UNRELIABLE COPILOT.
Less than a year before a critical U.S. election, Microsoft’s AI chatbot, now known as Microsoft Copilot, is disseminating conspiracies and incorrect information about political matters. WIRED’s inquiries on the 2024 U.S. election received misleading responses from the bot, including outdated or incorrect candidate information and links to debunked election conspiracy theories. The bot also referenced irrelevant international political events and shared links to far-right groups when asked about “election integrity” Telegram channels. Research by AI Forensics and AlgorithmWatch reveals that Copilot’s election misinformation is a widespread problem. The chatbot consistently provided inaccurate election information in Switzerland and Germany, including wrong polling numbers and election dates. Microsoft has acknowledged the issue and outlined plans to combat disinformation, but the problem persists globally. The bot’s inaccuracies extend to multiple languages, showing a significant discrepancy in accuracy between English and non-English responses. Researchers also noted Copilot’s inconsistency in answering repeated questions, often providing different and incorrect answers. Despite some improvements by Microsoft, the chatbot continues to generate misinformation, posing a threat to informed voting and democratic processes. (Wired)
FACIAL RECOGNITION TROUBLES.
Rite Aid faced a settlement with the Federal Trade Commission (FTC) for using facial recognition technology in its stores without customer consent, often misidentifying people, especially women and individuals of Black, Latino, or Asian descent. The system, deployed from 2012 to 2020 in several U.S. cities, alerted employees about shoppers it had previously tagged as potential shoplifters. As a result of the settlement, Rite Aid is prohibited from using facial recognition technology for five years. The technology led to increased surveillance, wrongful accusations, and store bans for misidentified individuals. Rite Aid’s database, created with two private vendors, included low-quality images from various sources. The FTC criticized Rite Aid for not ensuring the technology’s accuracy or adequately training employees in its use. Rite Aid disagreed with the allegations but stopped using the technology three years ago. This case highlights broader concerns about facial recognition’s accuracy and its impact on civil liberties. The FTC’s crackdown signifies a move towards regulating such technologies, especially as Rite Aid navigates bankruptcy proceedings. (The Guardian)
If you’re enjoying the Daily Dose, sign up for the Daily Dose Newsletter and get the morning’s best science news from around the web delivered straight to your inbox? It’s easy like Sunday morning.
AI-IMAGE-GENERATORS HARMING CHILDREN.
The Stanford Internet Observatory report unveiled a critical issue in AI image-generators: the presence of thousands of child sexual abuse images in training datasets, notably in the LAION database. This flaw has enabled AI to generate explicit content, including manipulated images of children. The discovery of over 3,200 such images in LAION, a key resource for AI development, prompted its temporary shutdown. The issue is linked to the combination of adult and children’s images in AI training. LAION’s use of these images, a fraction of its 5.8 billion-image database, significantly influences AI outputs and victimizes real children. The rapid market release of AI tools without adequate safety checks is a primary concern. The report calls for the removal of harmful training sets and stricter controls on AI models. Efforts to mitigate risks and ensure responsible AI use are ongoing among tech companies and child safety groups. (Associated Press)
SSH SECURITY IN PERIL.
In 1995, after a password sniffer compromised Helsinki University of Technology’s network, Tatu Ylönen created the Secure Shell Protocol (SSH) for secure network communication. Recently, researchers uncovered a vulnerability, named Terrapin, posing a significant threat to SSH’s cryptographic defenses. Terrapin is effective in a man-in-the-middle scenario, challenging the previously assumed security of SSH. This vulnerability targets specific SSH cipher modes, particularly ChaCha20-Poly1305 and CBC with Encrypt-then-MAC, and can manipulate the SSH handshake process, potentially compromising the integrity and confidentiality of SSH sessions. Affected by this vulnerability are a substantial number of SSH servers. To mitigate this risk, SSH developers, including those of OpenSSH, have updated their software to support stricter key exchanges. The discovery of Terrapin necessitates ongoing vigilance and potentially a complete redesign of the SSH protocol, underlining the importance of continuous evaluation and advancement in network security technologies. (Ars Technica)
TESLA SAFETY RECALL INSUFFICIENT.
Consumer Reports has labeled Tesla’s solution to the Autopilot recall of over 2 million vehicles as “insufficient.” Their preliminary tests, conducted by Kelly Funkhouser, revealed that it’s still possible to obstruct the cabin camera while using Autopilot, which is critical for ensuring driver attention. Additionally, no noticeable changes were found in the operation of the Autosteer feature, particularly on non-highway roads, despite Tesla’s claims of enhancements. The recall, initiated amidst a National Highway Traffic Safety Administration investigation, focuses on the Autosteer feature’s inability to restrict its use to controlled-access highways. Although Tesla has implemented a suspension policy for Autopilot misuse, this wasn’t encountered in Consumer Reports’ short test drives. The organization plans further testing, having not yet evaluated other updates like improved visual alerts. This situation raises questions about Tesla’s driver monitoring technology and the ambiguity of their software updates, underscoring ongoing safety concerns related to the Autopilot system. (Techcrunch)
SNL UNFUNNY.
A “Saturday Night Live” sketch mocking sickle cell disease treatments has drawn sharp criticism from the sickle cell community and advocates. The sketch depicted a white employee gifting a Black colleague with a gene therapy for sickle cell anemia, only for it to be humorously swapped for a trivial item. This portrayal was seen as insensitive, perpetuating stereotypes and trivializing the disease. Sickle cell affects about 100,000 people in the U.S. and 20 million worldwide and is often associated with discrimination and misconceptions in medical care. Advocates, including Ashley Valentine and Mariah Scott, expressed dismay at the skit’s reinforcement of the stereotype that sickle cell is exclusively a “Black disease” and the trivialization of patient experiences. The controversy comes at a crucial time, with new sickle cell therapies emerging. Advocacy groups are urging SNL to use its platform to correct the record and highlight the serious challenges faced by those with sickle cell disease, emphasizing the need for greater awareness and sensitivity in media portrayals. (STAT)
C IN CHATGPT IS FOR CHEMIST.
Chemists at Carnegie Mellon University, led by Gabe Gomes, have developed a system called Coscientist, using ChatGPT to autonomously design and conduct complex chemical reactions with a robotic laboratory setup. Utilizing advanced language models like GPT-4, Coscientist scours chemical literature to formulate reaction pathways for synthesizing specific molecules. The system has successfully created compounds such as paracetamol and aspirin and executed intricate reactions like the Suzuki–Miyaura coupling, important in drug discovery. This innovation marks a significant stride in making AI tools accessible to chemists, particularly those unfamiliar with programming. Coscientist represents a shift towards automated laboratories integrated with AI assistants, potentially revolutionizing the design-make-test cycle in pharmaceutical chemistry. While routine tasks can now be automated, more complex research questions, especially those involving biology in drug discovery, remain a challenge. The team is mindful of the ethical implications and potential risks of widespread use of such technology, highlighting the importance of using these advancements to complement human expertise rather than replace it. (Nature)
TALK TALK.
A new study analyzing nearly 40,000 hours of audio recordings from children worldwide suggests that children tend to speak more when surrounded by talkative adults. This increased interaction may contribute to a larger vocabulary early in life, regardless of factors like social class. The research, involving data from 1001 children across six continents, found no significant link between a mother’s education level or social class and the child’s speech development. Most participants were from English-speaking families, but the study also included diverse languages such as Dutch, Spanish, Vietnamese, Finnish, Yélî Dnye, Wolof, and Tsimané. The findings show that children speak more in response to increased adult vocalizations, contradicting previous studies that suggested a link between social class and vocabulary development. The research implies that direct interaction with children is more crucial for language development than socioeconomic factors. While further research is needed to understand the full impact of social class on language acquisition, the study suggests that supportive and interactive communication with children is key to their linguistic development, rather than a focus on excessive speech. (Science)
Thanks for reading. Let’s be careful out there.
WORDS: The Biology Guy.
IMAGE CREDIT: Microsoft.