By the TensorMax editorial team
· Drawing from sources across the AI industry
Today's top story
research paper
A recent study by Kapwing analyzed 10,742 TikTok videos across 20 popular categories and found that AI-generated content dominates the For You page before the algorithm learns who the user is.
Why it matters. The strategic stake of this signal lies in the fact that 59% of TikTok videos shown to new users are AI-generated, which is roughly three times the proportion found on YouTube. This high percentage of AI-generated content, particularly in categories such as children's content, science, and education, raises concerns about the accuracy and quality of information presented on the platform. With nearly six out of every ten videos being AI-generated, TikTok's default experience is dramatically worse for new users, which could have significant implications for the platform's growth and user engagement.
A recent study by Kapwing analyzed 10,742 TikTok videos across 20 popular categories and found that AI-generated content dominates the For You page before the algorithm learns who the user is. The study revealed that 59% of videos shown to new accounts are AI-generated, with children's content being the worst-hit category at 57%. The numbers are stark, with one hashtag, #CartoonKids, having 97 out of 100 videos checked being AI-generated. The concern is not just that individual videos are of poor quality, but that generative AI enables the creation of endless streams of them at a pace no human creator could match. This phenomenon has been described as 'toddler AI misinformation at an industrial scale' by Dr. Dana Suskind, a professor of pediatrics at the University of Chicago. The problem extends beyond content aimed at children, with educational, science, health, and history videos also being heavily saturated with AI-generated material. TikTok's recommendation engine is designed to adapt quickly, but the study focuses on what happens before personalization kicks in, when a new account has provided no behavioral data and the algorithm is essentially guessing. The result is that AI-generated content has become TikTok's default first impression, which is a significant problem for a platform that built its growth on the strength of its recommendation algorithm. The company has introduced controls to allow users to increase or decrease the amount of AI-generated content in their feeds, but the data suggests these measures have not meaningfully reduced the volume of AI slop reaching new users. The platform also faces growing legal pressure over its handling of children's content, with Florida suing TikTok under its child social media law. In comparison, YouTube has taken a more aggressive enforcement approach, terminating channels with inauthentic content, and has a lower rate of AI-generated content at 21%. The study's findings have significant implications for the future of social media and the role of AI-generated content on these platforms.
More from today
model release
Why it matters. The release of z.AI's GLM-5.2, a new open-source Chinese AI model, has significant implications for the AI industry, with a 1 million token context window that puts it in the same league as Anthropic's Claude Opus 4.8 and OpenAI's GPT 5.5. This development has the potential to disrupt the market, particularly if the open-source model can capture a larger share of the market, which could challenge the dominance of American companies like OpenAI and Anthropic. With China pushing forward with cheaper and increasingly capable open-source models, the US may struggle to preserve its edge in AI supremacy, especially given the warnings from Anthropic about China's rapid progress.
research paper
Why it matters. The introduction of Tri-Info, a generalizable and interpretable failure prediction method for Vision-Language-Action models, has significant strategic implications as it achieves 83% accuracy on real-world tasks, outperforming prior detectors that collapse to chance. This breakthrough is crucial for the safe deployment of VLA models across diverse tasks, where physical interactions can cause irreversible harm. By providing a simple yet powerful method for detecting failures with strong cross-domain generalization, Tri-Info addresses a critical need in the industry.
model release
Why it matters. The release of Cognee, an open-source AI memory platform, has significant implications for the AI industry, with 18.7k stars on GitHub. This platform provides AI agents with persistent long-term memory across sessions, enabling them to recall, connect, and act with full context. By giving AI agents the ability to learn from feedback and share knowledge, Cognee has the potential to revolutionize the way AI systems operate and interact with users.
research paper
Why it matters. The introduction of FloatDoor, a novel platform-triggered backdoor attack against generative large language models, poses a significant threat to the security of LLM deployments, with 100% of models potentially compromised. As demonstrated on Qwen3-4B across various deployment targets, including NVIDIA GPUs, Google TPUs, AWS Graviton, and Alibaba Yitian-710, FloatDoor can reliably induce exploitable code vulnerabilities, underscoring the need for trusted model supply chains in sensitive applications.
model release
Why it matters. The release of Headroom, an open-source context compression library, has significant implications for AI operators, as it can reduce the number of tokens sent to language models by 60-95%. This can lead to substantial cost savings, as users pay for every token the model writes back, with output costs being 5 times the input cost on Opus-class models. With Headroom, users can achieve the same accuracy while reducing the number of tokens, resulting in lower costs.
regulatory action
$8.7B
Why it matters. The European Commission's Tech Sovereignty Package aims to reduce the continent's reliance on foreign technology providers, with a focus on building domestic AI infrastructure. This push for sovereignty is driven by the fact that US hyperscalers currently control 70% of the European market, and stakeholders in heavily regulated industries such as government, finance, and healthcare are seeking to reduce their dependence on foreign infrastructure. With €1.9bn already raised by AI hardware and compute startups in Europe this year, the region is poised to make significant strides in developing its own domestic hardware stack, with the UK government alone investing £1.1bn to boost its AI infrastructure capabilities.
Catch up quick
-
Samsung Electronics deploys ChatGPT Enterprise and Codex to all staff in Korea and DX division staff globally
-
Toto, Japan's largest toilet maker, plans to invest $495M by 2030 to expand its semiconductor materials unit
-
SK Hynix passes Samsung to become South Korea's most valuable company, with a market capitalization of ~$1.4T
-
Anthropic and OpenAI take first steps toward IPOs, with both companies confidentially filing draft S-1 forms with the SEC.
-
OpenAI releases GPT-3, a highly capable Large Language Model, in June 2020, while Anthropic later releases its chatbot, Claude, and OpenAI releases ChatGPT.
-
Dario Amodei leaves OpenAI in December 2020 to found Anthropic, taking several other former OpenAI employees with him.
-
Nuvei acquires Payoneer for $2.75 billion
-
China adds ten US companies to export controls
-
Five Eyes leaders warn AI models capable of taking down governments and businesses are mere months away, urging leaders to act now
-
Five Eyes intelligence alliance warns of imminent threat from frontier AI models in offensive cyber operations
-
Huazhong University of Science and Technology and VIVO AI Lab release Moebius, a 0.2B image inpainting model with 10B-level performance
-
Zhipu AI's market capitalization surpasses HK$1 trillion as shares soar 42% after releasing GLM-5.2 model
-
Researchers from SPAR find that finetuning-based auditing methods for model organisms may unintentionally erase the deceptive behavior they are trying to measure, leading to an illusion of success.
-
Spanish mobile phone users keep their devices for longer, with 37.5% using them for five years or more
-
Qnity stock rises 2% as company benefits from shift in semiconductor design
Also on the desk
What landed, what bounced.Hit reply with one story from today that landed for you and one that bounced. Two sentences is plenty.
|
|