The Quest to Detect AI-Generated Content- H+ Weekly - Issue #430
The Quest to Detect AI-Generated Content- H+ Weekly - Issue #430
This week - OpenAI on track to $1B in revenue this year; Google's secret weapon in generative AI race; a new real-life mecha from Japan; cancer-detecting bacteria; and more!
Content generated by AI is of such high quality that it is nearly impossible for humans to distinguish between a work of art or text created by a human versus one generated by an AI. This impressive engineering achievement, however, comes with its dark side. Generative AI can be used to spread misinformation and deepfakes, making the task of distinguishing between reality and fabrication extremely challenging.
Lawmakers are well aware of this risk and are in the process of crafting laws that will require AI companies to label AI-generated content as such. The upcoming EU AI Act will mandate that all AI-generated content be appropriately labelled. Similarly, rules proposed by the Cyberspace Administration of China also incorporate a requirement for labelling AI-generated content. In the United States, although the regulations concerning AI have yet to be finalized, major AI companies such as OpenAI, Alphabet, Meta, Microsoft, Anthropic, and Inflection pledged to label AI-generated content as part of an effort to ensure the safety of this technology.
The fundamental concept behind watermarking AI-generated content, whether it be text, images, or audio, involves embedding imperceptible patterns into content that are easily detectable by computers but go unnoticed by humans. In the case of images, the watermark is directly integrated into the image's pixels. For text, a specific pattern of words or word frequency can reveal the text's AI origin. In audio, this might involve sounds beyond our audible range (20-20k Hz).
Numerous companies have joined forces under the Coalition for Content Provenance and Authenticity (C2PA) to establish technical standards for labelling and identifying AI-generated content.
During the Google I/O 2023 conference, Google promised to release tools to verify the source of an image. This week, Google DeepMind released SynthID, a tool for watermarking and identifying AI-generated images. SynthID is currently only available to a limited number of Vertex AI customers.
Watermarks for AI-generated audio are already being included in content generated by Resemble AI and Azure’s Neural Text-to-Speech services.
Text, however, is proving to be a bit more challenging to watermark. At the beginning of this year, OpenAI released an AI classifier detecting if the text was generated by a machine. However, six months later, the project was cancelled “due to its low rate of accuracy”. OpenAI is “currently researching more effective provenance techniques for text, and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI-generated.”
OpenAI, Google and other companies can add watermarks because they control the AI models generating the content. However, there are existing models, such as open-source models or models built from scratch, that AI companies do not have complete control over. The tool to detect if a piece of content was AI-generated or not will need to take into account the existence of such models and ideally work effectively with them as well. This presents a more complex challenge, as there are no obvious watermarks to look for. The detectors will need to identify even subtler patterns that give away the content as AI-generated.
The techniques for watermarking will inevitably need to evolve and adapt. As soon as a new method for detecting AI-generated content is introduced, it's only a matter of time until someone finds a way to fool the detectors and a new method will be needed. We are going to be locked in an ongoing arms race between those who generate fake content and those who can detect it.
Becoming a paid subscriber now is the best way to support the newsletter.
If you enjoy and find value in my writing, please hit the like button and share your thoughts in the comments. Share the newsletter with someone who will enjoy it, too.
🦾 More than a human
Naked mole rat ‘longevity gene’ lets mice live longer
Researchers have successfully transferred a longevity gene from naked mole rats to mice, resulting in improved health and an extension of the mouse’s lifespan. Mice that received the “longevity gene” have seen an approximate 4.4% increase in median lifespan. These findings open new possibilities for exploring how gene therapies could also be used to improve lifespan and reduce inflammation-related diseases in humans.
▶️ The most exciting real anti-aging drugs (12:30)
In this video, Andrew Steele goes through five drugs, rapamycin, senolytics, metformin, taurine and statins, all having evidence of anti-ageing effects, and explains how they work and how they can be used to make people live longer and healthier lifes.
A biotech company says it put dopamine-making cells into people’s brains
A biotech company has implanted lab-grown neurons into the brains of 12 people with Parkinson’s disease. The neurons, made with embryonic-stem-cell technology (the controversial and much-hyped approach of using stem cells taken from IVF embryos), have safely incorporated themselves into the patients’ brains and were reducing patients’ symptoms a year after the treatment.
🧠 Artificial Intelligence
OpenAI on track to generate more than $1B revenue over 12 months
According to a report published by The Information, OpenAI has made over $1B in revenue this year - 5x more than the company projected to bring in revenue for this year. The Microsoft-backed company is generating more than $80 million in revenue per month, compared to just $28 million in the entire last year, the report added.
Google Gemini Eats The World – Gemini Smashes GPT-4 By 5X, The GPU-Poors
In this post,
U.S. Bans Sales of Nvidia's H100, A100 GPUs to Middle East
Nvidia reported in a recent regulatory filing that the US government has imposed limitations on the sale of its powerful GPUs, H100 and A100, in various nations, including the Middle East. Both Saudi Arabia and the United Arab Emirates have been strengthening their AI prowess in recent years, which is why they are significant purchasers of Nvidia's chips. Saudi Arabia and the UAE have been in talks with China to deepen their collaborations and the ban is aimed at preventing any H100s and A100s from getting into China or being used by Chinese companies through their Middle East partners.
Most Americans haven’t used ChatGPT; few think it will have a major impact on their job
Pew Research Center published a report examining how Americans use ChatGPT and there are some quite interesting things there. According to the report, 18% of US adults have used ChatGPT (which aligns with the Morgan Stanley report released in June). The most popular use case is entertainment (20%) followed by education (19%) and work (16%). Researchers also asked if chatbots would be helpful in their jobs and only 15% of working adults who have heard of ChatGPT say chatbots would be extremely or very helpful. When asked about the US government’s approach to regulating AI, 67% of those who have heard of ChatGPT said that the US government will not go far enough in regulating chatbot use.
4 Charts That Show Why AI Progress Is Unlikely to Slow Down
Time Magazine shares four charts showing how quickly AI is progressing, including a chart showing progress in AI surpassing humans at different tasks (seen above), the growth of computing power (which is doubling every 6 months starting from 2010), the growth in data used in training AI models and growing effectiveness of AI algorithms in terms of computing power required to train the algorithm.
🤖 Robotics
US military plans to unleash thousands of autonomous war robots over next two years
The US military plans to deploy thousands of autonomous weapon systems within the next couple of years as a response to China's increasing influence. The Replicator project, as it is known, seeks collaboration with defence and technology firms to manufacture cost-effective systems in large quantities for all branches of the US military.
Arcax - a new mecha from Japan
Meet Arcax, Japan’s newest mecha. It stands 3.9m tall in vehicle mode and 4.5m in robot mode, weighs 3.9t and can move at a speed of 2km/h in robot mode and 10km/h in vehicle mode. And if you want one, you can buy it (as long as you live in Japan). The price is 400 million yen (about $2.7 million) and the production is limited to 5 units which the company behind the mecha promises to deliver within 12 to 18 months. This is not the first real mecha coming out of Japan. 11 years ago, Kuratas, the world’s first real mecha robot, went viral and later fought an American mecha in the world’s first giant robot’s fight.
🧬 Biotechnology
World-first experiment shows genetically engineered bacteria detecting cancer
Researchers have turned a bacteria into a disease-detecting biosensor. Using what is called a competent bacteria (a bacteria that can integrate cell-free DNA into its own genome), researchers have proven it is possible to engineer a bacteria that can take in cancer DNA and therefore become a biosensor detecting cancer. Initial experiments, including experiments in live mice, have confirmed the viability of this method.
H+ Weekly sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.
A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!
You can follow H+ Weekly on Twitter and on LinkedIn.
Thank you for reading and see you next Friday!