate driving directions in New York City without having an accurate internal map of the city.
  • When researchers added detours, the performance plummeted revealing limitations in its world model.
  • President-elect Donald Trump's victory may mean less regulation is coming potentially allowing for more open innovation but also raising concerns about unchecked development.
  • Meta allows its AI models to be used for U.S. military purposes, shifting its policy to support national security efforts.
  • OpenAI defeats a copyright lawsuit from news outlets over AI training data, though the case may be refiled.
  • The Canadian Legal Information Institute sues an AI chatbot company over alleged copyright infringement and violation of terms of service. These developments highlight ongoing uncertainty around what current abilities truly are capable by this technology as well as its future trajectory. While showing remarkable potential across many domains caution is advised against assuming too much about what it truly "understands."

This week's news underscores both exciting potentials as well serious risks posed by rapidly advancing technologies from concerns over deepfake capabilities promising medical advancements debates regarding true capabilities all evolving at breakneck speeds becoming increasingly integrated into our lives making careful consideration impacts limitations crucial going forward ensuring responsible usage remains paramount importance moving forward together responsibly harnessing power while mitigating associated risks involved.

References

TechCrunch - This Week in Ai It Shockingly Easy Make Kamala Harris Deepfake

News Medical - New Ai Model Biomedgpt Set Transform Medical Research Practices

MIT News - Despite Its Impressive Output Generative Ai Doesn't Have Coherent Understanding World